Accepted Papers

• A-GRAG: An Agentic GraphRAG Framework with Collaborative Agents for Temporal-Aware Multi-Hop Reasoning

  Haowei Mo, Fangling Leng, Yubin Bao, Tiancheng Zhang

  Abstract

  Retrieval-Augmented Generation (RAG) mitigates hallucinations and outdated knowledge in Large Language Models (LLMs) by incorporating external knowledge. As a structured extension of RAG, GraphRAG organizes entities and relations from documents into a knowledge graph, thereby enhancing the ability of LLMs to reason over multi-hop queries. However, existing GraphRAG methods typically rely on one-shot retrieval and struggle to dynamically adjust retrieval targets during reasoning, while existing agentic retrieval methods rarely model knowledge graph structure and temporal constraints jointly. To address these limitations, we propose Agentic GraphRAG (A-GRAG), a framework that combines a temporally aware knowledge graph with agentic iterative retrieval. A-GRAG consists of a Temporal Extraction Agent (TEA) and an Iterative Retrieval Agent (IRA). During graph construction, TEA extracts and normalizes temporal information from entity descriptions and writes it to the corresponding entities as explicit attributes. During querying, TEA extracts temporal constraints from the current query and attempts to locate matching anchor entities using these constraints; if no match is found, the system falls back to the standard graph-retrieval procedure. IRA acquires candidate evidence, assesses whether it is sufficient to answer the original query, and, when the evidence is insufficient, identifies information gaps, compresses the auxiliary context, and generates a subquery for the next round, with at most three retrieval rounds. Experiments on three multi-hop question answering benchmarks show that A-GRAG achieves the best results in both EM and F1. Ablation results further show that, compared with the variant without temporal awareness, the full A-GRAG model maintains slightly better answer quality while reducing the average token consumption per question by 23.7%--30.9% and average end-to-end latency by 12.8%--17.8%.

• Accelerating Meta-Reinforcement Learning with Predictive Curriculum Scheduling

  Fangjie Yang, Jing Yang, Yun Qu, Xingxing Liang, Guanlin Wu, Qi Wang, Longfei Zhang

  Abstract

  Meta-reinforcement learning (Meta-RL) enables rapid adaptation by leveraging shared structure across tasks, but suffers from poor sample efficiency due to costly meta-training and inefficient task sampling. In particular, uniform task sampling ignores the agent's evolving competence, wasting interaction on tasks that are either trivial or currently unsolvable. While curriculum learning offers a principled alternative, existing methods either rely on expensive online difficulty estimation or employ amortized predictors whose absolute difficulty estimates tend to drift as the learner improves, leaving them misaligned with the learner's competence. We propose Predictive Curriculum Scheduling (PCS), an adaptive framework that unifies amortized difficulty prediction with competence-aware curriculum design. PCS learns a predictive model of relative task difficulty with a rank-preserving objective, encouraging the task ordering used for scheduling to be less sensitive to shifts in absolute difficulty values during training. It further introduces a scheduling mechanism that dynamically tracks the agent's competence and prioritizes tasks just beyond its current mastery, enabling fully automated curriculum generation without additional environment interaction. Experiments on standard Meta-RL benchmarks demonstrate that PCS consistently improves sample efficiency, accelerates convergence, and enhances training stability over strong adaptive task-sampling baselines.

• ADRO: Adaptive Dual-Timescale Response Optimization for Transfer Reinforcement Learning across Diverse Opponent Styles

  Daojie Guo, Yixing Lan, Tenglong Liu, Long Cheng, Xuhui Xie, Xin Xu

  Abstract

  Policies trained against a finite set of opponents can lose performance when opponent competence, action preferences, or strategy-witching patterns change. Opponent-modeling methods infer behavior from interaction history, while policy-reuse methods select or combine learned source policies. Adapting to a changing opponent requires balancing evidence accumulated over longer interactions with recent observations. We propose Adaptive Dual-Timescale Response Optimization (ADRO). ADRO uses a shared history-conditioned model to construct a long-term reference response policy and a short-term candidate response policy from long and recent interaction windows, respectively. The centered log-probability difference between the two policies defines a state-dependent exponential-family path. Dual-critic directional evidence and Fisher curvature determine the nominal transfer coordinate, while a forward Kullback–Leibler (KL) constraint limits departure from the reference policy. Across nine competitive tasks and one cooperative task, ADRO achieves an aggregate out-of-distribution(OOD) task-outcome interquartile mean of 86.3%, compared with 72.2% for the strongest evaluated baseline. The aggregate combines competitive win rates with relative-return success in the cooperative task under predefined parameterized opponent shifts.

• Agentic Code Repair via Execution-Grounded Selective Planning

  Wei Wei, Teng Wang, Lin Li, Lijun Zhang

  Abstract

  Large language models can repair programs, but their outputs may still contain errors involving boundary conditions, interfaces, runtime behavior, or algorithmic logic. Existing approaches use execution feedback and planning to improve repair accuracy. However, repeated revisions increase inference cost, and limited visible tests may not provide sufficient evidence to justify replacing a partially correct program. To address these challenges, we propose Execution-Grounded Selective Repair (ESR), a centralized agentic framework that separates planning allocation from program updates under a bounded inference budget. ESR first generates direct repair candidates and evaluates them on visible tests. If no candidate passes all visible tests, the Planning Agent analyzes execution records from the complete candidate pool to identify recurring failure patterns. The resulting Failure Sketch, Repair Plan, and preservation constraints guide a subsequent Plan-Guided Repair stage. In both stages, PublicSelect ranks candidates by the number of visible tests they pass, while Guarded Update accepts a replacement only if it passes strictly more visible tests than the incumbent program. Experiments on TACO under matched generation budgets show that ESR outperforms additional direct sampling with both backbones and the evaluated PaT control with Qwen2.5-Coder-7B-Instruct. Across four Python benchmarks and two code backbones, ESR achieves the highest observed point estimate of Final Pass@1 in all eight combinations of backbone and benchmark. Selective planning reduces the number of model generations compared with a strategy that always executes both repair stages.

• AGES: Audit-Gated Evaluation Scheduler for Combinatorial LLM Context Evaluation

  Zhu Weilin, Mu Weimin, Fan Liu, Jian Chen, ChengHanchang, Weiping Wang, Yun Zhang

  Abstract

  AI systems, including LLM-based agents, often construct prompts from retrieved memories or few-shot demonstrations. Measuring how individual context items and their combinations affect performance supports regression testing, debugging, and attribution. For n candidate items, a complete analysis may require scoring all 2^n subsets. Existing methods pursue related but different objectives. Selection methods return only a small number of subsets, while reconstruction methods estimate missing scores under structural assumptions. Compression methods identify repeated information after the scores are available. These approaches do not directly address whether an observed score can replace a new evaluation while the score table is being acquired. We study this question using frozen few-shot demonstrations in a controlled, single-turn code-generation setting. We introduce the Audit-Gated Evaluation Scheduler (AGES) to reduce this cost. AGES uses score agreement on small subsets to propose reuse, then audits the proposal on larger subsets before copying any score. If the audit finds a mismatch, AGES retains all observed scores and evaluates every remaining subset. The fallback completes the table exactly without discarding previous evaluations. We evaluate AGES through offline replays of score tables obtained with deterministic Qwen2.5-Coder generation. With a shared demonstration pool, the frozen procedure reduces simulated unique score reads by 4.2% on LeetCodeDataset and 16.2% on 588 query-disjoint MBPP+ model–query score tables, compared with exhaustive evaluation. On MBPP+, 0.026% of copied scores are incorrect, and the selected subset attains a true maximum in all 588 tables.

• Aligning Context and Formulas: Multi-Stage Fine-Tuning for Large Language Models with A Novel Dataset ContextFormulas50K

  Xinyu Yin, Han Xiao

  Abstract

  Mathematics plays a crucial role in scientific research. However, it is never easy to formulate a problem using mathematical formulas, even for senior researchers. Fine-tuning pre-trained models on mathematical datasets to facilitate research has become a widely adopted approach. Although there are numerous mathematical datasets, only high-quality datasets could enhance the pre-trained model for a deep insight into formulas because mathematics is a precise discipline. To address this, we propose the ContextFormulas50K dataset, which consists of mathematical formulas paired with their contextual text. Based on this dataset, we fine-tune two pre-trained models (i.e. LLaMA3-8B-Instruct and Mistral-7B) to generate precise formulas for research work. But in this task, there emerges the information bottleneck problem, which means the parameter scale expands much faster than performance. To overcome this problem, we introduce a Multi-Stage Fine-Tuning(MSFT) approach to equip large language models (LLMs) with better mathematical comprehension. Specifically, there are three stages in our model, and by progressively inserting plug-and-play modules, model performance could be enhanced at each stage, which means our method effectively alleviates information bottlenecks. Experimental results show that our model effectively improves mathematical comprehension and achieves state-of-the-art performance in the formula generation task, outperforming multiple commercial baselines (i.e., GPT-4, GPT-4o, and Claude-3.5-sonnet).

• Applicability-Aware Evidence Reasoning for Knowledge Graph-Enabled Agents with Proposition Graphs

  Hejin Fang, Huihui Shao, Fanyu Wang, Zhenping Xie

  Abstract

  LLM-based agents increasingly interact with knowledge graphs (KGs) to acquire structured external knowledge for knowledge-intensive decision-making. However, reliable decisions require more than retrieving relevant entities or multi-hop paths: agents must also determine whether the acquired evidence is applicable under the current context, identify its prerequisites, resolve conflicts, and organize coherent support. Existing KG-enabled agents largely leave this evidence-reasoning stage to the implicit reasoning of the LLM, making them vulnerable to evidence that is factually plausible but conditionally valid, mutually conflicting, or contextually inappropriate. We propose PG-LLM, a plug-and-play applicability-aware evidence-reasoning module for KG-enabled agents. PG-LLM augments the evidence-reasoning process with a Proposition Graph (PG), a propositional reasoning substrate whose nodes represent self-contained propositions and whose typed edges encode support, requirement, inhibition, equivalence, and contradiction relations. Given a query and KG-acquired evidence, PG-LLM retrieves a query-specific PG context and uses its proposition content, graph structure, and typed relations to retain, qualify, reorganize, or suppress KG evidence. The PG context informs the refinement process but is not directly supplied to the answer generator, which receives the reorganized KG-derived evidence. We evaluate PG-LLM by isolating the evidence-reasoning stage in four KG-enabled reasoning pipelines while preserving their original KG-access mechanisms. Experiments on two medical benchmarks show consistent improvements in accuracy and safety-related outcomes, including an average accuracy gain of 7.84 percentage points (median +3.91) and a 24.18-percentage-point reduction in the Unsafe rate, while using substantially fewer evidence units. These results provide evidence that explicit applicability-aware evidence reasoning can improve the reliability of KG-enabled pipelines in the evaluated medical settings.

• AutoMind: Adaptive Knowledgeable Agent for Automated Data Science

  Yujie Luo, Yixin Ou, Jingsheng Zheng, Lanning Wei, Zhuoyun Yu, Shuofei Qiao, Jintian Zhang, Da Zheng, Yuren Mao, Yunjun Gao, Huajun Chen, Ningyu Zhang

  Abstract

  Large Language Model (LLM) agents have shown great potential in addressing real-world data science problems. LLM-driven data science agents promise to automate the entire machine learning pipeline, yet their real-world effectiveness remains limited. Existing frameworks depend on rigid, pre-defined workflows and inflexible coding strategies; consequently, they excel only on relatively simple, classical problems and fail to capture the empirical expertise that human practitioners bring to complex, innovative tasks. In this work, we introduce AutoMind, an adaptive, knowledgeable LLM-agent framework that overcomes these deficiencies through three key advances: (1) a curated expert knowledge base that grounds the agent in domain expert knowledge, (2) an agentic knowledgeable tree search algorithm that strategically explores possible solutions, and (3) a self-adaptive coding strategy that dynamically tailors code generation to task complexity. Evaluations on two automated data science benchmarks demonstrate that AutoMind delivers superior performance versus state-of-the-art baselines. Additional analyses confirm favorable effectiveness, efficiency, and qualitative solution quality, highlighting AutoMind as an efficient and robust step toward fully automated data science.

• Autonomous Prompt Evolution Framework for Combinatorial Auctions

  Haihui Zang, Heng Song

  Abstract

  LLM agents in dynamic competitive environments must both plan strategically and adapt from past interactions under substantial uncertainty. This paper studies combinatorial auctions, where agents bid under budget constraints while inferring hidden complement and substitute relations among items. Existing reflection and memory-augmented agents reuse cross-episode experience but summarize trajectories holistically, making it hard to identify which structural failure caused a missed bundle or redundant acquisition. We propose the Autonomous Prompt Evolution Framework (APEF), which treats the agent's natural-language intention as an evolvable policy optimized across episodes without weight updates, combining contrastive dual memory, in-context marginal-value simulation, and surprise-gated diagnostic evolution to convert synergistic failures and substitutable redundancies into executable prompt edits. On CABench, a synergy-aware extension of AucArena, APEF achieves the highest mean combinatorial profit across auction-order variation, randomized instances, backbone comparisons, and ablations, while maintaining low winner's-curse and no-acquisition rates—showing that structurally diagnosed prompt adaptation outperforms holistic reflection alone.

• Beyond Retrieval Failure: Understanding Why LLMs Miss Evidence in Multi-Hop RAG

  Jiaqi Wei, Wenxuan Huang, Chengdong Xu, Xiang Zhang, Wanli Ouyang

  Abstract

  Multi-hop retrieval-augmented generation (RAG) requires an agent to retrieve evidence for a sequence of dependent sub-questions. When a wrong answer lacks a required passage, the passage may be inaccessible to the retriever, or the agent may have failed to issue a query that would retrieve it. Standard retrieval evaluations conflate these two causes. We study how much apparent retrieval failure is instead caused by search drift: the search no longer tracks an unanswered sub-question. Holding the retriever, corpus, and generation model fixed, we introduce \methodname, a training-free requirement register that records open sub-questions and their supporting evidence. The register changes search guidance but not evidence access, so each gold passage it recovers certifies that the passage was reachable by the original retriever. On 1{,}000 MuSiQue questions, the number of wrong answers with unretrieved required evidence falls from 382 to 234 with all-MiniLM-L6-v2 and from 383 to 274 with Qwen3-Embedding-0.6B. These counts correspond to 69% to 44% and 66% to 48% of each system's wrong answers, respectively. Crucially, evidence recovery does not produce a proportional reduction in total error: premature resolutions and abstentions increase as missing-evidence failures decrease. The central finding is therefore a shift in the remaining bottleneck, from guiding search toward required evidence to verifying that the retrieved evidence is sufficient.

• Budgeted Guidance Refresh Control: A Paired Throughput--Call Study in Lifelong Multi-Agent Path Finding

  Xiaoxiao Ma, Quan Fang, Xiankun Jiang

  Abstract

  Lifelong multi-agent path finding (LMAPF) requires robot fleets to serve a continuous stream of goals while maintaining high throughput under changing traffic conditions. Learned global guidance can improve coordination, but existing systems commonly refresh guidance at fixed intervals: frequent refresh wastes generator calls when demand is stable, whereas infrequent refresh can leave guidance stale after workload shifts. We formulate refresh timing as an independent lifecycle-control problem, operationalizing resource use as the number of logical guidance-generator calls, and instantiate it with Context-Aware Refresh and Recall (CARR). Using only online signals—demand change, context similarity, guidance age, and route adoption—CARR chooses among retaining the active guidance, recalling a cached guidance version, and generating a new one. It wraps a frozen CNN guidance generator and a frozen GPIBT planner without retraining either component, isolating the effect of refresh control. We evaluate eight refresh policies on stationary and non-stationary warehouse workloads using paired, policy-independent task streams and 3,840 runs organized into 40 independent root clusters. The 26-call periodic reference (Exact-B25) retains a small mean-throughput advantage over CARR (CARR effect: −0.80%; 95% CI: [−1.19%, −0.40%]), so the prespecified 1% non-inferiority objective is not established. Nevertheless, CARR uses approximately 79% fewer calls, forms a non-dominated sample-mean throughput–call operating point, and significantly outperforms all five evaluated low-call comparators; historical recall primarily substitutes for fresh generations rather than improving throughput. These results identify guidance refresh timing as a distinct systems-design dimension and provide a reproducible basis for selecting refresh policies within a fixed LMAPF guidance backbone under call constraints.

• C-PAQ: A Counterfactual Provenance-Aware Quorum for Echo-Resistant Multi-Agent Fact Verification

  Lingdong Meng, Zeyu Song, Yu Guo

  Abstract

  Multi-agent large language model systems commonly aggregate agent judgments through majority voting, confidence weighting, or an LLM-based judge, yet these approaches often assume that agent reports are independent. In practice, communication can violate this assumption: downstream agents may copy or paraphrase an upstream conclusion without acquiring new evidence, causing the same information to be counted repeatedly and creating artificial corroboration. Distinguishing genuine support based on independent evidence from echo amplification caused by dependent communication, while ensuring that the final verdict covers all necessary evidence, remains a key challenge in multi-agent fact verification. To address this problem, we propose the Counterfactual Provenance-Aware Quorum (C-PAQ), a training-free aggregation mechanism. C-PAQ first performs a blind leave-one-source-out counterfactual audit to identify the evidence sources required for verifying a claim, then uses controller-recorded ancestry and direct source-access logs to contract dependent reports into evidence lineages. A claim is accepted as Supported only when the complete evidence packet passes verification, the required-source certificate is non-empty, and at least two distinct supporting lineages jointly cover all required sources. We evaluate C-PAQ on 120 matched complete and one-hop-missing evidence packets constructed from 60 HoVer claims with HotpotQA contexts, together with several controlled interventions. C-PAQ maintains an accuracy of 92.5% with both zero and seven dependent reports. In a conditioned stress test, seven descendants of an erroneous root reduce majority-voting accuracy from 100% to 0%, whereas C-PAQ retains 94.1% accuracy. Further experiments show that C-PAQ resists false support caused by copied citations and same-lineage re-access, while remaining responsive when an independent lineage acquires the missing evidence. These results demonstrate the effectiveness of C-PAQ in suppressing dependence-induced echo amplification and enforcing necessary-evidence coverage in multi-agent fact verification.

• CACR: A Reconstructible Context Plane for Cloud-Native Agent Serving

  Wujian Huang, Xueyan Yang, Wenhai Wang, Yanting Zhang

  Abstract

  Most agent context architectures are local-first: a resident process owns the transcript, orders memory updates, and assembles each prompt. This design can work on a single host but does not by itself provide reliable context guarantees in cloud-native serving, where requests move across stateless replicas. Shared caches, indexes, and background writers also evolve independently, so model calls may observe stale, incomplete, or cross-tenant evidence. Existing memory and retrieval mechanisms optimize what to retain or retrieve but do not specify which evidence each cloud request may present to the model. We present the Cloud Agent Context Runtime (CACR), a context plane that expresses scope, authority, ordering, and evidence requirements as a request-level context contract. CACR stores scoped events and authorized artifacts as durable authority, treats recent windows, retrieval results, and compact state as replaceable projections, and deterministically selects among three context-construction paths. It also records requested and realized evidence, making degradation explicit. In Kubernetes and deployments on three Elastic Compute Service (ECS) instances, CACR preserves exact state across all 30 A–B–A dual-control rounds. It rejects foreign evidence in all 20,661 scope probes, even when tenant embeddings are identical. On LongMemEval-Oracle, every policy receives the same frozen, question-specific annotated evidence, system and task instructions, reader, and judge, isolating model-bound context construction from evidence discovery. Under this protocol, the CACR context pipeline achieves 93.6% accuracy, 7.2 and 11.8 percentage points above Always Recall and Full History, respectively. With the same prepared CACR evidence, open-weight 8B–27B readers achieve 83.6–86.2% accuracy. Under these workloads, CACR enforces context correctness at runtime without relying on process affinity.

• Candidate Evidence Support Estimation and Routing from Single-Pass Generation Observations

  Shiyu Zhu, Hao Chen, Yichen Liu, Tang Shouyuan

  Abstract

  After candidate retrieval, a long-term memory system must decide which evidence windows fit within a limited generation context. Lexical or semantic similarity does not directly describe the generator's process state while using a candidate. We study candidate evidence-coverage readout between semantic-memory formation and final generation. For each question–candidate pair, a frozen model performs one deterministic trial answer; 20-dimensional predictive-entropy trajectories are extracted from the full vocabulary distribution and 22-dimensional stage statistics from attention allocations. An $L_2$ logistic regression forms the 42-dimensional Candidate Evidence Support Readout (CESR-42). On LoCoMo test candidates, CESR-42 obtains 0.8807 AUROC and 0.6680 AUPRC, exceeding the strongest scalar baseline, Sum NLL, at 0.7743 and 0.3300. On 351 test questions from four standard long-term-memory QA categories, it achieves 0.9164 Recall@5, 0.8871 MRR@5, and 0.8659 nDCG@5, with all point estimates above BM25's 0.9091, 0.8555, and 0.8429. Downstream answer metrics show different changes; in MultiHop, high evidence-identifier coverage does not guarantee nonzero answer-overlap scores. We position single-pass generation observations as an auditable and composable candidate-evidence routing interface between semantic memory and the final generator.

• CARVS: A Congestion-Aware Algorithm for Online Multi-Agent Pickup-and-Delivery in Human-Collaborative Warehouses

  Chaoran Li, Zeyuan Zhao, Shao Zhang

  Abstract

  Intelligent warehouse systems have transformed logistics by integrating robots with minimal infrastructure changes, boosting efficiency and optimizing workflows compared to traditional human-only operations. However, human-robot collaboration introduces new challenges, particularly due to human uncertainty, which leads to dynamic interactions and unpredictable behavior. To address these complexities, this paper presents the Multi-Agent Pickup-and-Delivery with Human Collaboration (MAPDwHC) problem, an extension of the Multi-Agent Pickup-and-Delivery (MAPD) problem that incorporates human unpredictability. This problem adds key constraints, such as dynamic obstacles caused by robots temporarily blocking paths when interacting with humans, and the need for real-time responsiveness. We formalize the MAPDwHC problem with additional constraints and properties. Based on this problem setting, we propose the CARVS algorithm, which improves task assignment and path planning by integrating task swapping, validation checks, and random initialization. CARVS reduces congestion, enhances adaptability to human-robot interactions, and ensures real-time responsiveness. Simulation results show that CARVS significantly reduces timeouts of path planning in large-scale scenarios where traditional MAPD algorithms fail, improving both efficiency and scalability.

• CGPST: A Multi-Step Creativity Benchmark for Complex Creative Problem-Solving of Large Language Models

  Xiangyu Wang, Jin Wu, Wei Xia, Haoran Shi, Chanjin Zheng

  Abstract

  Recent rapid advancements in Large Language Models (LLMs) have highlighted their promising potential across various creativity tasks. However, effectively measuring the creativity of LLMs remains a critical challenge. Existing paradigms predominantly rely on traditional psychological tests (e.g., the Alternative Uses Test) or simple tasks focusing on specific domains (e.g., creative story generation). These tasks lack rich contextual stimuli to evaluate deep creative thinking and focus narrowly on single-step divergence or conceptual association, thereby ignoring the process-oriented characteristics of real-world creative problem-solving. To address these limitations, we propose Contextually-Grounded and Procedurally-Structured Tasks (CGPST), a novel benchmark for evaluating creative problem-solving abilities of LLMs with future scenario and multi-step design. CGPST is defined by four core features: contextual grounding, procedural structure, process-oriented evaluation, and multi-dimensional assessment. Furthermore, we conduct extensive experiments with 9 popular LLMs on CGPST coupled with rigorous human scoring, constructing a high-quality dataset comprising 270 full-process responses and 540 expert human ratings. Additionally, we explore preliminary automated scoring for such complex multi-step creativity tasks, providing practical directions for automated creativity evaluation.

• ChronoMem: When Relevance Is Not Validity in Long-Term Assistant Memory

  Yujie Shen

  Abstract

  Long-term memory is becoming a practical component of language-model assistants, agent runtimes, and personalized interactive systems. Yet a remembered preference is not simply a fact to retrieve. A user may replace it, make a temporary exception, scope it to work or personal use, or delete it. A memory can therefore be semantically relevant to a later query while being invalid for the current time and context. We introduce ChronoMem, a controlled benchmark for evaluating whether assistant-memory policies select the currently valid preference after temporally scoped memory events. The benchmark contains 360 auditable event sequences spanning 12 preference domains, five event types, and five temporal lags. We compare a transparent rule-based temporal filter, ChronoGate, with full-history prompting, recency, keyword retrieval, TF-IDF retrieval, time decay, and an Ollama dense-retrieval baseline across DeepSeek R1 8B, Qwen 2.5 7B, and Gemma 3 4B. The story is deliberately two-sided. On the original controlled benchmark, ChronoGate is strongest for both full-matrix models, reaching 1.000 accuracy with Qwen and 0.406 with DeepSeek, with positive paired-bootstrap intervals against every evaluated baseline. A harder lexical held-out split changes the conclusion: ChronoGate drops to 0.350, 0.708, and 0.722 on DeepSeek, Qwen, and Gemma, while full history or dense retrieval often matches or exceeds it. These results support ChronoMem as a useful benchmark for temporal validity in agent memory, but they also show that lexical heuristics are not a sufficient basis for general long-term memory systems.

• Chronos-Lite: Causality-Aware Message Integration for Order-Robust LLM Agent Teams

  Chenxu Zhu, Zhiran Wang

  Abstract

  Parallel LLM agents can produce related updates under asynchronous execution, so message arrival order need not reflect information order. An integration policy based on arrival order may therefore allow a delayed ancestor to overwrite a causally newer revision. We formalize this failure as stale-message overwrite and introduce Chronos-Lite, a deterministic message-integration protocol that records parent dependencies and vector timestamps, maintains a maximal causal frontier, rejects causally dominated messages, and reconciles incomparable messages using a permutation-invariant rule. Under trusted causal metadata, Chronos-Lite satisfies a stale non-overwrite property; this structural guarantee does not imply semantic correctness. We evaluate eight integration methods using fixed cached outputs from heterogeneous DeepSeek and Qwen agents on 100 GSM8K and 100 ARC-Challenge tasks. Controlled replay covers 20,000 delivery schedules and 160,000 method traces, with paired bootstrap inference performed at the level of the 200 independent tasks. Across a balanced mixture of natural and diagnostic stress schedules, Chronos-Lite attains 94.51% accuracy compared with 91.79% for arrival-order last-write-wins, a difference of 2.72 percentage points with a 95% confidence interval of [1.15, 4.05]. It also reduces order sensitivity from 3.87% to 0.04% and correct-to-wrong transitions from 5.31% to 1.00%. The mean-accuracy difference is small under natural schedules, 94.50% versus 94.21%, and larger under targeted stale-message stress, 94.52% versus 89.36%. Lamport reaches 95.48% accuracy, so the current evidence does not support a mean-accuracy advantage for Chronos-Lite over a scalar total-order baseline. The guarantees are stale non-overwrite and same-set permutation invariance; the measured mixed-schedule robustness also reflects omissions. Final correctness remains dependent on the semantic resolver and the underlying agents.

• Co-opetition Equilibrium in Adversarial Team Games with Heterogeneous Utilities

  Youzhi Zhang

  Abstract

  The United Nations' 2030 Agenda for Sustainable Development requires all countries to collaborate against adversarial factors, a scenario that can be formalized as an adversarial team game. However, existing solution concepts assume team players share identical utility functions—an assumption inconsistent with real-world settings where countries have divergent objectives. This paper argues that studying adversarial team games must account for heterogeneous utilities among team players. We show that ignoring utility differences can render computed equilibria unstable in the original game, and we formalize this degradation via the Price of Ignoring Heterogeneity, which can be unbounded. To address this, we introduce the Co-opetition Equilibrium (CoE), where team players with heterogeneous utilities correlate their strategies (cooperation) against an adversary (competition). We establish existence via reduction from Nash equilibrium, and prove that finding a CoE is PPAD-complete while computing a Team-Maximizing CoE (TMCoE) is NP-hard. Nevertheless, we identify a broad class of zero-sum adversarial team games—those satisfying a consistent-constraint condition that generalizes identical utilities—where TMCoEs are exchangeable and computable in polynomial time via linear programming. We outline directions for algorithm design, MARL, and extensions to extensive-form and multi-team games.

• Collab-Solver: Collaborative Solving Policy Learning for Mixed-Integer Linear Programming

  Siyuan Li, Yifan Yu, Zhihao Zhang, Mengjing Chen, Fangzhou Zhu, Tao Zhong, Jianye Hao, Peng Liu

  Abstract

  Mixed-integer linear programming (MILP) is a fundamental problem in combinatorial optimization, and conventional MILP solving primarily relies on carefully designed heuristics within the branch-and-bound framework. Driven by the strong capabilities of neural networks, recent research is exploring machine learning techniques alongside conventional MILP solving. Although learning-based MILP methods have shown great promise, existing work typically learns policies for individual modules in MILP solvers in isolation, without considering their interdependency, which limits both solving efficiency and solution quality. To address this limitation, we propose Collab-Solver, a novel multi-agent-based policy learning framework for MILP that enables collaborative policy optimization across multiple modules. Specifically, we formulate the interaction between cut selection and branching as a Stackelberg-inspired leader-follower framework. Under this formulation, we develop a two-phase learning paradigm to stabilize collaborative policy learning: the first phase performs data-communicated policy pretraining, and the second phase further orchestrates the policy learning for various modules. Extensive experiments on both synthetic and large-scale real-world MILP datasets demonstrate that the jointly learned policies significantly improve solving performance. Moreover, the policies learned by Collab-Solver have also demonstrated excellent generalization abilities across different instance sets.

• CoMAT A Collaborative Multi-Agent Framework for Black-Box Jailbreak Attacks via Attack-Trajectory Learning

  Fumin Cai, Jianing Li, Hao Li, Aimin Hao

  Abstract

  As large language models are widely deployed in conversational assistants, content generation, code development, and agentic systems, their safety-alignment mechanisms continue to face substantial challenges. Automated jailbreak attacks are an important tool for identifying safety weaknesses in models. However, existing white-box methods usually rely on internal model signals and are therefore sensitive to changes in model architecture and alignment strategies. Although black-box methods are more broadly applicable, they often require a new multi-round search for each incoming instruction, leading to high target-model query costs and external model invocation costs. To address these issues, this paper proposes a multi-agent black-box jailbreak attack framework for iterative optimization of attack instructions. The framework uses a teacher agent to collect successful and failed attack trajectories, converts successful trajectories into round-by-round supervised data for training a local attack agent, and thereby reuses historical instruction-transformation experience. Meanwhile, successful and failed candidates are used to train a screening agent, which filters low-potential instructions before querying the target model and reduces ineffective exploration. We conduct systematic evaluations on two safety benchmarks, HarmBench and AdvBench, against three closed-source models and three open-weight models. Experimental results show that the proposed method achieves cross-model average attack success rates of 93.3% and 92.8% on HarmBench and AdvBench, respectively, improving over the strongest baseline by 22.1 and 23.6 percentage points. On HarmBench, the proposed framework requires an average of 11.12 target-model queries, and candidate screening reduces the average number of queries on Qwen3.5-9B from 28.91 to 13.83. These results demonstrate that attack-trajectory learning and candidate pre-screening can improve the generalization ability and query efficiency of black-box jailbreak attacks while maintaining high attack effectiveness.

• Conditional Best-Agent Discovery with Horizon-Limited Peer Audits: Overshoot Closure, Whole-Batch Path Certification, and Scoped Query Barriers

  Qifan Yang, Yujiao Peng

  Abstract

  We study how a finite archive of agents can identify its best member without assuming that any evaluator is reliable across all candidates. Starting from one trusted seed agent, our whole-batch path walk (C1-PW) promotes each accepted candidate to serve as the next evaluator. The core challenge is that an evaluator may reliably recognize modest improvements yet misjudge a much stronger candidate beyond its unknown verification horizon. We formulate an overshoot-closed verification hull (OCVH) condition ensuring that every non-best agent reachable through true improvements retains a locally reliable outgoing step. Under OCVH, a known local margin, and one-sided response bounds, we prove that following the first accepted certificate identifies the unique best agent with probability at least $1-\delta$, even under arbitrary history-dependent super-horizon feedback. The protocol completes one edge batch at a time with exact query cost $Q=mD$, where $m$ is the batch size and $D$ is the number of tested edges. Complementary lower bounds separate the statistical cost of bridge verification from the combinatorial cost of certified successor search. Across 6,400 compliant synthetic trials (17 OCVH-compliant cells), C1-PW made zero identification errors, with mean query savings relative to full branch saturation ranging from 0% (on seed-best and chain controls) to 79.7%. Because OCVH concerns latent utilities and horizons that cannot in general be verified from audit transcripts alone, these results establish an explicitly conditional guarantee.

• Constitutional Filtering of Manipulative Influence in Grounded LLM Multi-Agent Systems

  J. de Curtò, I. de Zarzà

  Abstract

  Existing frameworks for governing large language model (LLM) multi-agent systems are typically evaluated within closed simulations, where cooperation and manipulation are defined only relative to internal model quantities and lack an external referent against which they can be assessed as true or false. We address this grounding gap with TRACE (Truth-Referenced Auditable Cooperative Epistemics), in which agents are legible epistemic subjects holding persistent natural-language beliefs about synthetic questions with a known ground truth, an LLM influence compiler issues structured directives, and a constitutional layer filters them. “Cooperation” becomes collective truth-tracking, “manipulation” becomes displacement toward a false anchor, and agents’ reported rationales provide an auditable record. Across 706 simulation runs, with Llama-3.3-70B-Instruct as the default model, constitutional governance preserves collective accuracy (+0.033, p = 0.023) and suppresses manipulation drift (−0.070, p = 0.017), whereas a false anchor labelled FACTUAL clears the constitution on every run yet still degrades truth-tracking (p = 0.030). This exposes the insufficiency of filtering by self-declared honesty labels. The accuracy benefit decreases as the informed fraction grows, with the reported unpaired tests significant through ρ = 0.40 and perfect rank separation at ρ = 0.10 and 0.25. Accuracy benefits are positive for all four tested models and individually significant for one. Benefits are also positive across thirteen language conditions with translated question stems and a language-invariant policy schema (pooled p = 0.0002). Finally, we port the framework to a simulated 2D drone swarm, where governance more than doubles mission success under command-channel spoofing (0.31 to 0.69) and limits Byzantine insider pressure when the malicious channel is correctly identified. In the tested configurations, external commands cause greater degradation than lateral peer pressure; the attack magnitudes are not matched. Results are reported with bootstrap confidence intervals and rank-based tests.

• Control Under Compression: Reliability Frontiers for Tool-Using Agents

  Yinghan Hou, Zongyou Yang

  Abstract

  Tool-using language-model agents are governed not only by task prompts but also by persistent system-side instructions that specify tools, arguments, policies, execution protocols, and recovery. Compressing these agent control contexts (ACCs) can reduce input cost and context use, yet existing prompt-compression evaluations do not reveal whether the resulting control remains operationally reliable. We introduce CompressAgent, an environment-verified benchmark for ACC compression across nine independently constructed ACCs, three task families, three fixed Qwen API model identifiers, six retained-context budgets, and 15,525 runs. We uncover a nonlinear, method-dependent reliability frontier. At 75% retained context, generic rewriting and section-based compression achieve 92.7% and 92.4% success, close to the 93.8% full-context baseline. Between 50% and 35%, methods diverge sharply; at 35%, section-based, obligation-aware, and generic rewriting achieve 47.0%, 39.0%, and 19.9%. At 25–10%, executable protocols become fragile. Reliability also varies substantially across ACCs, making universal compressor rankings inappropriate and motivating per-context qualification. Failure analysis shows that compression primarily surfaces as tool-execution and action-parsing errors. These findings recast ACC compression from token reduction into a runtime-reliability problem that must be evaluated through executable outcomes.

• Coordination Graph Memory for Structure-Aware Offline-to-Online Multi-Agent Reinforcement Learning

  Chenxing Li, Chin-Jui Chang, Yinlong Liu, Yichen Zhao, Cong Wang, Jan R. Seyler, Shahram Eivazi

  Abstract

  Offline-to-online multi-agent reinforcement learning reuses offline value estimates to accelerate online adaptation, but these estimates also depend on inter-agent coordination. When disturbances or task shifts occur during the online phase, this cooperative structure may be disrupted, potentially leading to catastrophic forgetting, especially after pretraining on suboptimal data. This paper proposes Graph Coordination-Aware Retention and Exploration with Offline Value Memory (GCARE-OVM), a graph-memory-based method for structure-aware offline-to-online multi-agent reinforcement learning. During offline training, it encodes cooperative patterns among agents into retrievable graph memories. During online adaptation, it adaptively regulates the retention strength of offline value memories and prioritizes exploration by agents that exhibit structural shifts. Experiments on the StarCraft Multi-Agent Challenge show that GCARE-OVM preserves competitive performance in same-environment transfer. Under coordination-related task shifts, including observation noise and communication dropout, it generally enables more stable online adaptation and improves the final win rate, the area under the online learning curve, or both, across multiple scenarios. By leveraging coordination graph memory, this work provides a new perspective on reliable offline-to-online multi-agent reinforcement learning.

• Correlated Forecast Errors and the Bullwhip Effect: Evidence from Time-Series Foundation Models in Decentralized Supply Chains

  Yunyang Zhi, Haoran Zhang, Yuhang Li

  Abstract

  Firms along supply chains increasingly use pretrained time-series foundation models (TSFMs) for demand forecasting. We study a systemic risk that local accuracy does not reveal: cross-echelon forecast-error correlation. In an order-up-to serial chain, we prove that upstream order variance increases with cross-echelon error correlation $\rho$ at fixed marginal accuracy. A $\rho$-dialable controlled probe then isolates this effect causally (top-echelon bullwhip 61% higher and system cost 17% higher at $\rho=1$ than at $\rho=0$). We next place Chronos-Bolt, Chronos-2, and TimesFM 2.5 in a four-echelon closed loop under homogeneous and heterogeneous assignments. Across the evaluated pool, these deployments exhibit persistently high conditional error correlation; moving from a homogeneous assignment to cross-family diversity changes it by only +0.01 on average in the structural-change scenarios. This observational fleet evidence measures correlation, not its cause: possible contributors include overlapping training data, architectures, objectives, training recipes, inductive biases, and shared or similar weights. Assignment-level cost differences are small and regime-dependent (-3.5% to +2.7%). Under the seasonal and M5 retail-demand regimes tested, homogeneous TSFM assignments also reverse the benefit of centralized information sharing (seasonal system cost +20.0%), an interaction effect that does not imply that information sharing is generally harmful. An empirical adoption game provides supporting evidence that operationally rational assignments can remain exposed to this correlation channel. Together, the results establish the mechanism and existence of the channel, not its prevalence or average magnitude in deployed supply chains.

• CreditTrace: Budgeted Counterfactual Replay for Root-Cause Attribution in Tool-Using Agent Graphs

  Taoda Zou, Jing Zhang, Zhiwen Zeng, Yaokai Guo

  Abstract

  When a tool-using agent workflow fails, the last anomalous step is often a downstream symptom rather than the step that introduced the error. Log-only failure attribution cannot resolve this ambiguity, while replaying every step is expensive. Given access to validated replacement outputs, we present CreditTrace, a training-free active diagnosis method for executable agent directed acyclic graphs (DAGs). CreditTrace initializes a causal-node posterior from local trace anomalies, models whether replaying a node should mask each candidate cause through post-dominance, and chooses the next trusted replay by expected information gain. A margin gate avoids intervention when local evidence is already decisive. We evaluate on a held-out benchmark of 2,880 executable workflows spanning four topology families, 12–48 nodes, and six tool, memory, routing, and communication faults. Dynamic routers make the declared DAG an imperfect causal model: structural predictions match 98.28% rather than all replay outcomes. With at most three replays, CreditTrace reaches 77.1% exact causal-node accuracy while using 1.43 replays on average, compared with 56.3% for anomaly-only attribution, 72.1% for a greedy posterior replay, and 69.5% for chronological replay with the same gate. At a five-replay budget it reaches 86.6% using 2.39 replays, 6.0 points above greedy replay. The benefit is largest for silent-value, schema, and stale-memory faults, while the margin gate resolves clear exceptions without replay. These results establish executable, budget-aware intervention as an effective basis for root-cause attribution in agent DAGs.

• CRSR: Complexity- and Risk-Aware Step-Level Model Routing for Multi-Step LLM Agents

  Yongzhuang Xiao

  Abstract

  Multi-step large-language-model (LLM) agents execute heterogeneous steps: some are shallow retrievals, whereas others integrate dependent evidence, satisfy structured output contracts, or prepare consequential tool actions. Query-level routing assigns one model to the whole request and cannot exploit this within-plan variation. We present CRSR, a transparent step-level router that selects a small, medium, or large worker before every executable step. It combines structural and semantic complexity with tool and consequence risk, applies auditable promotion rules, and records the decision evidence in the execution trace.

  We evaluate nine strategies on 248 DAG-executed tasks spanning 14 families, then repeat the comparison after replacing all three workers with a non-overlapping model pool. In Pools A/B, CRSR attains adjudicated pass rates of 0.702/0.738. Its strict-pass rates, which exclude the judge and deterministic relaxation, are 0.536/0.597, at CNY 0.0224/0.0254 per task. Relative to fixed-large, CRSR reduces mean cost by 35.5% in Pool A and 27.8% in Pool B; both paired reductions remain significant after Holm correction. Its observed pass rates are lower by 3.6 and 2.8 percentage points, respectively. Latency is 7.0% lower/82.4% higher than fixed-large, but 49.0%/44.4% lower than the repeated-attempt cascade. The 106 multi-solve tasks contribute 80.7% of routed nodes; CRSR reaches 69/106 and 79/106 pass while costing 34.8%/27.3% less than fixed-large. On 102 consequence-sensitive tasks, it reduces small-tier calls relative to complexity-only from 27 to 4 in Pool A and from 29 to 4 in Pool B, while shifting the error balance without eliminating conservative errors. Taken together, these results position CRSR as a practical, auditable cost-capability control layer that preserves competitive observed task success while substantially reducing expensive-worker use across two distinct pools under shared orchestration.

• DACE: Decoupled Adaptive Configuration Escalation for Agentic Image Generation

  Ruiqi Zhou, Qian'ang Mao

  Abstract

  Agentic image generation couples a multimodal planner-critic with a frozen diffusion model, but current agents mostly act through prompt refinement and leave the diffusion sampler fixed. We propose Decoupled Adaptive Configuration Escalation (DACE). This framework exposes sampler configuration as an agent action. DACE edits the generator's sampler configuration and uses per-constraint feedback to localize residual violations of user-specified requirements. DACE uses two-stage escalation: it first refines the prompt under the default sampler, then unlocks configuration edits for residual failures. We evaluate DACE on ParamBench, a benchmark of 6 fidelity-critical prompt families graded by user-stated constraints. On FLUX.1-dev, DACE raises instance success from 59.2% to 75.8% and constraint accuracy from 79.4% to 89.4% over direct generation from the user prompt. It also improves over prompt-only refinement, increasing instance success from 72.5% to 75.8% and constraint accuracy from 88.2% to 89.4%. The mechanism is generator-agnostic: ported unchanged to Stable Diffusion XL (SDXL), a weaker backbone, DACE reproduces the same direct-to-escalate ordering.

• DAGGuard: Budgeted Release-Gate Verification for Manager-Agent Workflows

  Juanna Zhai, Chao Tan, Kang Wu, Mengjiao Kai

  Abstract

  Manager agents decompose objectives into dependent tasks, so an erroneous upstream product can contaminate multiple descendants while uniform auditing eliminates much of the efficiency of delegation. We present DAGGuard, a budgeted release-gate framework that freezes versioned products and selects audits by conditional reduction in downstream harm and operational cost per unit budget. DAGGuard uses fast noisy-OR propagation when parent faults are approximately independent and routes shared-cause cases to correlation-aware sample-average planning. A dependency-evidence verifier grounds each verdict in the prerequisite versions actually consumed. Controlled experiments measure budget sensitivity, correlation error, model misspecification, and scalability. Human-adjudicated and cross-model evaluations quantify the recall--specificity--cost trade-off of richer evidence. At the primary budget, the complete policy reduces end-to-end loss relative to the strongest static baseline. Verification is therefore treated as a budgeted release-time decision rather than a uniform post-hoc check.

• Depth or Width? Series–Parallel Oversight of LLM Agents under Correlated Errors and Protocol-Dependent Cascades

  HikaruMatsuoka

  Abstract

  Overseeing an LLM agent increasingly means standing up a distributed system of other LLMs: parallel reviewers vote, and chains of critics delegate and verify. Under a fixed inference budget it is unclear which organizational geometry of this evaluator collective best prevents unsafe acceptance without over-rejecting useful outputs. We model oversight as a budget-matched series-parallel decision network with depth $D$, width $W$, threshold $k$, and composition order, whose evaluators have asymmetric false-accept/reject rates, correlated errors, and susceptibility to compromise. We prove three results. (i) Composition order is non-commutative: holding the joint law of primitive decisions fixed, wide-of-deep acceptance is contained in deep-of-wide, so wide-of-deep is uniformly more conservative (it lowers false and true accepts). (ii) Under an exchangeable beta-binomial model, majority vote has a strictly positive error floor, so width buys independent information, not agents. (iii) Under a per-stage exposure model, cascading compromise yields a finite optimal depth with a closed-form minimizer. A numerical design map partitions the (correlation, compromise) plane into committee-, deep-of-wide-, and wide-of-deep-optimal regions. On 1,920 real-LLM judgments over 80 screening items, model diversity cut false-accept correlation by a paired 0.66 but raised the marginal false-accept rate by 0.12, and no selector exceeded 40% top-1 across all 20 template-disjoint splits. A sequential-oversight pilot suggests protocol-dependent vertical failure ($0/12$ vs. $3/12$ injected accepts; underpowered). We thus recast oversight-architecture choice as organizational design for a distributed evaluator system.

• Diagnosing Memory-Unit Granularity in Long-Term Memory for LLM Agents

  Zhiqiang Mo, Qifeng Liang, Xiaohua Xu

  Abstract

  Long-horizon agents rely on persistent memory, yet memory-unit granularity is typically evaluated only within end-to-end architectures, conflating representation choice with extraction coverage, memory-bank composition, retrieval behavior, and evaluation criteria. We study these interactions through a controlled comparison of sentence-, belief-, event-, and summary-level memory units on 1,986 questions from multi-session LoCoMo dialogues, using two model backbones, each applied throughout memory extraction and answer generation under shared retrieval settings. Our evaluation decomposes memory performance into write availability, conditional retrieval, and answer utility. On LoCoMo, the answer-optimal granularity shifts from beliefs in the Qwen3-14B pipeline (0.410) to events in the DeepSeek-V4-Flash pipeline (0.557). A complementary diagnostic on a balanced LongMemEval subset shows the same qualitative pattern, with different evaluation targets favoring different representations. Summaries achieve the highest conditional support-unit recall in the induced LoCoMo banks, but this advantage narrows or reverses after accounting for write availability, bank composition, and evaluation target. These findings show that no single memory-unit granularity consistently dominates across extraction pipelines and evaluation criteria, and that effective memory design requires matching representations to system objectives.

• Distributed Best-Tool Identification under Credentialed Statistical-Record Custody

  Qifan Yang, Yujiao Peng

  Abstract

  Distributed tool selection can separate who compares tools, who samples them, and who may retain their statistical evidence. We formalize this separation through the CREST interface: three independent incidence relations govern decision access, sampling eligibility, and authorized one-copy record custody. An exact capacitated-Hall characterization gives the minimum residency depth, and a destructive fresh-batch learner is jointly δ-PAC with finite unconditional expected resources on every feasible finite cover. When all relevant records admit one-layer residency, a deadline merge of restarting nested trials and fresh escape sweeps yields arm-adaptive expected queries and an all-history exponential escape tail. We connect these learners to completion-driven service with unknown stopping times, sharp source-edge cover comparisons, and a full-alternative heterogeneous information benchmark. The evaluation contains 57,600 complete repetitions across 195 algorithm configurations. In prespecified synthetic Bernoulli comparisons, DMNF reduces mean queries by 44.79% relative to Fresh for K=32, gap 1/32, and q=1; FirstFree reduces logical service time by 27.89% relative to wave scheduling for d=2 and p=2. The evidence also exposes permission costs and cases where classical centralized methods are stronger. Explicit model boundaries, complete outcome accounting, reproduced execution traces, and rationally checked optimization certificates distinguish statistical, service, and state claims.

• Distributed-Utility Team Bandits: Learning When No Agent Observes the Team Reward

  Yujiao Peng, Qifan Yang

  Abstract

  We study cooperative bandit learning in which no agent observes the team reward. Pulling an arm gives two agents independent private outcomes, while a public, potentially nonseparable evaluator defines the latent team utility. Because neither agent sees the reward, communication must first construct a shared utility signal before the team can learn from it. We formalize this information structure with public actions, finite common utility certificates, and deterministic transcript accounting. Within this interface, a general precision-to-gap compiler converts distributed estimation into successive elimination, while a two-way estimation-resolution reduction connects offline communication complexity to genuine anytime target-anchor learning. For equality utility, combining signed heavy-light certificates, finite enumerative codes, and a permanent cache yields expected regret at most 77 sqrt(KT) with O-tilde(K^(2/3) T^(1/3)) worst-path communication. Joint-transcript lower bounds establish matching communication powers in a stated small-regret window, whereas at large regret budgets an exact two-bit interval exposes a distinct frontier. For known rational anchors, variance-sensitive policies adapt to the scale theta(1-theta). Reproducible synthetic experiments measure realized messages and regret, quantifying both the communication savings and the local computational costs of compact encoding, persistent caching, and endpoint adaptation.

• Does Synthetic Data Generation Strategy Affect Agent Memory Evaluation?

  Sebastian-Antonio Toma, Oliviu Matei

  Abstract

  Synthetic benchmarks are increasingly used to evaluate AI agent memory systems, yet while recent work has examined bias in synthetic test collections for information retrieval and RAG evaluation, we find none examining whether document generation strategy affects intrinsic evaluation of agent memory systems, where generated artifacts must encode temporal fact evolution across multi-document sequences. We ablate over 11 generation strategies spanning 5 axes (verification strictness, generator model size, style grounding, model diversity, and document realism), holding ground truth, scoring, and the system under test constant. Results show that generation strategy affects absolute scores, with up to 8.1 percentage points difference in fact accuracy across strategies on the same ground truth. Multi-model rotation nearly doubles corpus vocabulary diversity over single-model generation at moderate additional cost, reducing the stylistic homogeneity that can inflate scores. We provide actionable guidance for practitioners building synthetic benchmarks for agent memory evaluation.

• Does the Game Need the Science? Validity-First Evidence Audits for Finite Multi-Agent Science-Communication Games

  Zimo Zhang, Chunxuan Gao

  Abstract

  A science-communication game can present correct scientific content without actually requiring an agent to use it. The answer may still be recoverable from a default action or from an unintended clue such as packet presence, length, route, or timing. We introduce TruthFlow, an executable audit for finite, partially observable multi-agent games. It asks one narrow question: after neutralizing the registered scientific evidence and every registered, checked visible carrier derived from it within the declared finite model, while preserving the communication task, does one legal protocol and one total view-based policy still satisfy the action contract in every admitted world? TruthFlow searches again after removal and returns TRUE, FALSE, or INCONCLUSIVE, together with a replayable certificate. A failure-preserving heterogeneous Z3 study retained its initial 35/40 STOP; a newly frozen post-fix suite then achieved 40/40 exact agreement, with SAT witnesses replayed. For each of six safeguards, removing it produced the preregistered wrong conclusion on all four targeted cases. Two source-bound applications produced necessary, surviving-fallback, and invalid-removal outcomes. A 38-point CPU profile completed through 262,144 explicitly stored contract-correct policy witnesses, identifying an implementation-specific time and memory boundary rather than an intrinsic lower bound. TruthFlow therefore supplies a mechanism check before participant research; it does not measure learning, trust, engagement, behavior, or operational safety.

• Don't Trust Retrieved Evidence: Falsification Testing for Robust Retrieval Agents

  Shunyuan Zhou, Jiayi Han, Zekun Xu, Haiying Zhao

  Abstract

  Retrieval agents often incorporate noisy, conflicting, or adversarial passages into reasoning without first testing their reliability. Direct credibility judgments can be unreliable when fluent falsehoods appear plausible. We introduce FALCON, a training-free falsification controller that tests retrieved claims before downstream reasoning. For each passage, FALCON extracts atomic claims and assigns $K$ critics to generate targeted counterarguments for each claim. The controller retrieves counter-evidence and evaluates each counterargument's relevance and support. Claims whose fraction of successful rebuttals reaches a quorum $q$ are removed. Passages are reconstructed from surviving claims, then filtered and re-ranked using aggregated claim survival weights. FALCON integrates into ReAct-style agents without updating model parameters. Across three LLM backbones on TriviaQA and StrategyQA with injected random, conflicting, and adversarial distractors, FALCON improves answer accuracy and evidence precision over ReAct, ReAct+CRITIC, and ReAct+SC. Its accuracy advantage over ReAct increases with noise intensity. Ablations support the contributions of grounded counter-evidence, claim-level testing, and multiple critics. A retrieval-matched control indicates that additional retrieval alone does not explain the gains. On labeled evaluation data, surviving claims are substantially more likely to be correct than dropped claims, supporting survival as a useful but fallible reliability signal.

• DPI-CoT: Deductive Path Refinement for Instance-Aware Automated Chain-of-Thought Reasoning

  Mengyao Wang, Dian Huang, Junquan Gu, Shaorong Xie

  Abstract

  Large Language Models (LLMs) have achieved strong performance on complex reasoning tasks through Chain-of-Thought (CoT) prompting, yet automated CoT construction faces critical challenges. First, existing selection strategies struggle with an inherent trade-off between example diversity and instance-level specificity. Second, CoT quality control predominantly relies on answer-level correctness, ignoring intermediate logical flaws that propagate errors during few-shot inference. To address these limitations, we propose DPI-CoT, an automated CoT framework that elevates quality control to the step level and dynamically adapts prompt selection to individual test queries. Specifically, we first initialize the semantic space and partition the dataset via Semantic Space Partitioning (SSP) for diverse query coverage, which enables Deductive Path Verification (DPV) to generate both structurally diverse and step-wise faithful rationales, before Instance-Aware Dynamic Prompting (IADP) ultimately constructs adaptive prompts guided by semantic relevance and step-level logical consistency. Extensive experiments across seven reasoning benchmarks demonstrate that our framework outperforms the strongest evaluated automated CoT baselines by an average of 5.16 and 6.80 percentage points on Llama3-8B and Qwen2.5-7B, respectively.

• Dual-Channel Service Operations with AI Agents: Centralized Allocation versus Customer Self-Selection

  Ruiqi Zhou

  Abstract

  AI-enabled service systems can organize access to AI agent and human agent channels through two distinct mechanisms: centralized allocation by the system and customer self-selection. A queueing model compares these mechanisms for customers with heterogeneous task complexity and an AI agent success probability that decreases in complexity. Customers who try the AI agent channel first but experience failure are referred to a congestion-prone human agent channel. The analysis characterizes the socially optimal threshold under centralized allocation and the equilibrium threshold under customer self-selection. In both mechanisms, customers with task complexity below the relevant threshold try the AI agent channel first, whereas those above it go directly to the human agent channel. The analysis establishes that centralized allocation has a higher threshold than customer self-selection. Thus, whenever the congestion externality is active, customer self-selection leads to underutilization of the AI agent channel relative to the socially optimal allocation. The gap arises because each customer who chooses the AI agent channel imposes a positive externality by reducing congestion in the human agent queue, an externality that self-interested customers ignore. A Pigouvian price differential is shown to restore the first-best allocation, and the welfare loss from decentralized choice is quantified. Numerical experiments show that the welfare gap narrows as AI agent capability improves and widens as the demand rate increases. The results provide guidance on how a central service coordinator can manage customers’ decentralized choices to improve system-wide welfare in AI-enabled multi-agent service systems.

• Dynamics-Aware Federated Learning for Privacy-Aware Distributed Behavioral Risk Screening

  Muhua Tan, Zhaohui Yuan

  Abstract

  Privacy-sensitive behavioral time-series are natively distributed across personal devices and exhibit substantial user-level non-IID. In distributed artificial intelligence, each client can be viewed as a privacy-constrained learning agent that must collaborate without sharing raw data. Existing personalized federated learning methods usually partition shared and local components by network layers or generic feature subspaces, which does not directly match the dynamical distinction between cross-user stable rhythms and individual non-stationary perturbations. This paper proposes FedState, a dynamics-aware federated state-space framework in which the S4 shared rhythm stream participates in server aggregation, the Mamba local perturbation stream remains on clients, and cross-covariance regularization reduces second-order redundancy between the two representations. In user-independent evaluations on MentalTest and GLOBEM, FedState achieves Macro-F1 scores of 76.37% and 61.29%, outperforming FedAvg by 6.26 and 5.55 percentage points in Macro-F1, and by 7.25 and 6.70 percentage points in AUROC. User-level paired tests on MentalTest show Holm-corrected improvements over FedAvg, TMBL-FL, and FediOS. Ablations and offline representation separation analysis support the empirical roles of the dynamical division of labor and decoupling constraint. FedState is intended as a dynamics-aware principle for organizing federated sharing boundaries, rather than a claim of universal performance dominance.

• E2P: Effective and Efficient Planning for Multi-Agent Robots with Large Language Models

  Chenbing Wang, Shitong Luo, Sheng Zhou

  Abstract

  Large language models (LLMs), endowed with strong natural language understanding and commonsense reasoning capabilities, can effectively parse complex human instructions and perform autonomous task planning for multi-robot systems. However, existing approaches to long-horizon planning either require the LLM to generate a complete plan in a single pass—without accounting for environmental perturbations, which often leads to execution failure—or interact with the LLM at every step, improving success rates at the cost of substantial computational overhead and increased inference latency. To address these limitations, we propose E2P(Effective and Efficient Planner), a language model–based planner for long-horizon multi-agent systems. E2P follows a Plan-Adjustment–Act–Feedback framework. Instead of producing single step actions or a full global plan, the LLM generates localized action sequences, thereby reducing computational cost and inference latency. In addition, we introduce a learning-based dynamic adjustment module that refines these action sequences online, ensuring robustness under environmental disturbances. We conduct extensive experiments on the MAP-THOR and TDW-MAT benchmarks. The results demonstrate that E2P improves task success rates by up to 13 percentage points over state-of-the-art multi-agent planning frameworks, while substantially reducing computational cost and inference overhead.

• ECHO-Code: Execution-Coupled Hierarchical Orchestration for Reliable Project-Level Code Generation

  Danyang Zhang

  Abstract

  Project-level code generation must turn a natural-language specification into a coherent, executable multi-file project. Existing approaches often separate planning from execution feedback, use one policy for incompatible responsibilities, or detect integration failures only after most code is written. We present ECHO-Code (Execution-Coupled Hierarchical Orchestration for Code Generation), an execution-coupled multi-agent framework that (i) translates specifications into executable constraints, (ii) coordinates architect, developer, and tester roles through a shared semantic workspace, and (iii) uses execution, test, dependency, and specification signals to select iterative repairs. On the 500-specification ProjectBench-X benchmark, the reported results are 94.7% execution success, 91.3% test pass rate, 89.6% dependency resolution, and a 91.9% aggregate score. Component ablations and refinement analyses indicate that shared execution evidence and role specialization are both important. We also identify reproducibility, cost, and evaluation-validity limitations that must be addressed before broad deployment.

• Efficient Teammate Adaptation with Language-assisted Progressive Intention Alignment

  Zhichao Wu, Ruiqi Xue, Yi-Chen Li, Lei Yuan, Yang Yu, Jing-Wen Yang

  Abstract

  Collaborating effectively with diverse and previously unseen teammates is a cornerstone of multi-agent reinforcement learning (MARL). While existing research has made significant strides in adapting to diverse teammate behaviors under a fixed shared reward, the challenge of collaborating with partners who pursue distinct and unobserved personal rewards (intentions) remains unexplored. In these scenarios, the ego agent cannot rely on a pre-defined joint objective; instead, it must actively infer the teammate’s intention to provide aligned assistance. Traditionally, agents can deduce intentions from scratch using interaction histories, a process that is often inefficient because it lacks an informative starting point. In real-world collaboration, natural language serves as a direct channel to provide such an initial signal, yet robustly integrating these semantic signals with interaction histories remains a non-trivial challenge. To address this, we propose TALP, a principled Bayesian framework that facilitates rapid goal alignment by synergistically integrating semantic signals with low-level interaction histories. Specifically, TALP leverages LLMs to ground natural language into informative Bayesian priors, which are then dynamically refined via a distilled neural model that estimate likelihood from interaction histories. By employing an intention-agnostic policy for unbiased inference and an intention-aligned policy for cooperation, TALP enables rapid adaptation within a single episode. Extensive experiments in Grid-world and Overcooked environments demonstrate that TALP better fulfills the teammate’s intentions compared to state-of-the-art context-based meta-learning and teammate modeling baselines.

• Embodied Science Agents: Making Embodied Experimental Design Learnable

  Yihang Chen, Zhongwei Yu, Yang Li, Xingyu Lu, Hongxia Yu, Jun Wang

  Abstract

  Recent advances in large models have accelerated AI for Science toward agentic science, where AI agents autonomously conduct scientific discovery. Realising this vision requires closing the loop between computational reasoning and physical experimentation, positioning self-driving laboratories (SDLs) as critical enablers. However, existing SDLs exhibit a fundamental paradox: foundation model integration has increased linguistic sophistication while severing physical grounding. By treating models as disembodied planners that output symbolic workflows for separate robotic systems, current architectures inherently block the development of embodied experimental intuition. We are guided by a vision of embodied scientific intelligence in which autonomous discovery emerges through the co-evolution of three interdependent dimensions: a Digital Twin bridging theoretical models with simulation, a Cerebellum developing embodied skills from manipulation to procedural learning, and a Cerebrum advancing from automation to autonomous reasoning. Building on this vision, we introduce the Embodied Science Agent (ESA), which unifies digital intelligence with physical embodiment through three synergistic components: a Digital Agent for reasoning and strategy optimisation, an Embodied Agent for generalisable instrument control, and an Embodied Unified Interface for seamless coordination. We validated ESA in autonomous wastewater treatment experiments, navigating over 200 million possible conditions. Through 15 continuous rounds, ESA raised the final solution transmittance from 82% to 99.7%, corresponding to a 98.5% reduction in residual dye concentration and near-complete decolourisation, while demonstrating robust generalisation across diverse chemical operations, proving a transformative step toward next-generation SDLs that enable breakthrough scientific discoveries.

• EngineTailor:Minimal Complete Inference Engines for Agent-Driven Optimization

  Haiyan Qin, Rui Wan, Jinchen Han, Shulin Zeng, Guohao Dai, Yu Wang, Wang Kang

  Abstract

  Production inference frameworks such as vLLM and SGLang span models, devices, schedulers, and workloads. Their breadth creates a large, coupled optimization surface: retrieving editable fragments does not give a coding agent a coherent view of the complete execution path or a change's end-to-end effects. EngineTailor instead provides each model--device--workload target with a minimal but complete inference engine. It excludes unrelated serving functionality while retaining the target's complete execution and verification path. A Manager Agent turns measured bottlenecks into bounded tasks for fresh Worker Agents, while a protected acceptance plane admits only independently verified changes that pass correctness and satisfy a predeclared win or enable contract. Verified changes strengthen the current target engine, while reusable implementations and their applicability conditions accumulate in a shared substrate for re-qualification on later targets. We evaluate seven targets spanning autoregressive LLM, embodied VLA, and interactive world-model inference; four are edge-oriented. On a matched Qwen3-4B/B200 BF16 task under the same agent configuration, EngineTailor obtains verified prefill gains after matching token-generation latency, whereas five direct SGLang attempts yield no admitted improvement. On Jetson AGX Orin, the accepted strict-BF16 Pi0.5 product reduces end-to-end action latency by 43.0% relative to PhyAI (190.3 versus 333.9,ms). On B200, re-qualifying a fixed-contract graph boundary for Qwen raises concurrency-16 (c16) throughput from 3078.6 to 4654.8 tokens/s (51.2%). These results show that target-scoped engines can evolve without discarding reusable implementation knowledge. Source code and artifacts are available at https://github.com/qhy991/EngineTailor.

• Enhancing Communication Efficiency for FedLLM Training via Loss-Gap Weighted Router Aggregation

  Yi Hou, Chen Yang, Danyang Zheng, Fei Teng

  Abstract

  Federated large language model (FedLLM) offers a promising approach for continuously adapting on-device LLMs without centralizing raw data. However, practical deployment remains challenging due to two intertwined bottlenecks: the high communication overhead of transmitting massive model updates, and the substantial performance degradation that arises when models are constrained by limited on-device resources. Existing methods typically address only one of these challenges at a time. To overcome both, we propose FedMoR, a FedLLM framework tailored for communication-constrained settings that reduces communication cost while preserving model performance. FedMoR integrates a Mixture-of-Recursions (MoR) architecture to recursively share parameters, significantly shrinking the communication footprint per round. We further identify that naively aggregating MoR with FedAVG causes severe router shift, degrading global performance. To address this, we design Loss-Gap Weighted Router Aggregation (LGRA), a lightweight strategy that leverages the loss gap between the first and last local epochs as a proxy for update quality and performs weighted aggregation accordingly. Experiments demonstrate that FedMoR reduces communication volume by up to 87.5% while improving accuracy by up to 1.56X compared with baselines, achieving an improved trade-off between communication efficiency and model performance.

• Entropy Guidance and Exploration Rectification for Agentic Reinforcement Learning

  Haozhe Sun, Hengtong Lu, Bingwei Li, Zheng Yang, Huimin Ren, Xiaojie Wang, Caixia Yuan

  Abstract

  Reinforcement learning (RL) with step-wise credit assignment has shown promising potential for training LLM agents. Existing group-based methods typically construct comparison groups from fixed or overly local state representations, making long-horizon reasoning vulnerable to state aliasing, and discard isolated exploratory states because singleton-group baselines yield zero advantages. To address these limitations, we propose DEEPO, a policy optimization algorithm for agentic reinforcement learning. DEEPO couples Dynamic Entropy Guidance and Exploration Rectification: it uses action entropy to adaptively select context history for representing each state, and applies a query-relative baseline to assign advantages to isolated exploratory states that would otherwise be ignored. These designs offer three key advantages: (1) reducing state aliasing by adapting the entropy-guided context scope to each step, (2) improving exploration by assigning advantages to isolated but informative states, and (3) maintaining computational efficiency by relying only on readily available action-entropy proxy and lightweight baseline adjustment. Across multiple benchmarks, DEEPO achieves competitive or superior performance on most benchmarks and demonstrates notable out-of-domain generalization.

• EvoMAS: Learning Execution-Time Workflows for Multi-Agent System

  Chengdong Xu, Kaiqiang Ke, Ziheng Liu, Jiaqi Wei, Zibo Shao, Jungang Li, Weile Guo, Yiling Wu, Chao Yu

  Abstract

  Large language model (LLM)-based multi-agent systems have shown strong potential on complex tasks through agent specialization, tool use, and collaborative reasoning. However, most automated multi-agent system design methods still follow a one-shot paradigm: a workflow is optimized or selected before execution and then reused unchanged throughout the task. This static coordination strategy is ill-suited for long-horizon tasks whose subgoals, intermediate evidence, and information needs evolve over multiple execution stages. We propose EvoMAS, a framework for execution-time multi-agent workflow construction. EvoMAS formulates workflow construction as a meta-level sequential decision problem along a single task trajectory. At each stage, it constructs an explicit task state through a Planner--Evaluator--Updater pipeline and uses a learned Workflow Adapter to instantiate a stage-specific layered workflow from a fixed pool of candidate agents. The adapter is trained with policy gradients using sparse, verifiable terminal task success as the main supervision signal, while evaluator-based process reward is analyzed separately under very-hard sparse-reward settings. Experiments on GAIA, HLE, and DeepResearcher show that EvoMAS outperforms single-agent baselines and recent automated multi-agent workflow design methods. Our analyses further show that explicit task-state construction and learned workflow adaptation provide complementary benefits. Additional results indicate that process reward is most useful when terminal success is extremely sparse, and qualitative case studies illustrate that EvoMAS adapts agent coordination as the task state evolves.

• EvoSense: A Self-Evolving Multi-Agent System for Multi-Task Remote Sensing Interpretation

  Yimeng Wang, Ao Xu, Ziyou Guo, Tieru Wu

  Abstract

  Vision-language models (VLMs) have advanced the development of disaster-oriented remote sensing understanding. However, existing methods still face challenges in multi-task scenarios, including high training costs, fixed inference pipelines, and limited adaptability. We propose EvoSense, a training-free dynamic multi-agent framework for multi-task disaster remote sensing visual understanding. The framework dynamically constructs and optimizes the inference graph structure on a per-sample basis. Rather than relying on a fixed reasoning pipeline, it adaptively organizes the reasoning structure according to task requirements and sample-specific characteristics. Systematic evaluations on the DisasterM3_Bench benchmark demonstrate that EvoSense achieves competitive performance while effectively balancing task accuracy and computational cost.

• Exact Locality of Coalitional-Stability Auditing in Graphical Games

  Weihan Long, Yujiao Peng

  Abstract

  We study how much communication is needed to audit a given pure profile against joint deviations by coalitions of at most $k$ agents. The game is graphical on an undirected utility host graph $H$, while messages travel on a potentially different graph $G$. This separation exposes an information question that is absent from centralized strong-equilibrium verification: where can the utility tables and current actions needed to verify one coalition be brought together?

  For every fixed pair $(H,G)$, we give an exact class-worst-case characterization of deterministic LOCAL complexity. Every profitable coalition contains an $H$-connected profitable component. For such a component $C$, the necessary information support is $N_H[C]$ when current actions are private, and is $C$ when agents have trusted observations of the current actions in their utility neighborhoods. The audit complexity is precisely the maximum $G$-covering radius of these supports over all connected $C$ of size at most $k$. The lower bound holds with binary actions and full topology knowledge. When $H=G$, the two information models differ by at most one round; paths attain $\lfloor\min\{n,k+2\}/2\rfloor$ and $\lfloor\min\{n,k\}/2\rfloor$, respectively.

  Finally, unbounded LOCAL messages are essential. For $k=2$, binary actions, $\{0,1\}$ utilities, a matching utility graph, and a diameter-three communication graph, two LOCAL rounds suffice whereas standard CONGEST requires $\Theta(n/\log n)$ rounds. Thus propagation distance and cut bandwidth impose distinct barriers to auditing coalitional stability.

• Fair Dynamic Auctions via Equity-Constrained Pivot Mechanisms

  Meihong Zhang, Mengxiao Zhang, Bakh Khoussainov, Jiamou Liu

  Abstract

  We study repeated allocation of a scarce indivisible resource in dynamic auctions where strategic agents' private valuations evolve as a Markov process and the population is partitioned into exogenous groups. Although the dynamic pivot mechanism is efficient and satisfies periodic ex-post incentive compatibility and individual rationality, it can create persistent inter-group disparities when valuation processes differ structurally. We introduce \emph{$\epsilon$-periodic group fairness} ($\epsilon$-GF) to bound gaps in groups' expected discounted social welfare from every time step onward, and propose a fairness-aware mechanism that decouples equity control from truthful allocation: a state-dependent policy randomises group selection to satisfy $\epsilon$-GF, while an adapted pivot rule runs within the selected group. To limit manipulation of the equity controller, group selection follows a report-independent virtual-state trajectory. We provide an exact dynamic-programming solution for the equity-constrained policy and a scalable neural approximation, $\mathsf{FairPivotNet}$. Experiments show that our approach attains near-target fairness with modest welfare and revenue loss relative to the unconstrained pivot benchmark.

• Faithful, Not Corrective: Model Capability Governs Message-Format Effects in Multi-Hop Agent Relays

  Sicheng Zeng

  Abstract

  When LLM agents hand information to one another, does the message format matter? Two literatures disagree: format-optimization work reports that structured messages cut cost without hurting accuracy, while format-restriction studies find that imposing structure degrades generation. Neither line has measured what happens when messages traverse multiple hops, where copy fidelity, rather than one-shot generation quality, dominates. We introduce a controlled relay testbed in which briefs of twelve programmatic atomic facts are re-encoded hop by hop in five formats (free natural language, precision-instructed NL, JSON, triples, key-value) over six hops, scored against programmatic ground truth by a fixed strong grader, across two relay-capability tiers, a cognitive-load condition, and a paired-fork error injection. We find that (i) a strong relay is nearly lossless for every format (hop-6 QA recall ≥ 0.973), with residual loss concentrated at the first encoding step; (ii) per-hop cognitive load raises generation cost by 24–53% while fidelity changes stay within ±1.8 points; (iii) under a weak 1.5B relay, the across-format dispersion of hop-6 recall grows by a factor of 8.7 (CI 5.3–15.5), driven by an encode–drift trade-off that flips the format ranking in transit; and (iv) once an injected error is present, every format propagates it faithfully (surface persistence 83–100%) and no format cascades collateral damage onto neighboring facts. Structure buys a faithful, error-localizing channel, not an error-correcting code.

• Feasibility Propagation: Resolving Cost Credit Assignment in Almost Sure Constrained RL

  Chenyang Wu, Mingjun Cao, Zongzhang Zhang, Tianyu Wang, Zhilin Zhang, Chuan Yu, Jian Xu, Bo Zheng

  Abstract

  Reinforcement learning in safety-critical domains often requires satisfying constraints on every single deployment trajectory (almost sure constraints), rather than merely in expectation. Existing methods for almost sure constrained MDPs typically employ state augmentation combined with terminal penalties for constraint violations, creating sparse, uninformative learning signals. This paradigm suffers from a structural failure in cost credit assignment: a uniform penalty upon violation provides no gradient information regarding the magnitude of the failure, leaving the agent without a directional signal to reduce costs and recover feasibility. In this paper, we propose feasibility propagation, a framework that solves the cost credit assignment problem by learning a robust cost critic. Instead of relying on violation penalties, this critic explicitly estimates the worst-case accumulated discounted cost, providing dense, magnitude-aware gradients across the augmented state space. We prove that a finite-horizon margin-priority policy iteration scheme, which minimizes the robust cost when infeasible and maximizes reward otherwise, monotonically expands the feasible region and converges to the globally optimal policy. We introduce Solver-Guided Feasibility Propagation (SGFP). At the safety boundary, our method utilizes numerical solvers to compute precise Lagrangian multipliers that optimally balances cost versus value. Empirical results on continuous control benchmarks demonstrate that by rectifying the cost credit assignment failure during constraint violations, our approach not only achieves near-zero constraint violations but also yields significant gains in complex environments.

• Federated Diffusion Priors with Local Forward-Model Adaptation for Heterogeneous Snapshot Compressive Imaging

  Ge Wang, Xin Yuan

  Abstract

  Snapshot Compressive Imaging (SCI) reconstructs a spatial-temporal or spatial-spectral signal cube from a single coded measurement. Its reconstruction accuracy depends on both the learned signal prior and the fidelity of the sensing operator. In distributed settings, clients may have different data distributions and acquisition parameters, while direct data pooling is strictly restricted by privacy concerns. To address these competing challenges, we present FedCal-DiffSCI, a federated reconstruction framework that explicitly separates transferable diffusion-prior parameters from client-specific forward-model parameters. To prevent negative transfer, video and spectral SCI are handled by independent modality-specific federations. Within each federation, clients aggregate only the diffusion backbone while retaining their measurements, target cubes, conditioning adapters, and forward-model parameters locally. During inference, the proposed method iteratively alternates an HQS data-consistency update, diffusion posterior sampling, and projected local operator adaptation to seamlessly integrate the global generative prior with local physical constraints. Experiments on simulated measurements generated from hyperspectral histology images and surveillance videos evaluate the method under joint statistical and operator heterogeneity. FedCal-DiffSCI notably improves mean reconstruction metrics over existing distributed diffusion baselines across both video and spectral modalities, demonstrating the benefit of combining a shared diffusion prior with client-specific forward-model adaptation.

• FlockAct: From Local Proposals to Collective Progress in Multi-Agent Coordination

  Lingfeng Zhou, Shuaixing Chen, Jin Gao, Dequan Wang

  Abstract

  When agents share constrained resources, resolving the current conflict does not ensure that every member can finish its task. We present FlockAct, a closed-loop coordination system that connects local proposals to complete joint assignments and follows their consequences through execution. Each assignment covers every current group member, satisfies encoded resource checks, and passes as a whole to shared vehicle controllers. State feedback determines whether further coordination or ordinary recovery is needed. The selection policy is replaceable while construction and control remain fixed. A unified evaluation compares four systems in 720 runs across six fixed traffic configurations. Explicit coordination improves destination success over direct execution of the observed local proposals. Within FlockAct, rule-based and LLM-based selection both achieve 180/180 interaction successes without recorded vehicle collisions or off-road events, but all-vehicle destination success within the horizon differs: 180/180 versus 155/180. Interaction success therefore does not substitute for completion of the common task. Recorded paths connect this gap to residual conflicts, member release order, and time remaining for arrival. A separate diagnostic executes all listed candidates at seven states—15 branches—with deterministic proposals and a fixed feedback continuation. It distinguishes missed successful choices from states where no current candidate succeeds under that continuation. Together, these results support evaluating joint actions by both their current compatibility and the subsequent opportunity for every member to finish.

• Flow Policy Optimization with Distributional Value Guidance for Reinforcement Learning

  Ruijie Hao, Longfei Zhang, Xingxing Liang, Guangquan Cheng

  Abstract

  Reinforcement learning algorithms often rely on unimodal Gaussian policies and expected return optimization, which limits their ability to represent multimodal action distributions and exploit the rich structure of stochastic returns. Such limitations become particularly problematic in tasks with multiple optimal behaviors, where insufficient policy expressiveness and value guidance can hinder effective exploration and optimization. To address these challenges, we propose FP-DRL, a reinforcement learning framework that integrates a Flow Matching-based policy with a distributional critic. FP-DRL leverages the expressive capability of flow models to represent complex multimodal policies while retaining efficient action sampling, and employs quantile-based distributional reinforcement learning to capture the full return distribution rather than only its expectation, providing more informative guidance for policy improvement. Unlike existing generative policies that rely on supervised behavior modeling or diffusion-style objectives, FP-DRL directly optimizes the flow-based policy using reinforcement learning objectives within an off-policy actor-critic framework. Extensive experiments on MuJoCo continuous control benchmarks demonstrate that FP-DRL consistently achieves superior performance compared with strong baselines, including SAC, TD3, DSAC-T, DACER, and SAC-Flow, while exhibiting improved stability and multimodal representation capability.

• From Probe Scores to Alarm Policies: Operational Validity of Activation Monitors for Language-Model Agents

  Xueping Gao

  Abstract

  Activation probes can predict safety-relevant properties of language models with high area under the receiver-operating-characteristic curve (AUROC), but deployed agent monitors make thresholded alarm decisions under tight false-alarm budgets. These are different estimands. We introduce an Operational Validity Contract that fixes a monitor's target, observability, identity, timing, intervention unit, comparator, calibration, and cost. We formalize risk at the semantic request or trajectory level: when one task contains repeated alarm opportunities, row-level AUROC and false positive rate do not identify semantic-unit any-alarm risk. A confidence-certified threshold also requires enough independent negative units, a tie-safe rule, and transport to deployment. Across Models Under Pressure, LASR refusal prediction, immutable AgentDojo, and a prospectively protocol-frozen ST-WebAgentBench replication, joint activation--observable monitors reach AUROC 0.957 and 0.935 on the first two benchmarks, yet their locked 5%/10% detection rates are only .642/.742 and .719/.782, respectively. On AgentDojo, the secondary mean-activation rollout monitor reaches AUROC 0.922 but detects none of 38 positive semantic cases at the locked 5% operating point; thresholds intended for 10% false alarms realize 18.7--20.0% on test. Because the test misses its prospectively frozen 40-positive support gate, we label it support-insufficient. On ST-WebAgentBench, activation reaches AUROC .874, but 23 independent calibration negatives cannot identify even a 10% controller; the locked policy abstains rather than reporting its mechanical zero FPR as a success. An exploratory counterexample also lowers full AUROC while improving realized 10% utility. The fail-closed compiler caps MUP and LASR at restricted predictive value and AgentDojo and ST-Web at representation accessibility; no setting reaches alarm-policy validity. A reproducibility audit recomputes these claim levels and all four settings' reported score-derived metrics from frozen predictions and identities.

• Ground, Diagnose, and Refine: A Hierarchical Perception Agent for Aerial Visual Grounding

  Qin Liu, Hao Yu, Qize Guo, Chao Yang, Chenyang Wang

  Abstract

  Aerial visual grounding must localize language-referred objects in small, dense, and visually repetitive aerial scenes. We present \method, a hierarchical perception agent that combines an explicitness-aware one-pass grounder with budgeted test-time diagnosis and refinement. The grounder introduces Adaptive Query Clarification and Query-Tuned Spatial Anchoring (QTSA), which injects query-conditioned structural and semantic anchor tokens into Qwen2.5-VL-7B. A parent controller parses a query constraint graph, diagnoses the initial box using inference-time evidence, dispatches Target and crop-based Zoom units under a hard visual-call budget, and fuses candidates with a false-repair guard. Context and relation constraints are evaluated deterministically, yielding auditable Accept, Refine, and Escalate outcomes. On DVGBench's implicit UAV protocol, QTSA raises mIoU from 0.1930 for the task-adapted Qwen2.5-VL-7B baseline to 0.3653, and guarded refinement further reaches 0.3747, improving the matched one-pass result by 0.93 points. On RSVG-HR's direct-answer protocol, QTSA reaches 0.3915 mIoU versus 0.2993 for Qwen2.5-VL-7B; the explicit Agent produces 0.3870 mIoU with zero parse failures. These results show how query clarification, expert-aligned spatial anchoring, and evidence-guided re-observation support complementary grounding regimes under bounded computation.

• GSR-Agent: Graph-Guided Solver Repair for Inference-Time Self-Improvement of Reinforcement Learning Schedulers

  Zixiao Luo, Dandan Liu, Yi Chai

  Abstract

  Reinforcement learning schedulers construct feasible solutions quickly, but their one-shot decisions can leave coupled machine assignments and queue orders that cannot be improved through an isolated move. Global exact re-optimization can address these dependencies, but is often too expensive under a short inference budget. We present GSR-Agent, a post-policy self-improvement framework that leaves a pretrained scheduler frozen. Given its incumbent schedule, GSR-Agent builds a dynamic critical-conflict graph, uses a repair-gain critic to prioritize a graph-coherent neighborhood, and invokes warm-started CP-SAT to revise machine assignments and operation orders jointly. An independent decoder accepts a candidate only when it is feasible and strictly improves the incumbent. On 72 protocol-frozen flexible job-shop instances, after averaging each instance across three CP-SAT seeds, GSR-Agent improves HGNN-PPO schedules by 3.553%, compared with 2.627% for the matched Critical LNS baseline. Its mean paired advantage is 0.966 percentage points ($p=1.04\times10^{-5}$). Experiments on 130 public FJSPLib instances provide additional transfer evidence. Ablations support the roles of learned graph selection and progressive expansion, while scale analysis identifies CP-SAT as the bottleneck at 30- and 40-job scales under a one-second budget. These results establish a verified interface through which a frozen RL agent can use structured critique and bounded optimization tools to improve its decisions at inference time.

• Guard Before You Remember: Receiver-Guided Cached Prompt Relay for Decentralized Continual Learning

  Ziming Wang, Jianrong Wang, Dengcheng Hu, Qi Li, Shiwen Lu

  Abstract

  In decentralized continual learning, agents learn evolving private task streams under costly model exchange and intermittent contacts that separate when useful knowledge is produced, encountered, and locally needed. Prompt tuning offers an established parameter-efficient approach by representing task-specific knowledge as compact trainable visual parameters attached to a frozen model. Store-carry-forward caching complements this representation by retaining and relaying task modules beyond the contacts in which they were produced. However, caching can also deliver stale or multi-source prompt states to an agent whose task history and local evidence differ from those of their senders. Existing mechanisms therefore improve the efficiency and reachability of knowledge transfer, but do not determine whether a delayed proposal should modify the receiver's persistent task memory.

  We introduce Receiver-Guided Cached Prompt Relay (RGCPR), a server-free protocol that separates four states: reachable, proposed, authorized, and persisted. A byte-bounded least-recently-used (LRU) cache retrieves task-matched packets; the receiver forms a count-weighted convex proposal anchored by its local prompt, then checks active-task gain and cross-context compatibility before committing it. Failed proposals use bounded mass-to-local repair, returning removed remote weight to the local anchor while preserving convexity and avoiding amplification of retained remote states.

  Across two datasets, two budgets, two heterogeneity levels, and three seeds, RGCPR has the highest mean Average Incremental Accuracy (AIA) among decentralized methods in all eight settings. Mechanism analysis identifies exchange and caching as the main utility sources, while repair recovers rejected proposals. Same-state replay shows that receiver authorization prevents damaging writebacks under stale and unreliable knowledge and improves post-event target accuracy.

• HERO: Hypervolume-guided Exploration and Utility Gap Rescue Optimization for Multi-Objective Reinforcement Learning

  Hao Tang, Kai Xu, Guanxun Zhao, Zhicheng Wang, Gang Wei

  Abstract

  Multi-objective reinforcement learning (MORL) aims to learn a set of policies that represent diverse trade-offs among conflicting objectives. Recent Pareto front discovery methods improve the efficiency of obtaining trade-off solutions by explicitly searching for diverse Pareto policies. However, under limited interaction budgets, existing approaches often allocate optimization resources uniformly, ignoring that different preference policies and exploration directions may have different potentials for improving the Pareto front. Therefore, adaptively allocating optimization resources according to the potential contributions of different policies and exploration directions is essential for efficient Pareto front discovery. In this paper, we propose HERO (Hypervolume-guided Exploration and Utility-Gap Rescue Optimization), a budget-efficient framework that adaptively allocates optimization resources during both stages of Pareto discovery. During initialization, Utility-Gap Rescue identifies under-optimized preference policies through utility gaps and transfers knowledge from nearby high-quality solutions. During extension, Hypervolume-guided Exploration evaluates candidate directions based on their potential hypervolume improvement and adaptively adjusts exploration strength for Pareto expansion. Experiments on five multi-objective continuous-control benchmarks show that HERO achieves higher final hypervolume and expected utility than the evaluated MORL baselines under matched, benchmark-specific interaction budgets.

• HypeTree: Neuro-Symbolic Decision Fusion and Claim-Level Verification for Explainable Poker Agents

  Yuxiang Zhang, Heng Song, Junwu Zhu

  Abstract

  Strategic agents operating in imperfect-information games must choose legal actions while producing explanations consistent with game rules and quantitative state variables. Existing LLM-based agents often misidentify hands, draws, or pot odds, whereas unconstrained neural policies provide limited guarantees of factual consistency. We present HypeTree, a neuro-symbolic framework for explainable heads-up No-Limit Texas Hold’em. HypeTree combines typed Symbolic Logic Units that compute hand, draw, pot-odds, and legal-action facts with a graph neural Action-Policy Neural Intuition Unit through a validation-selected log-linear fusion layer. A claim-level Generate Verify-Refine procedure checks structured explanation claims against symbolic facts and performs bounded targeted repair before rendering text. We further construct PokerBench-E, a leakage-controlled benchmark of 9,998 post-flop states with board-group-disjoint splits and rule-verifiable factual explanations. On the 1,029-state held-out test split, HypeTree achieves 63.27% action accuracy and 61.85% macro-F1, compared with 60.84% and 56.72% for the NIU-only model, and obtains the best scores among nine evaluated methods. Across 12 machine-verifiable explanation claim types, full GVR raises the pass rate from 12.73% for generate-only to 100.00%, using deterministic fallback on 3.50% of samples. Together, these results indicate improved class-balanced benchmark agreement from neuro-symbolic fusion, while the accuracy gain over NIU-only remains directional. Claim-level verification with deterministic fallback enforces schema-level consistency over the 12 predefined explanation claim types.

• Identifiability Before Replay: Agent Workflow Recovery under Dependency Loss

  Guoqing Zhang

  Abstract

  Replay answers how to re-execute a compromised agent workflow. It does not determine what semantic result remains safe when dependency evidence is incomplete. Recovery is identifiable when a nonempty set of attacker-compatible worlds has a nonempty intersection of safe semantic classes. This criterion is distinct from checkpoint, version, and replay-plan identity. Validated evidence refines that set. Provenance-Guided Escalating Replay (PGER) then instantiates runtime recovery. It escalates from local replay through wider scopes and Full Restart to rejection without enumerating compatible worlds online. Across 75 fixed multi-world tasks, four ordinary evidence channels leave 20 unresolved. An optional validated pre-fault binding resolves all 75 under its trust assumption, and 54 tasks require at least two channels. Exact enumeration and a finite conflict-world solver agree on all tasks. In a frozen two-model realization diagnostic, only one of 200 configurations achieves Strict Safe Completion. The sole success occurs in the Clean non-recovery control, and none of 59 verifier-accepted model-authored PGER outputs satisfies recovery adequacy. We interpret this probe as a diagnostic rather than a causal comparison. At the default normalized cost setting, PGER has lower cost on partial-scope tasks. A LangGraph case shows lower aggregate resource use across five pairs retained after post-hoc strict rescoring. Neither result establishes a universal advantage over Full Restart. These findings motivate separating identifiability, evidence refinement, PGER instantiation, and realization.

• Improving Collaboration in Multi-Agent Reinforcement Learning via Limited Egoistic Altruism Preference

  Guang Yang, Tianpei Yang, Qixin Long, Xingguo Chen, Yang Gao

  Abstract

  Social dilemmas, characterized by the conflict between individual rationality and collective welfare, are ubiquitous in mixed-motive games. While recent advances in social preferences have shown promise in preventing multi-agent reinforcement learning (MARL) from converging to sub-optimal equilibria, existing approaches face a critical trade-off: preference-homogeneous methods (e.g., focusing solely on fairness) often lack sufficient incentives for robust collaboration, whereas preference-heterogeneous models suffer from prohibitive search costs and poor scalability. To address these challenges, we propose a limited egoistic-altruism framework. By distilling complex heterogeneous preferences into a tractable egoism-altruism dualism, our method facilitates the efficient learning of collaborative policies. Theoretical analysis using Schelling diagrams, complemented by extensive empirical evaluations, demonstrates that our approach effectively mitigates sequential social dilemmas and significantly enhances cooperation among agents.

• Incentives for Client-Diversity Caps in Proof-of-Stake Networks

  Manvir Schneider

  Abstract

  When a new node implementation (client) lowers operating costs, proof-of-stake operators have an incentive to adopt it. If most stake uses this client, a single software defect can disrupt the network. We study incentives for limiting each client's share of total stake to a chosen cap. Each operator assigns its indivisible stake amount to one client and pays a fixed operating cost. The cap is at least the share per client when total stake is divided equally among all available clients. We compare stake-share bonuses and penalties with charges for each operator's contribution to concentration above the cap. These charges guarantee a pure Nash equilibrium, where no operator can gain by switching clients alone, and convergence when operators make profitable switches one at a time. A quadratic version measures concentration using squared excess shares above the cap. Each client's equilibrium share is bounded by the cap plus the largest operator's stake share and an allowance for operating costs that decreases as pricing strength increases. Simulations illustrate the effects of incentives and indivisible stake.

• JudgeOrbit: Calibrating LLM Judges with Functional-Equivalence Orbits

  Yujie Shen

  Abstract

  Distributed and agentic AI systems increasingly delegate evaluation to language-model judges, yet those judges are usually assessed only by aggregate agreement on one presentation of each candidate. This leaves a structural risk: a judge can appear accurate while changing its verdict after a behavior-preserving rewrite, or remain unchanged after a tested behavior is broken. We introduce JudgeOrbit, a training-free calibration protocol that evaluates a judge on functional-equivalence orbits and paired behavior-changing mutants. JudgeOrbit reports invariance error, causal sensitivity, position-swap consistency, and selective accuracy under disagreement-triggered refusal. Across five open-source judges, mean invariance error is 18.7% on 100 test-validated HumanEval tasks and 17.8% on 100 MBPP tasks. On 100 static WebDevJudge pairs, comment invariance error is 19.2% and swap inconsistency reaches 49.0%. A browser oracle retains 48 fully validated pairs, on which comment and formatting invariance errors are 13.3% and 12.9%. Requiring orbit unanimity raises selective accuracy by about one point but reduces coverage by 13--14 points. A 200-item blinded synthetic reviewer audit yields Fleiss' k=0.068, revealing low agreement about observed violations but providing no human ground truth. The results show that functional invariance is an independent calibration axis and that evaluator instability should be exposed, rather than hidden inside a single benchmark score, before a judge controls an agentic workflow.

• Learning to Observe-then-Predict: Efficient Multi-agent Collaboration via Latent Dynamics Learning

  Zhaohan Feng, Zhijie Liu, Lei Yuan, Gang Wang, Yang Yu

  Abstract

  Improving data efficiency remains a fundamental challenge in multi-agent reinforcement learning. Existing approaches often train decentralized policies end-to-end using shared reward signals without explicitly exploiting the structure of agents' observations, which can lead to inefficient and suboptimal policy learning. Inspired by the observer--predictor paradigm in control systems, we argue that effective observation encoders provide an important foundation for policy optimization. In particular, learning environment dynamics can encourage encoders to extract decision-relevant features. We propose a MAmba model for multi-agent self-supeRVised rEpresentation Learning (MARVEL) that constructs global latent states by aggregating locally encoded observations and models their evolution through action-conditioned recurrent prediction. This latent dynamics objective provides a temporally structured self-supervised signal that guides the observation encoders toward dynamics-aware representations for downstream policy learning. The recurrent latent model is designed to limit error accumulation over multi-step prediction horizons. Experiments on challenging SMAC and SMACv2 tasks show that augmenting HAPPO and MAPPO with MARVEL generally improves their sample efficiency and task performance and yields competitive results against several strong MARL baselines.

• LESM: Reliability-Gated Latent Episodic Strategy Memory for Cooperative Multi-Agent Reinforcement Learning

  Yuran Gong, Jinqiao Wang

  Abstract

  Cooperative multi-agent reinforcement learning under partial observability often suffers from redundant exploration, as agents repeatedly rediscover coordination patterns that have already succeeded. Existing episodic-memory approaches typically exploit past experience through value estimation or reward-level learning signals, but provide limited support for directly reusing successful coordination strategies within decentralized recurrent decision making. We introduce LESM, a reliability-gated latent episodic strategy memory framework for cooperative multi-agent reinforcement learning. LESM infers latent situations from each agent's local observations and recurrent context, and compresses trajectory windows from qualified episodes into latent strategies consisting of discounted action preferences and normalized hidden-state summaries. During decision making, strategies associated with similar inferred situations are retrieved, aggregated, projected into the recurrent feature space, and selectively integrated before the GRU update through a Reliability Gate conditioned on local decision information and retrieval statistics. Across six SMAC tasks and two GRF scenarios, LESM achieves the best late-stage mean performance on six of eight tasks and ranks first overall with an average rank of $1.62$. Ablation studies show that episodic strategy reuse accelerates learning and that gated integration is particularly important on GRF; additional analyses reveal task-dependent strategy integration behaviors and indicate that LESM maintains wall-clock training efficiency comparable to QMIX.

• LLM-MANDy: Agentic Discovery of Sparse Tensor-Product Representations for Multidimensional Nonlinear Dynamics

  Junchen Wang, Rongye Shi, Haoheng Wu, Bingrun Chen, Tianchen Zhu, Wenjun Wu

  Abstract

  Inferring the governing equations of dynamical systems is a fundamental challenge across a wide range of scientific fields. Sparse regression methods are efficient and interpretable, but their performance depends strongly on predefined candidate libraries, whose size grows rapidly when interactions among multiple state variables are considered. To address this limitation, we introduce LLM-MANDy, a single-Agent framework for iterative coordinate-wise basis discovery. The proposed basis functions are tensorized to construct a shared multivariate candidate library, from which MANDy-based candidate pre-screening and equation-wise sparse support search identify candidate supports. The final support is selected using the complexity-aware criterion $\operatorname{BIC}_{C}$, and its coefficients are refitted from trajectory data. Numerical evaluations of the equations fitted from the sparsified multivariate library are fed back to the Agent, guiding the iterative refinement of the basis libraries. Under the low-separation-rank assumption, LLM-MANDy recasts multivariate library design as coordinate-wise basis discovery, thereby mitigating the combinatorial search burden of explicit multivariate library design. Experiments on 68 systems with two to four state variables demonstrate the effectiveness of LLM-MANDy against both LLM-based and conventional baselines.

• MAC-Tree：Task-Adaptive Graph-Constrained Multi-Agent Search for Multi-Hop Question Answering

  Zhaoliang Zhang, Bo Tian, Shuai Lei, Yuan Huang

  Abstract

  Multi-hop question answering requires large language models to connect distributed evidence while satisfying question-specific entity, temporal, causal, comparative, and negation constraints. Existing tree-based methods mainly search over free-form natural-language states, while graph-enhanced methods commonly use graphs for retrieval or evidence storage; these lines of work typically do not expose question-specific constraints as a shared control interface across expansion, evaluation, and pruning. We propose MAC-Tree, a task-adaptive graph-constrained multi-agent search method. MAC-Tree compiles each question into a Task-Adaptive Event Logic Graph (TELG), a typed control interface that couples constrained candidate expansion, shared-state role evaluation, and adaptive pruning. TELG represents entities, events, relations, temporal and causal links, and conditional constraints, remains fixed during a single search, and guides candidate generation, path evaluation, and branch retention. Available evidence is incorporated during subsequent path expansion and evidence grounding. Four role-specialized evaluators independently assess a shared pool of candidate states for task progress, evidence support, logical consistency, and answer sufficiency. We evaluate MAC-Tree on a fixed benchmark of 900 instances evenly distributed across StrategyQA, HotpotQA, and MuSiQue, with 300 instances from a public split of each dataset, covering implicit judgment, cross-document multi-hop question answering, and compositional multi-hop reasoning. We further construct a 519-instance cross-dataset disagreement subset from frozen baseline outputs to analyze mechanism behavior when baseline predictions diverge. MAC-Tree achieves 84.00% accuracy on StrategyQA, 62.67% exact match (EM) and 75.97% F1 on HotpotQA, and 38.67% EM and 55.18% F1 on MuSiQue. Under the fixed evaluation subsets and experimental configuration used in this study, the results support the use of task-adaptive graph constraints and role-based shared-state evaluation for more targeted multi-hop reasoning search.

• MARA: Flow-Matching-Guided Multi-Agent Resource Allocation for Computational Resource Efficient Learning

  Hanye Zhao, Muning Wen, Yong Yu, Weinan Zhang

  Abstract

  Allocating limited computation among concurrent learning tasks is difficult when each task must reach a target loss before a deadline but its required training effort is unknown. Existing approaches combine online loss prediction with adaptive resource allocation, yet commonly treat computation as continuously divisible throughput. We instead study a practical setting in which tasks arrive over time and computation is provided by discrete nodes. This setting introduces both uncertain demand and constrained sequential decisions. We propose MARA, which predicts future loss trajectories with conditional flow matching and coordinates compute nodes through a cooperative multi-agent autoregressive policy. A potential-based progress reward supplies intermediate training feedback while preserving the undiscounted task-completion objective. Across in-distribution, reinforcement-learning, and vision workloads, flow matching reduces remaining-resource prediction error relative to weighted least squares. At the scheduler's training load, MARA completes 63.46% of tasks on average, 8.54 percentage points above strong baseline Learning with Adaptive Resource Allocation (LARA), and remains ahead under unseen heavier workloads.

• MARFT: Multi-Agent Reinforcement Fine-Tuning

  Junwei Liao, Muning Wen, Jun Wang, Weinan Zhang

  Abstract

  The rapid rise of LLM-based agents has led to LLM-based Multi-Agent Systems (LaMAS), which show strong potential in complex collaborative tasks. While reinforcement learning is well established for improving LLM agents, its success has focused largely on single-agent settings. Applying multi-agent reinforcement learning to LaMAS remains less developed because traditional MARL assumptions do not directly match LaMAS decision processes, which can involve dependency-triggered actions, heterogeneous roles, changing organization, and long autoregressive outputs. To characterize these challenges, we formalize LaMAS optimization as a Flexible Markov Game (Flex-MG), which represents structured and potentially varying dependencies among agents, and introduce Multi-Agent Reinforcement Fine-Tuning (MARFT). Our MARFT instantiation combines sequential action generation and predecessor-conditioned PPO objectives, a centralized critic, parameter-efficient adapters, and token-level credit assignment derived from action-level multi-agent advantage decomposition. Across fixed two- to four-agent math and coding workflows, both MARFT variants report higher mean evaluation scores than IPPO in all 24 settings, while the variant without parameter sharing shows the clearest advantage in final training reward. Role-specification ablations further examine how prescribed roles affect optimization. We conclude by identifying dynamic workflows, larger-scale evaluation, and system-level reliability as important future directions.

• Meta-CBPO: Meta Contrastive Bilevel Prompt Optimization with Repair–Regression Evidence

  Yi Jiang, Yongli Wang

  Abstract

  Automatic prompt optimization adapts black-box models through natural language, but reusable system prompts must separate behavior shared across related tasks from task-specific user-prompt changes. MetaSPO frames this setting as bilevel optimization, but its outer loop receives only residual errors after local adaptation. This signal conflates persistent failures with adaptation-induced regressions and discards successful repairs. Meta-CBPO introduces Repair–Regression Evidence (RRE), which records seed-to-adapted correctness transitions on the same examples. Evidence-Aware Adaptation (EAA) retries local adaptation with expanded residual-example context when transition support is sparse, while Prompt-Pool Reranking (PPR) reranks user prompts separately under each system-prompt candidate. Under the reported source-optimization configuration, Meta-CBPO raises the unweighted average of task-native scores from 39.5±2.2 to 43.3±1.2 across 14 held-out tasks in five families and obtains higher three-seed means than MetaSPO on 12 tasks. On Review at seed 42, Meta-CBPO uses 1.66× as many base-model and 1.57× as many optimizer-model requests as MetaSPO. On Review, reference-seed diagnostics rank Meta-CBPO first across all tested base models, optimizers, and source-set sizes; its system prompts are shorter than MetaSPO's in all five families.

• More Papers More Slots: AI-Enabled Sybil Attacks in Peer Review

  Zhicheng Du, Zihe Wang

  Abstract

  Generative AI can reduce the effort required for research and writing, making high-volume submission strategies increasingly feasible. We study the vulnerability created by this shift: under proportional acceptance, an additional paper can enlarge the screened pool and create more acceptance slots. We formalize this strategic behavior of authors as a Sybil attack and develop a Bayesian game in which heterogeneous authors allocate a fixed effort budget across a bounded number of papers. We show how lower production costs can make attacks feasible and characterize an optimal attack for a rejected author in a two-paper best-response setting. For every finite portfolio menu, a mixed Bayes--Nash equilibrium exists; in a certain binary-type benchmark, the unique symmetric pure equilibrium is separating when productive authors are sufficiently rare: high types attack, whereas low types concentrate effort. Using equilibrium enumeration, we study how review-mechanism parameters shape attack incentives and conference outcomes. In the computed participatory symmetric pure equilibria, attacks arise in a localized, nonmonotone region. Counterfactual comparisons show that attack portfolios can increase review load and reduce accepted-paper quality.

• Motif-Risk-Aware Distributed Coordination for Fragmentation-Resilient Multi-Agent Systems

  Xie Guangqiang, Ruikai Chen, Sicheng Zhao, Yang Li, Xiaohui Li

  Abstract

  Distributed multi-agent consensus under distance-based communication is prone to topology fragmentation when explicit connectivity-preserving constraints are absent. A key cause is density-biased local interaction: densely populated directions exert excessive aggregate influence, while sparse but structurally critical connections are underweighted, with their fragility further shaped by heterogeneous relay redundancy. To address this issue, this paper proposes Motif-Risk-Aware Distributed Coordination (MRDC), which selects representative supports to suppress redundant density contributions and uses relay-aware motif risks to adapt coordination weights according to local structural fragility. Without constructing explicit connectivity-preserving regions, MRDC mitigates fragmentation through local structural adaptation. Simulations across diverse topology and density settings provide empirical evidence of fragmentation-risk mitigation and competitive convergence-step performance under the tested configurations.

• Multi-modality Stealthy Backdoor Attack on Medical Visual-Language Model under Federated Pre-training

  Zhengkai Yan, Shiting Wen, Dongang Wang, Fanyi Wang, Haotian Hu, Zhiwang Zhang

  Abstract

  Medical Visual-Language Models, empowered by large-scale medical data pre-training, have shown great potential in enhancing clinical procedures. To safeguard data privacy, federated learning is commonly employed during the pre-training process. However, due to limited cybersecurity measures in many medical institutions, some clients are vulnerable to hijacking (e.g., backdoor attacks) during the federated learning pre-training phase. To reveal this matter, in this work, we propose a novel backdoor attack on medical visual-language models during the federated pre-training phase, which causes the models to generate erroneous diagnostic results during the testing stage. This work pioneers an underexplored area, as prior backdoor attacks have not addressed federated pre-training settings in medical visual-language models. To further improve stealthiness, we introduce a Multi-modality Stealthy Backdoor Attack (MSBA) framework that leverages both visual and textual triggers. This approach ensures minimal alteration of medical images and textual prompts while activating the backdoor triggers. We conduct extensive experiments on the MIMIC-CXR dataset to validate our method in the context of medical visual-language model pre-training. Our approach achieves high attack success rates and superior stealthiness compared to vanilla methods. The proposed method is further evaluated in the task of generating medical reports, demonstrating its generalizability. The code will be released after a safety review upon acceptance.

• NegoMerge: Resolving Silent Semantic Merge Conflicts via Execution-Gated Multi-Agent Negotiation

  Liran Wang, Zhoujun Li

  Abstract

  Version control decides whether a merge "succeeded" at the text layer. When two branches modify interdependent but textually disjoint code, three-way merge completes cleanly, yet the merged program silently violates a crosscutting assumption and fails in continuous integration: a silent semantic conflict. Merge tools cannot see such conflicts, and learned merge resolution targets conflicts the toolchain has already reported, scored by agreement with the historical human resolution, so neither its task interface nor its metrics transfer. Repair-style agents pointed at the failing merged state can reach a passing suite by reverting one branch's changes, which destroys intent instead of reconciling it. We first formalize silent semantic conflicts and their witness-preserving resolution, with a success criterion defined entirely by execution: a resolution must pass the existing tests and each branch's supplied witness test, a test that pins the branch's behavioral increment and makes silently discarding it a measurable failure. We then build $\textbf{SilentMerge-Bench}$, to our knowledge the first executable benchmark for the resolution (not detection) of silent semantic conflicts. Its fully automatic inject-and-filter pipeline plants divergent edit pairs, guided by eight interference patterns distilled from the empirical literature, into fixed containerized environments of six real open-source projects, and admits an instance only when execution verifies every defining condition, with no human annotation; because the injected conflicts are new constructs, no historical resolution exists for a model to have memorized. Finally we present $\textbf{NegoMerge}$, a multi-agent framework in which two branch advocates defend the intents of the two branches through structured briefs and conditional concessions under a bounded negotiation protocol, an arbiter distills a conflict diagnosis and synthesizes a mediating patch, and a sandboxed execution gate validates every patch, feeding failure reports back into the next round as evidence. Across 64 instances built from eight verified interference mechanisms in six real Python projects, evaluated with one frozen run per configuration, NegoMerge reaches a 93.8% resolution rate on each of two model families, Grok 4.3 and DeepSeek-V3.2, while a single-agent repair loop given the same context and execution gate degrades from 87.5% to 75.0% across the same pair, with the gains concentrated in the contract-level interference patterns. On the weaker family (DeepSeek-V3.2), 15.6% of the repair loop's outcomes are suite-green yet fail a witness test, silently discarding one branch's intent, versus 3.1% for NegoMerge; the difficulty of the task lies in preserving intent rather than in reaching a passing suite.

• Order-Preserving Trajectory Reuse with Selective Replanning for Lifelong Multi-Agent Path Finding

  Xueyuan Wen, Zhoutao Zhang, Zhixin Zhang, Chengzhen Kuang, Lei Shang, Zhen Wang

  Abstract

  Lifelong Multi-Agent Path Finding (LMAPF) requires agents to continuously generate collision-free trajectories for new goals under a finite planning budget. To address online planning time pressure, concurrent planning and execution frameworks---like Planning and Improving while Executing with Delay Probabilities (PIE-D)---are leading solutions. However, current PIE-D instantiations still exhibit inherent limitations. The default approach replans all agents from scratch without reusing spatiotemporal structures across episodes, while reactive priority-based strategies lack lookahead. Additionally, traditional ''Simulate then Replan'' strategies combined with complete MCP execution inherit obsolete waits, decreasing throughput. To address this, we propose a novel planning strategy, Order-Preserving Trajectory Reuse with Selective Replanning (OPTR-SR), and couple it with a wait-filtered MCP executor within the PIE-D framework. By preserving original visit orders to advance retained trajectories and selectively replanning only affected agents, OPTR-SR reuses spatiotemporal structures while avoiding obsolete wait accumulation. Evaluated against the default PIE-D and enhanced PIBT (E-PIBT) across five maps and 25 configurations, OPTR-SR achieves higher throughput than both baselines in most settings, while substantially reducing memory usage in most settings.

• Owl-Gaze: A Zero-Shot PTZ Tracker with Tightly-Coupled Perception and Control

  Zhikang Jiang, Lihua Zheng, Yang Shi, Yaohua Wang

  Abstract

  Pan-Tilt-Zoom (PTZ) cameras are widely deployed for active target following, yet existing tracking pipelines remain tightly coupled with closed-set, pre-trained detectors that offer no generalization to unseen categories without costly retraining. The Segment Anything Model (SAM) provides an escape through class-agnostic segmentation, but integrating this single-frame foundation model into a real-time active camera loop introduces critical challenges---the absence of temporal associations, the disruption of spatial correspondences under continuous gimbal rotation, and the computational burden that hinders direct edge deployment. To bridge these gaps, we propose a tightly-coupled perception-and-control framework that combines NanoSAM's zero-shot perception with reinforcement learning-based active camera control. Our core insight, Physical Attention, treats gimbal rotation as a physical analog of an attention operation: by continuously re-centering the target, the controller reduces the need for dense full-frame search under our virtual PTZ protocol. Supported by ego-motion-compensated Hint Box Propagation and LSTM-driven blind prediction during occlusion, the system maintains continuous tracking despite detection loss. Extensive evaluations under simulated PTZ dynamics with real-world video content demonstrate that distilling dense visual data into compact geometric states enables robust, low-latency, zero-shot cross-domain generalization, supporting a perception-control co-design paradigm for active vision on edge devices.

• ParaCiteSlide: Paragraph Evidence Binding for Agentic Presentation Generation

  Shuaiwei Zhang, Yuhan Lin, Zhongtian Hua, Yi Luo, Yingjie Han, Lei Li

  Abstract

  Generating presentations from long documents requires content compression, visual organization, and traceability from factual statements to their sources. Existing systems address content selection, planning, template adaptation, and rendering, but attribution is often attached at slide level or reconstructed after generation. These designs do not necessarily preserve, as part of the generation state, which source paragraphs conditioned each title, bullet, or caption. We present ParaCiteSlide, a stateful approach that makes paragraph evidence binding and slot provenance explicit operational objects in agentic presentation generation. Given a source document and a PPTX template, ParaCiteSlide constructs typed factual slots and maintains shared input, plan, evidence, content, quality, and rendering states. For each slot, it binds source paragraphs before generation, records native provenance with the generated text, and checks both structural provenance validity and semantic support before editable rendering. When either check fails, bounded recovery corrects provenance references, rereads source evidence, and regenerates the affected slot text and provenance before revalidation. Final factual support is evaluated separately through a unified post-hoc protocol that ignores method-supplied provenance. Under the unified post-hoc protocol on the DOC2PPT-derived benchmark, ParaCiteSlide records higher PSR and lower UCR than PPTAgent across the tested model configurations, while comparisons with KCTV remain model dependent. The method preserves competitive presentation quality, exposes native paragraph provenance unavailable from the compared baselines, and returns editable decks for all inputs in the evaluated configurations. These findings illustrate how generation-time provenance can be maintained and inspected independently of external factuality and presentation-quality evaluation.

• Parallelism or Concession? Concurrency-Aware Procurement Negotiation for Agentic Commerce

  Xiaolin Xu, Donghao Zhu

  Abstract

  Agentic buyers can cheaply fork a procurement task into many parallel negotiations, but concurrency is not free: every thread consumes resources, and simultaneous agreements create cancellation and commitment risk. We study a one-unit post-order sourcing problem with a single hard-deadline negotiation window, in which a planner jointly chooses the number of seller-facing negotiators and a common procurement price cap. The model combines a product-specific acceptance curve with fulfillment loss, per-thread cost, and excess-commitment cost. We establish three structural results. First, holding the per-thread acceptance target fixed, the marginal value of another negotiator decays geometrically, yielding a conditional concurrency threshold. Second, under a convex quantile curve, parallelism substitutes for concession: more concurrent negotiators imply a weakly lower per-thread acceptance target and price cap. Third, when prices are more dispersed, Agentic buyers benefit by searching harder for bargains, but suffer when they instead try to guarantee procurement by offering higher prices. We operationalize these results in the Concurrency-Aware Negotiation Optimizer ($\text{Cano}$), a deterministic optimizer that jointly determines the optimal negotiation concurrency and procurement price cap for an agentic procurement system. Across different analytic market configurations and extensive Monte Carlo, finite-data, non-Gaussian, and correlated-seller stress tests, $\text{Cano}$ consistently outperforms common heuristic policies while validating the predicted structural properties.

• PerfectDan: High-Performance AI for GuanDan from End-to-End Reinforcement Learning

  Caixin Cai, Chao Yu, Zijian Fang, Yucong Zhang

  Abstract

  GuanDan is a complex four-player partnership card game characterised by imperfect information and cooperative–competitive play. Due to partial observability, one player-view observation corresponds to multiple underlying states that differ in how unseen cards are distributed among the other players. The resulting ambiguity makes reliable value estimation from local observations a central difficulty for self-play reinforcement learning. We propose perfect-information guided self-play (PGSP) to address this problem. PGSP trains a perfect-information response (PR) policy on global states to estimate action values and guides a deployable main response (MR) policy trained on local observations through an adaptive stop-gradient loss. To support scalable training and consistent evaluation, PerfectDan is then presented as an end-to-end GuanDan artificial intelligence (AI) framework that integrates a PGSP-based algorithmic core with GuanDanPlatform, an open-source environment, distributed runtime, and evaluation component. The framework combines multidimensional state–action representations and temporal modelling with dual observation modes and inspectable trajectories. In the reported evaluation, PerfectDan wins 95% against DanZero and at least 97% against each of eight rule-based agents, ranking first at 2318 Elo. PerfectDan achieves the best performance among the publicly available agents evaluated in this study. This project is publicly available at https://github.com/Choysang/rlcard-guandan.

• PhyWorld: Unified Semantic and Physical Reasoning for Physics-Native 3D Scene Generation

  Chao Shen, Yunfan Zhou, Zhihai Wang, Xijun Li, Jianguo Yao, Haibing Guan

  Abstract

  Embodied Artificial Intelligence (EAI) requires robust spatio-physical grounding to operate effectively in the physical world. However, existing scene synthesis paradigms predominantly follow an "outside-in" methodology that prioritizes semantic plausibility over physical fidelity. This design leads to (1) inadequate treatment of implicit physical violations— such as kinematic obstructions and unstable equilibria— and (2) a poor balance between semantic consistency and physical constraints, thereby limiting coherent alignment between these two objectives. To address these limitations, we introduce PhyWorld, a programmatic scene generation pipeline based on a physics-native "inside-out" design philosophy. Specifically, PhyWorld formulates scene construction as a sequence of executable programs and employs a code agent to translate high-level design intent into low-level scene manipulation and physical reasoning. The framework combines code-driven layout calibration with rigorous runtime physical verification, thereby casting scene generation as a rejection sampling process governed by deterministic physical checks. In addition to synthesizing high-fidelity scenes with stronger alignment between semantic plausibility and physical realism, PhyWorld automatically produces dense physical annotations. Experimental results demonstrate that PhyWorld achieves a consistently better balance between semantic and physical requirements than state-of-the-art baselines at scale, while also providing substantial improvements in downstream spatio-physical tasks.

• R-DOCP: Conserving Contract-Defined Responsibility in Multi-Agent Tool Use

  Yikai Zhang, Guojunling, Qirui Chen, Dong Jin, Quan Zheng, Shuangwu Chen

  Abstract

  Call-level guards can validate only actions represented in an executable plan. If a required external effect never enters that plan, every observed call may be valid while the request remains incomplete. We study finite, closed-world requests for which a reviewed typed operator derives the required effect frontier from an authenticated snapshot. We introduce the Refinement-aware Distributed Obligation Custody Protocol (R-DOCP). A candidate manifest specifies the proposed effects, and a deterministic checker certifies it only when its members, source coverage, dependencies, and bound versions exactly match the required frontier. After certification, atomic expansion creates one identity-stable responsibility leaf per effect. During execution, leaves may change custodian and lifecycle state but not semantic identity or multiplicity; the request closes only after every leaf is discharged by admissible evidence or explicitly reported as residual. With full typed checking, no nonexact candidate is admitted in the controlled audit. The reference implementation also satisfies every tested lifecycle and closure invariant under injected structural faults, adversarial transitions, and interrupted settlement. Local replay shows that independent effects can still execute in parallel, and all integrated-path multi-process trials reach accountable closure without duplicate effects or unresolved outcomes. These results support contract-relative safety within the supported scope; unsupported cases fail closed, and unrestricted intent grounding remains outside our claim.

• RAAC: A State-Aware Retrieval-Aided Adversarial Commentary Attack on Discussion-Based LLM Multi-Agent Systems

  Yuhao Zhao, Cheng Xie

  Abstract

  Large Language Model-based Multi-Agent Systems (LLM-MAS), where multiple LLM agents interact and collaborate to solve complex tasks, have shown great potential across a wide range of applications. However, such collaborative systems may contain compromised or untrusted agents that can exploit legitimate interaction channels to mislead benign agents and distort collective decisions. Although recent studies have examined adversarial agents in multi-agent collaboration, existing attacks typically pursue a fixed target answer throughout the discussion and rely primarily on predefined prompts to produce adversarial content. They do not fully exploit the evolving discussion state, nor do they draw on prior successful persuasion cases. In this work, we propose Retrieval-Aided Adversarial Commentary Attack (RAAC), an adaptive attack framework that exploits the evolving discussion state and prior successful persuasion cases. RAAC first employs a State-Aware Wrong-Cluster Targeting Strategy to identify an effective incorrect-answer cluster and determine which agents should be supported or opposed. It then retrieves relevant successful persuasion cases from the Adversarial Persuasion Corpus as references to guide the generation of targeted adversarial comments for the agents identified for opposition. Experiments across four downstream tasks and two LLM families show that RAAC consistently degrades the downstream task performance of discussion-based LLM-MAS and causes larger accuracy reductions than those caused by the prompt-based MCA. In particular, on MedMCQA, RAAC reduces accuracy by 32.79% relative to the Clean MAS. These findings expose a critical vulnerability in discussion-based LLM-MAS: feedback intended to improve collective reasoning can instead be exploited by an internal adversarial agent to progressively influence benign agents' answer updates and ultimately induce incorrect system-level decisions.

• Rank Without an Oracle: Deviation-Aware Interaction-Rank Selection from Offline Multi-Agent Logs

  Xiangwu Wang, Chengwei Cao, Hongyuan Tang

  Abstract

  Offline multi-agent payoff models are estimated under a logging distribution but used on distributions induced by learned solutions and unilateral deviations. Standard held-out loss can therefore favor an interaction class that predicts logged play well while distorting strategic incentives. We introduce Selective Interaction-Rank Validation (SIRV) for finite games with known logging distributions. A training split fits nested payoff models and constructs a common union of all candidate deployment and unilateral-replacement distributions; an independent calibration split evaluates every candidate on this same union. SIRV returns the smallest rank whose simultaneous upper worst-target risk is within tolerance of the best upper score, and abstains when a declared target is unsupported or too imprecisely estimated. A common coverage event yields a finite-candidate target-risk bound and a candidate-specific coarse correlated equilibrium (CCE) gap certificate. We also isolate an exact two-point off-support non-identifiability result. In a controlled factorial study with 2,048 independent games per family, empirical-Bernstein bounds reduce the median CCE-gap certificate by 42.5% relative to Hoeffding bounds on common returns, with a 1.36-point reduction in supported return. Under paired rank misspecification and in a separately generated congestion family, the SIRV-EB fallback rule lowers mean true candidate-selection CCE regret relative to ID-Mean, while retaining game-level losses. Across 384 games at $N=3,5,8$, ID-Mean-relative mean CCE-regret effects stay positive while certified return falls sharply under weak coverage. These results separate certifiable model selection from universal strategic improvement.

• Reachability-Certified Subteam Decomposition for Locally Interacting Multi-Agent MDPs

  Xiangwu Wang, Chengwei Cao, Hongyuan Tang

  Abstract

  Persistent communication limits force a multi-agent system to decide which agents may coordinate throughout a rollout. Current proximity alone is insufficient: separated agents may interact later, whereas a large pair reward may remain unreachable until it is heavily discounted. We introduce Reachability-Certified Subteam Decomposition (RCSD) for finite multi-agent Markov decision processes with factorized physical dynamics, finite-range ordered pair rewards, and almost-sure motion bounds. RCSD combines a speed-limit lower bound on pairwise contact time with a reward envelope to form a current-state affinity. For any capacity-valid persistent partition, the sum of cut affinities bounds the reward-deletion error of every unchanged stationary Markov state-feedback policy. A product of team-optimal policies for the resulting cut MDP incurs at most twice this certificate in regret against the centralized optimum. Both bounds are worst-case tight. On a controlled five-agent family, RCSD-Exact reduces aggregate normalized execution regret by 56.0%, 28.8%, and 25.3% relative to uniform, distance-only, and envelope-only partitions. A separate stochastic two-dimensional study finds no bound violation over 384 exact-partition and 1,440 restricted-controller evaluations. Exact four-agent evidence favors RCSD over uniform and distance-only grouping; raw evidence for current contact is borderline and envelope-only is unresolved. Across balanced 8--20-agent strata, controller-library utility is mixed: pointwise paired intervals favor RCSD over distance and current contact, include zero for uniform, and favor envelope-only and Value-MIP over RCSD. Partition construction remains subsecond in median up to 100 agents; this last result does not include affinity formation or MDP planning.

• REIN: Revision Engineering Infrastructure for Non-Code Assets

  Liran Wang, Zhoujun Li

  Abstract

  Engineering work often happens inside specialized editors whose outputs, including game levels, CAD parts, and video projects, are versioned as opaque binary files. AI agents can operate these editors through structured tools, but existing asset-aware mechanisms provide specialized merge or live-collaboration paths rather than a common branch workflow spanning editor-exposed asset types with per-commit provenance and machine verification. We present $\textbf{REIN}$, an agent harness that supplies that workflow over two layers: an execution layer of structured editor operations, and a harness layer that confirms intent, logs every operation, and records each commit as intent, operation script, semantic diff, and verification evidence. On this record $\textbf{REIN}$ merges at the intent level: a read/write/effect analysis of the recorded operations decides compatibility, both sides replay onto the common ancestor in an isolated workspace, and validators gate the result; cases outside the analysis degrade to a conflict report or a branch-aware lock rather than a silent merge. We further build $\textbf{REIN}$-Bench, to our knowledge the first benchmark jointly evaluating commit-message generation, change question answering, and operation-replay merging for opaque binary assets, pairing scripted instances that carry machine-checkable ground truth with a challenge set designed independently of $\textbf{REIN}$'s rules, over two Unreal Engine 5 projects with a human-adjudicated oracle. Counting every input against the system, $\textbf{REIN}$ decides the full scripted merge set correctly, resolves nearly all challenge inputs, and commits no erroneous merge in either set; ablations attribute the scripted merges to the deterministic rule engine and the unlogged-edit recoveries to the agent. On the understanding tasks, the on-demand semantic diff is the operative mechanism, raising commit-message coverage and question-answering recall well above tool-only agents at a fraction of their token cost. Because $\textbf{REIN}$ assumes only a structured editor interface, instantiating it for another creative tool replaces the execution layer alone, a design property whose evaluation beyond Unreal Engine 5 remains future work.

• Resource-Conditioned Epoch Allocation for Tail-Reliable Personalized Federated Remote Sensing

  Chengfei Xu, Yunrui Jiang

  Abstract

  Federated remote-sensing systems operate across clients with different data distributions and model capacities. A uniform personalization budget ignores this resource asymmetry and often leaves narrow models behind. We study resource-conditioned effort allocation after shared federated pretraining. Our method, resource-adaptive epoch budgeting (RAEB), assigns a fixed local epoch budget to each resource class while keeping client adapters and classifiers private. Across 60 paired runs on FedRS and EuroSAT with TinyCNN and ResNet18 backbones, two non-IID maps, two participation rates, and five seeds, RAEB improves client mean, worst-10%, and tiny-client accuracy over uniform two-epoch adaptation by 1.85, 3.88, and 5.05 percentage points, respectively. All three endpoints improve in all 12 condition cells. The gain is concentrated among resource-constrained clients. Estimated training FLOPs remain similar at 1.007 times, while sample passes and wall time increase to 1.293 times and 1.468 times. These results characterize a tail-oriented accuracy–cost operating point and show how its value changes with client capacity and population composition.

• Revenue Maximization in Diffusion Auctions via Reduced-Form Oracles

  Wenhao Xue, Bin Li, Dengji Zhao

  Abstract

  Designing revenue-maximizing diffusion auctions, in which buyers strategically propagate auction information over a social network, remains an open problem even for a single item. We study Bayesian diffusion auctions with independent finite-support type distributions and extend the Cai, Daskalakis, and Weinberg (CDW) reduced-form framework by incorporating diffusion feasibility into the feasible interim allocation space. Within this framework, the feasible interim allocation rules form a convex polytope whose extreme points correspond to simple virtual VCG rules. This characterization reduces revenue maximization to a linear program over interim allocations and payments, with virtual welfare maximization providing the basis for feasibility separation and implementation. For arbitrary networks, we provide a sampling-based algorithm that runs in polynomial time and, with high probability, returns an approximately revenue-maximizing auction. The algorithm approximates the feasible reduced-form polytope via empirical sampling, thereby enabling approximate feasibility separation and decomposition. These results extend the CDW reduction to diffusion auctions and yield a polynomial-time approximation scheme for revenue maximization on general networks.

• Revocation Is Not Complete at Receipt: An Executable Outcome Contract for Distributed Agent-Tool Gateways

  Lifei Liu, Haoran Yu, Danping Zhang

  Abstract

  Distributed agent workflows cache delegated authority at tool gate- ways. Revocation may be reported at issuer commit, transport ac- knowledgment, or receiver persistence, although none of these ob- servations establishes that all gateways enforced the change or that guarded effects stopped. We define C0–C5, a bounded outcome contract spanning issuer commit through effect prevention, and implement it as a persis- tent monitor over a frozen gateway set, call set, and observation window. A 14-case fault matrix checks conservative classification, sticky violations, restart persistence, and evidence immutability. In a derived audit of its eight semantic cases, treating C3 or C4 as completion produces seven or five false positives, respectively; the bounded C0–C5 verdict produces none. Descriptive scaling covers 1,024 gateways and 10,000 effect vectors. The primary experiment uses the maintained Keycloak 26.7 identity server and measures the contract across two durable receivers, eight physical agent-tool gateways, Filesystem and PostgreSQL effects, and 16 AWS hosts in three availability zones. Normal delivery and recovery commit no effects. Under receiver- selective non-delivery, transport acknowledgment and receiver persistence still succeed, but 112 of 225 calls (49.78%) are admitted and all 112 effects commit in every treatment block. A separately provisioned execution repeats the result. OPA/Toxiproxy and a live three-role LLM workflow with networked MCP tools reproduce the persisted-but-not-enforced boundary on narrower surfaces. Fi- nally, an unmodified external MCPMark agent executes 60 closed- loop task arms. Normal C4 denies all 31 exposed post-barrier calls and commits no effect; selective C3 allows 344 calls and commits 145 effects across Filesystem and PostgreSQL. Transport success is therefore evidence of progress, not a revocation-completion verdict.

• Rewarding the Heroes: Hierarchical Gradient-Based Shapley for Fair Federated Valuation

  Zehua Cheng, Wei Dai, Zhipeng Wang, Jiahao Sun

  Abstract

  Federated Learning (FL) enables collaborative model training across decentralized data silos, yet remains structurally fragile because no scalable mechanism exists to weigh each participant's true contribution-an essential prerequisite for fairness, robustness, and data-quality assurance. The Shapley Value (SV) is the canonical answer in cooperative game theory, but its exact computation is prohibitive (O(2")), and existing approximations all collapse one of two ways: sampling-based methods buy fidelity at the price of thousands of forward passes, while heuristic methods buy speed at the price of misranking the very "Hero" clients FL most needs to retain. We propose Hierarchical Gradient-based Shapley (HGS-Shapley), a framework that breaks this dichotomy by decoupling the valuation problem into a divide-and-conquer pipeline. HGS-Shapley projects high-dimensional gradients into a compact subspace via Johnson-Lindenstrauss transforms, clusters clients into metaagents by gradient affinity, computes a coarse-grained Shapley value over these clusters via stratified permutation sampling, and redistributes each cluster's budget to its members through a magnitude-aware Gated Softmax allocation that exponentially amplifies aligned Heroes while strictly nullifying adversaries. The hierarchical design provably reduces per-round complexity from O(2") to O(Nd + N1.5, dominated in practice by the linear projection term. Across four cross-modal benchmarks (CIFAR- 10, FEMNIST, CheXpert, StackOverflow), HGS-Shapley matches the fidelity of expensive Monte Carlo oracles (p>0.93) at up to three orders of magnitude lower wall-clock cost—a 1200x speedup on CheXpert-and reliably identifies free-riders, labelflipping attackers, backdoors, and white-box adaptive adversaries (Adversary AUROC ≥ 0.92) where standard robust aggregation fails.

• RGRL-PID: Reference-Guided Residual Reinforcement Learning with PID for Task-Constrained UAV Laser Evasion

  Jiaxian Liu, Mingkai Li, Xingyu Duan, Ying Zhang

  Abstract

  Parallel advances in unmanned aerial vehicles (UAVs) and counter-UAV directed-energy systems make survivability under predictive laser tracking a timely security problem. Target-side evasion supports authorized-UAV resilience and adversarial evaluation of counter-UAV trackers. Unlike terminal interception, laser evasion couples predictive tracking with cumulative thermal damage. We study a UAV crossing a bounded corridor that rules out range-avoidance detours against a predictive opponent emitting continuously in range. The UAV observes motion, goal state, and laser warning, but not the opponent's tracker or beam direction. We propose RGRL-PID, which assigns feasibility to classical feedback and opponent uncertainty to learning: a reference and proportional-integral-derivative (PID) controller form a closed-loop mission backbone, a fixed maneuver supplies partial survivability, and residual reinforcement learning (RL) adapts it to parameter variation within the modeled fire-control family. The RL policy modulates timing, lateral strength, and vertical motion; zero output recovers the base controller. In simulation, on 60 previously unused, paired test scenarios spanning six tracker/latency settings, RGRL-PID succeeds in 86.7% of missions, versus 65.0% for Tuned Open-Loop Maneuver and 16.7% for the zero-residual Base-Maneuver. Mechanism analysis shows that the learned residual improves evasion at two scales: it coordinates threat-region traversal with non-repeating 3-D maneuvers that increase prediction error and slow thermal accumulation, while the PID backbone preserves goal progress. On six additional sensor-noise/pointing-rate settings unused for training or selection, RGRL-PID succeeds in 53/60 missions (88.3%), versus 31/60 (51.7%) for Tuned Open-Loop Maneuver and 11/60 (18.3%) for Base-Maneuver, providing further evidence that learned residual adaptation improves performance under previously unused parameter settings within the same opponent family.

• RMCP-Distill:Residual-Aware Multi-Candidate Preference Distillation

  Zihao Wang, Lin Li, Wei Wei, Lijun Zhang

  Abstract

  Knowledge distillation transfers knowledge from a large teacher model to a small student model. Preference distillation often relies on isolated pairwise comparisons, which describe local preference relations but do not directly represent preference strengths over an entire candidate set. Multi-candidate preference distillation provides richer supervision, but existing methods often construct preference targets from teacher evaluations without adapting them to the evolving student state across training rounds. This may repeatedly optimize preference relations that the student model has already learned. To address these challenges, we propose Residual-Aware Multi-Candidate Preference Distillation (RMCP-Distill), a novel multi-round preference distillation method grounded in residual curriculum learning and adaptive target interpolation. RMCP-Distill employs two complementary components: (1) a residual curriculum that stratifies training examples based on the ratio of cross-entropy losses computed by the teacher and anchor models, with the proportions of difficulty groups adjusted under a fixed data budget; and (2) an adaptive target that uses the divergence between the preference distributions of the teacher model and the student model from the previous round to determine how strongly each training target is adjusted toward the teacher model’s preferences. This design focuses training on preference relations that remain insufficiently learned while preserving previously acquired preferences. Experiments across three model families and five downstream tasks show that RMCP-Distill achieves the highest average score in each family.

• RoboGuardian: Fault-Aware Execution and Memory Maintenance for Embodied Agent Harness Frameworks

  Zijie Gao, Chenxu Li, Zixuan Chen, Jiapeng Xu, Wenquan Zhang, Yizhou Zhu, Rongjin Zheng, Xinke Wang, Jieqi Shi, Jing Huo, Yang Gao

  Abstract

  Embodied-agent harnesses connect high-level planners to low-level execution policies and use execution results to update task state and memory. However, current harness mechanisms often fail to verify physical state changes reliably. This limitation applies to both low-level execution feedback and VLM-based evaluations. Unverified outcomes may therefore enter memory and mislead subsequent actions. In this paper, we present RoboGuardian, a fault-aware execution and memory-maintenance framework for embodied-agent harnesses. It organizes execution exceptions into three response-oriented categories: outdated plan steps, uncertain outcomes, and execution failures. This categorization enables robust decision-making from multi-modal evidence. RoboGuardian checks plan validity and entity identity before execution and combines semantic, proprioceptive, and effect-level evidence to produce a three-way physical-effect judgment where only physically confirmed actions update its embodied memory and progress. It also maintains a bounded trusted context and re-observes before retrying unresolved actions, preventing blind retry and replay. We evaluate RoboGuardian across diverse simulation environments, including robotic arm manipulation on LIBERO-10, dexterous grasping on Clutter-Bench, and continuous vision-language navigation in Habitat. Furthermore, we deploy the system in the real world on both dual-arm and dexterous-hand manipulation platforms. Experimental results demonstrate that RoboGuardian effectively enhances execution reliability across diverse tasks and hardware embodiments. The framework maintains a lightweight and effective memory mechanism while introducing minimal computational overhead during execution. Videos and open-source code can be found on our project website: https://tsuki-gor.github.io/RoboGuardian/.

• Robust Adversarial Reinforcement Learning with Risk Sensitivity and Critic Consistency Regularization

  Jiaxi Wu, Tiantian Zhang, Yuxing Wang, Yongzhe Chang, Xueqian Wang

  Abstract

  Reinforcement learning (RL) achieves strong performance in sequential decision-making but remains brittle under dynamic uncertainty and distributional shifts. Robust Adversarial Reinforcement Learning (RARL) improves robustness via worst-case perturbations, but existing approaches frequently suffer from unstable optimization and degraded value estimation. In particular, overly aggressive adversaries can drive the agent toward uninformative failure states, while adversarial perturbations amplify disagreement between double critics and introduce biased value targets. We propose a unified framework, RACER (Risk-sensitive robust Adversarial critic ConsistEncy-regularized Reinforcement learning), that revisits adversarial RL from a risk-sensitive perspective. First, we introduce a state-dependent adversarial objective that adaptively regulates perturbation strength, suppressing harmful disturbances while preserving informative exploration. Second, we propose critic consistency regularization to reduce disagreement between Q-value estimators and stabilize learning. Comprehensive experiments on challenging continuous control benchmarks demonstrate that RACER consistently improves performance, robustness, and training stability over strong robust RL baselines.

• Routing by Auction: Budget-Constrained Multi-LLM Inference

  Ziying Wu, Hongtao Lv, Lei Liu, Lizhen Cui

  Abstract

  Deploying large language models under a limited inference budget requires balancing answer quality against heterogeneous model costs. Existing routing methods typically assign each query to a single model, while multi-model collaboration approaches often lack explicit budget control and incentive-aware model selection. We formulate budget-constrained multi-LLM routing as a reverse-auction problem in which model providers submit bids for serving a query. The proposed mechanism represents query requirements and model capabilities in a shared five-dimensional space, defines a task-aware model-set utility with diminishing marginal returns, and selects models according to marginal utility per bid under a proportional-share rule. Threshold payments are then used to guarantee individual rationality and truthful bidding under the single-parameter cost setting. We further establish a guarantee on total threshold payments. We evaluate the mechanism on five heterogeneous LLMs across knowledge, reasoning, mathematics, coding, and long-context tasks under multiple budget levels. Compared with greedy multi-model selection, the proposed method reduces average selected-bid cost by 12.4%--38.9%, while keeping the accuracy gap within two percentage points. Under tight budgets, it achieves higher accuracy with lower selected-bid cost than the single-model baselines. These results show that the proposed mechanism achieves lower selected-bid cost with comparable observed accuracy under the evaluated budgets.

• SACE: Source-Aware Context Engineering for Multi-Hop Question Answering Agents

  Congcong Tian, Kang Yao, Yufei Han, Wang Fei, Xu Chu

  Abstract

  Large language model (LLM) agents answer multi-hop questions by decomposing them into dependent sub-questions, retrieving passages for each step, and finally answering from a context window that holds only a few passages. What the agent remembers between steps and what it places in this bounded context are therefore questions of context engineering. In current agents, the memory carried between steps holds only bare intermediate answers: the agent does not remember which retrieved passage each answer came from. As a result, each sub-question is answered without knowing how its answer will be used, and relevance-based context selection may discard the very passages on which earlier steps relied. We propose SACE, a training-free, source-aware context engineering method that keeps the source passage of every intermediate answer in the agent's memory and uses it to fill the final context. When answering a sub-question, SACE shows the reader the original question, the answers obtained so far and the later step that will use the current answer, and asks it to cite the passage that states the answer. When filling the final context, SACE checks whether the source passage of each step is retained and restores missing ones by replacement within the same passage budget, without additional LLM calls or additional retrieval. Experiments on MuSiQue, 2WikiMultiHopQA and HotpotQA show that, relative to the same agent without these two operations, SACE improves EM@5 by 2.7, 0.8 and 0.4 points and complete-evidence recall by 4.9, 2.4 and 1.0 points. Ablations show that the two operations are complementary: usage-aware reading yields more accurate intermediate answers, and source-preserving assembly turns their recorded sources into a more complete final context.

• SALT-GUI: A Benchmark for Salience- and Turn-Controlled Red-Teaming of Mobile GUI Agents

  Haoxin An, Zihao Bai, Qianyi Wang, Mengfan Xu, Yunpeng Song, Zhongmin Cai, Guojun Xiong, Wentao Chen, Feng Wei

  Abstract

  The safety of a mobile GUI agent depends on how it is asked, yet current benchmarks measure only single-turn, explicitly harmful instructions, leaving two dimensions that any ordinary user can traverse uncontrolled: the salience of adversarial intent and the number of turns over which a request is spread. We introduce SALT-GUI, a screen-grounded red-teaming benchmark that treats intent salience and turn structure as controlled axes. SALT-GUI contains 79 harmful cases captured from real Android screens together with their UI-automator hierarchies, complemented by 49 matched benign look-alike controls. Every harmful case is instantiated in four cells crossing concealed vs. explicit intent with single-shot vs. four-turn escalation, further crossed with a prompt-level defense, yielding 948 adversarial utterances per model and defense condition under a user-side persuasion threat model with no environment injection. Evaluating three proprietary and four open-weight GUI agents across 7,322 judged units shows that a one-line safety prompt removes about 20 points of attack success in each regime (-20.6 single-shot, -20.5 escalation), yet leaves intact the +20.3-point gap between single-shot probes and four-turn chains, which count as successful if any turn is unsafe; averaged over models, only 37% of chain-level attack success has surfaced by the first turn of a chain. The defense also reshapes the risk it fails to remove, widening the salience gap by a mean +12.7 points and making concealed intent the more effective tactic for most agents. Reporting a scalar attack-success rate therefore overstates deployed robustness and hides where the remaining risk has moved. The harness and benign controls will be released publicly, and the adversarial probes to verified researchers.

• Scaling Reinforcement Learning via Agent Ensemble

  Yi Ma, Hongyao Tang, Jinyi Liu, Chenjun Xiao

  Abstract

  Existing studies on scaling for deep reinforcement learning predominantly focus on architectural scaling of monolithic agents, leaving the potential of scaling of ensembled agents underexplored. This paper introduces Agent Ensemble Reinforcement Learning (AERL), a framework that achieves performance scaling by orchestrating coordinated learning across multiple agents. The key insight lies in transforming agent quantity into quality through three synergistic mechanisms: 1) Collaborative Replay Buffer where agents jointly maintain a shared experience buffer through diversity-driven exploration --- each agent proposes its individual action evaluated by Q-value, followed by Boltzmann exploration for interaction action selection to enrich the buffer; 2) Follow-the-Leader Policy Regularization, a policy optimization method where each agent undergoes standard policy iteration while simultaneously regularized by the action of the current best policy, enforced through a L2-distance minimization objective across the ensemble; 3) Test-time Best Action Selection that obtains optimal actions during deployment by selecting the proposal with the maximum Q-value across all agents' policy outputs. Extensive evaluations on challenging locomotion benchmarks reveal scaling properties: the system exhibits performance growth as the agent number expands. This work explore the potential of establishing agent ensemble as a new scaling axis for deep reinforcement learning, complementing with architectural scaling.

• SciRelay: Calibrated Bid-and-Relay for Communication-Bounded Scientific Evidence Retrieval

  Wenbin Ma

  Abstract

  Distributed retrieval agents may hold disjoint corpora and use heterogeneous scorers, making raw relevance scores incomparable. Sending every candidate defeats the communication boundary. We introduce SciRelay, a training-free bid-and-relay protocol: each agent emits at most four 20-byte sketches containing a document fingerprint, rank-normalized position, and compact query-term coverage; a coordinator greedily maximizes a monotone submodular surrogate and relays at most $B$ documents. On 5,183 SciFact abstracts partitioned among 4--16 agents, SciRelay improves nDCG over the strongest decentralized baseline by 0.038--0.049 under lexical heterogeneity and 0.108--0.150 with one score-inflating agent. A held-out 809-query split confirms the six gains at 0.042--0.052 and 0.125--0.183. A frozen SPECTER extension confirms gains of 0.059--0.076 with dense agents and 0.164--0.223 when lexical and dense agents coexist. On the 505 confirmation claims with SUPPORT/CONTRADICT labels, relayed evidence raises a fixed FEVER-trained NLI verifier's accuracy by 1.2--3.8 points; only the fault-condition interval excludes zero. A 171,332-document TREC-COVID study reproduces fault robustness but leaves the smaller heterogeneity transfer uncertain. These results support a lightweight protocol contract, not universal superiority or Byzantine security. Code, pinned model revisions, per-query outputs, and executable claim checks accompany the submission.

• SiamFlow-KF: Mitigating Temporal Misalignment for Robust Visual Target Following on Quadrupedal Robots

  Kaiming Xu, Yuehan Wang, Junhao Xiao, Deyu Song, Shuai Wang, Zhiyu Wu, Xiaoxu Zhang, Jian Xu

  Abstract

  Target following on quadrupedal robots suffers from temporal misalignment between local visual search and continuous control state. Rhythmic ego-motion can move the target beyond the search region, while blur, jitter, and occlusion corrupt frame-wise observations; these errors perturb motion and degrade subsequent images. This coupling motivates a tracker that anticipates displacement and preserves state continuity. We propose SiamFlow-KF, a predict--match--refine framework. RoI-restricted optical flow estimates inter-frame displacement to pre-align the search region. A resolution-preserving hierarchical Siamese network with an anchor-free Box Adaptive Head then generates the visual observation, which a constant-velocity Kalman filter refines and propagates across short interruptions. SiamFlow-KF achieves 0.447 EAO and 0.178 Robustness on VOT2018, 0.698 success AUC and 0.910 precision on OTB100, and 0.627 success and 0.690 precision on LaSOT. Unitree A1 trials further demonstrate robust following under viewpoint change, distractors, scale variation, and partial occlusion.

• SLAM: Slot-based Latent Alignment for Multi-Image Multimodal Named Entity Recognition

  Jia Huang, Shuli Zhao, Yaru Zhao, Binyang Li

  Abstract

  Multi-image multimodal named entity recognition (MNER) aims to identify entity spans and types from social media posts that contain text and multiple related images. Compared with the single-image setting, useful visual evidence in multi-image MNER is often scattered across images and mixed with irrelevant visual content, making it difficult to select reliable evidence for each entity mention. To address this issue, we propose SLAM, a slot-based latent alignment framework for multi-image MNER. SLAM uses a set of learnable latent semantic slots to summarize fine-grained evidence from text tokens and image patches, and aligns each candidate mention with the most relevant slots for entity type prediction. We further introduce an contribution-aware slot masking objective to reduce the model's reliance on dominant but potentially noisy slots under limited supervision. Experiments on MNER-MI show that SLAM achieves a new state-of-the-art F1 score of 82.62 under the full-resource setting, outperforming the previous best multi-image baseline by 2.57 points. Under limited-supervision settings, SLAM also obtains an F1 score of 75.85 with only 10% training data, outperforming TPM-MI by 13.89 absolute points. These results suggest that structured slot-based evidence abstraction is effective for multi-image MNER, especially when cross-image evidence is noisy or supervision is limited. Code is available at https://github.com/Jialy-lily/SLAM-MNER.

• SPACE: State-Guided Workflow and Evidence Management for Multi-Agent Troubleshooting of Cloud Systems

  Ruofei Hou

  Abstract

  Autonomous cloud-system troubleshooting requires agents to maintain a coherent diagnostic state, interact reliably with operational tools, and verify corrective actions over extended trajectories. Although lightweight language models are suitable for private deployments with limited resources, their performance in long-horizon AIOps tasks is often affected by unstable workflows and uncontrolled context accumulation. This paper presents Space, a state-guided multi-agent framework that coordinates workflow progression with evidence management. Space combines bounded parallel exploration for initial evidence collection, intent-guided refinement of tool outputs, and a state-controlled diagnosis loop that separates decision making, tool execution, observation analysis, and execution-error handling. Instead of retaining complete interaction histories, the framework maintains compact action-evidence-analysis records in a curated working memory. We evaluate Space on 48 AIOpsLab tasks using Qwen3.5-9B as the common backbone. Space achieves a 75.0% overall success rate, exceeding the strongest controlled baseline by 22.9 percentage points, with success rates of 63.6% for root-cause analysis and 45.5% for mitigation. It also reduces execution failures and redundant actions to 4.3% and 8.7%, respectively. The results demonstrate that structured workflow control and selective evidence retention improve the reliability of lightweight agents in long-horizon cloud-system troubleshooting.

• SPC: Selective Process Control in Multi-Agent Recommendation

  Jiayan Zhuang, Li-e Wang, Zhigang Sun, Qiuxia Gong, Jinke Xu

  Abstract

  In recent years, LLM-based agents have become an important research direction in recommender systems, providing a new paradigm for modeling complex user preferences and interactive recommendation.Multi-agent recommendation organizes recommendation decisions as a collaborative process involving multiple language agents, allowing the final ranking to emerge through continuous information exchange.Existing methods typically use reasoning and reflection to provide opportunities to re-examine and adjust recommendation results.However, existing methods mainly focus on enhancing agent collaboration or revising recommendation results, and have not used observable candidate states during collaboration to determine whether further processing is needed.To address this issue, we propose Selective Process Control (SPC) for multi-agent recommendation.Specifically, SPC converts the candidate rankings progressively formed during collaboration into online branch states and selectively activates localized re-execution based on these states.For tasks that satisfy the control conditions, SPC determines the suffix re-execution boundary from continuous monitoring records, surfaces task-valid alternative directions using the candidate topology, and re-executes the affected collaboration process while preserving the original task, candidate set, and agent configuration.Experimental results show that SPC effectively improves the terminal ranking utility of the underlying multi-agent recommendation workflow. On the Office Sparse setting, its NDCG@1 improves by 16.67% over the best-performing baseline for this metric.

• SSGM: A Framework and Evaluation Protocol for Governed Memory in Agentic AI Systems

  Chingkwun Lam, Jiaxin Li, KoPang, Lingfei Zhang, Zhao Kuo

  Abstract

  Persistent memory creates risks beyond retrieval error: agent systems must govern what they store, expose and repair. We present Stability and Safety Governed Memory (SSGM), a framework and evaluation protocol for gated writes, scoped reads, contradiction quarantine and ledger-grounded repair. The protocol evaluates structured governance, raw-to-record induction, provenance stress and benign long-memory utility separately. With trusted structured control fields, SSGM achieves 0% observed risky-write acceptance and leakage success across three stratified 500-scenario LME-Gov resamples, while preserving 99.63--100.00% benign-write acceptance. In the representative Qwen3.5-9B setting, a memory-utility proxy on LongMemEval decreases by 0.10 percentage points relative to write-through, while a separate LoCoMo QA evaluation yields near-baseline answer F1. Raw LME-Gov identifies errors in recovering keys and provenance as the main observed source of unsafe admission. These results support auditable governance of persistent memory under explicit trust assumptions.

• Stability and Last-Iterate Convergence of Hedge Dynamics with Best-Response Oracles in Zero-Sum Games

  Junyue Zhang, Xinxiang Guo, Yifen Mu

  Abstract

  In this paper, we study the stability and last-iterate convergence of Hedge dynamics with best-response oracles in repeated two-player zero-sum (min-max) games. Unlike previous studies, our results do not require the Nash equilibrium to be unique or interior, owing to a KL error bound that we establish to control the KL divergence linearly by the strategy gap through a suitable choice of Nash strategy. In Hedge-Myopic dynamics, the minimizer updates by Hedge while the maximizer plays a best response to the current strategy. We prove stability under a constant learning rate and last-iterate convergence under a decreasing learning rate. In adaptive Hedge-Best Oracle dynamics, both players have access to best-response oracles and update by Hedge using a learning rate determined by joint exploitability. We obtain an $O(1/t)$ rate for the joint KL energy and an $O(t^{-1/2})$ exploitability bound. An example with a unique boundary Nash equilibrium attains the $\Theta(1/t)$ KL rate, showing that this rate is tight in the boundary case. This gives a negative answer to the open question raised by Munos et al. 2020 of whether exponential convergence can still be guaranteed without an interior Nash equilibrium. Numerical experiments support our results.

• StylePlan: Movement-Style-Guided Multi-Agent LLM Planning for Quadrupedal Locomotion

  Lin Li, Yuze Yuan, Wei Wei, Jianguo Wu

  Abstract

  Reliable quadruped locomotion in unstructured environments requires not only safe traversal of complex terrain but also effective optimization of task utility under varying rules, time constraints, and reward conditions. Existing locomotion methods primarily focus on local stability and terrain traversability, whereas planning methods based on large language models generally produce semantic actions or continuous control parameters that cannot be reliably mapped to low-level control interfaces. To address these limitations, this paper proposes StylePlan, a hierarchical multi-agent framework based on large language models. StylePlan decomposes the decision-making process into task planning, path refinement, movement-style planning, and structured encoding, thereby transforming natural-language task rules and terrain descriptions into executable sequences of waypoints and gait parameters. The framework employs movement style as an intermediate representation between high-level reasoning and low-level control, enabling coordinated adjustment of body posture, forward velocity, step frequency, and gait pattern. This design allows the robot to adapt to changes in task rules without retraining the low-level controller. In simulated tasks with easy, normal, and high levels of difficulty, StylePlan achieved higher task scores than the evaluated baselines in the normal- and high-difficulty settings. In real-world experiments involving gravel surfaces, slopes, and 12-cm steps, StylePlan increased the overall success rate from 53.3% with a fixed-parameter baseline to 93.3%. These results demonstrate that the proposed framework enables terrain-adaptive quadruped locomotion while effectively balancing task utility, execution robustness, and practical deployability.

• Tacit Algorithmic Collusion under Gradient Play: Nash Equilibrium Stabilization via One-Firm Algorithm Redesign

  Junyue Zhang, Yifen Mu

  Abstract

  Independent pricing algorithms may sustain supracompetitive outcomes without communication, a phenomenon known as tacit algorithmic collusion. We study this possibility under simultaneous constant-step gradient play in a logit-Bertrand market and ask when redesigning only one firm's algorithm can restore the unique Nash equilibrium. We find that sufficiently aggressive steps can destabilize the equilibrium through a period-doubling bifurcation. We derive local conditions under which the resulting two-cycle exists, is attracting, and raises every firm's average profit, thereby constituting an outcome-based tacitly collusive price cycle under a fixed profile of independently executed algorithms. We then characterize exactly when an admissible finite-dimensional controller at one firm can locally exponentially stabilize the Nash equilibrium while all rivals retain gradient ascent. The controller uses only that firm's price, profit gradient, and finite private memory online. Guided by this characterization, we construct a one-memory two-tap price-update rule for one firm. In a three-firm example, we certify stability of the full price-and-memory dynamics and show that this rule succeeds where merely reducing one firm's step size or replacing its algorithm with standard optimistic gradient ascent fails.

• TAPP: Transparent PII Pseudonymization Between AI Assistants and Cloud LLMs

  Dennis Itzel, Danila Mamontov

  Abstract

  Agents built on cloud-hosted LLMs transmit their entire context on every turn, personal data included. We present TAPP, an open-source privacy proxy, that pseudonymizes personally identifiable information (PII) on every message an agent sends and restores it on the return path, covering streamed output and the tool-call arguments through which restored values re-enter the model on later turns, and leaving both agent and model unmodified. An end-to-end study across seven LLMs and 120 bilingual tool-calling tasks (conversation, e-mail, calendar) compares three conditions: no proxy (Control), the proxy with ground-truth PII spans (Mocked), and the proxy with the Microsoft Presidio Analyzer (Presidio). Interposition alone costs nothing measurable. Complete masking costs 20 to 35 percentage points of task success per model, while the real detector narrows the gap to 1 to 19 points by leaking undetected values, so functionality is non-monotone in detection completeness. The loss falls almost entirely on arguments the assistant must transform into a tool's format: calendar success collapses from 99.8% to 17.8% under Mocked and recovers to 86.7% under Presidio, because the date and time values the transformation consumes are precisely what Presidio misses, while e-mail, whose recipient is passed through, never drops below 98.6%. The failure boundary is therefore not masking volume but the pairing of opaque surrogates with strict tool schemas, a coupling that applies to any mechanism obscuring agent context while leaving its structure intact. The source code is publicly available at https://github.com/PlueschtierBaum/TAPP-PrivacyProxy.

• Task as Gradient: Implicit Geometric Embedding for Differentiable World Models of Embodied Agents

  Mingyu Jia, Junchi Yan, Haobin Zhu

  Abstract

  Differentiable spatial world models constitute a foundational component of model-based reinforcement learning for embodied agents, supporting gradient-enabled planning through learned environment simulation. Nevertheless, prevailing approaches formulate task objectives as externally imposed ad-hoc constraints, typically realised via auxiliary reward functions or explicit tracking losses. Such an explicit modelling paradigm leads to structural decoupling between environmental representations and task intent, introducing pronounced hyperparameter sensitivity and destructive gradient conflicts. These issues frequently yield local deadlocks and undesirable edge-following behaviours within complex spatial topologies. In this work, we present a principled paradigm shift termed Task-as-Gradient built upon implicit geometric embedding. Our method transforms discrete task specifications into geometric virtual boundaries and jointly encodes these task constructs alongside real-time physical observations within a unified, analytically continuous feature field. Within this shared latent space, task intent manifests as intrinsic gradient signals that implicitly steer trajectory optimisation, removing the need for manual balancing of multiple explicit cost terms. Rooted in three core design principles — differentiability, implicit integration, and parameter decoupling — we construct a general planning framework and implement its concrete instantiation, Neu-Star, for autonomous robotic navigation. Theoretical analysis and empirical evaluations verify that our framework effectively mitigates local deadlocks and overcomes the conventional trade-off between safety and efficiency. Extensive experiments show consistent gains in planning efficiency, trajectory optimality, and zero-shot generalisation robustness for generalist embodied agents. We will release our implementation upon acceptance.

• Test Before You Trust: Falsification-Aware Agent Coalitions for Stock Movement Forecasting

  Yuanjian Zhang, Chuanchang Liu, Yifan Yuan, Xiang Xu, Bo Cheng

  Abstract

  Agents that forecast a stock together must decide whose evidence to believe, and multi-agent systems answer that question with a quantity each member produces about itself: a confidence score, an attention weight, a router logit. This basis is unsound: confidence turns out to place no constraint on robustness, so at equal confidence a fragile claim and a sound one are indistinguishable to any weight computed from the member's own forward pass, and evidence that is stale, duplicated, or carried by a correlation that will not persist enters the decision unchallenged. Separating these cases calls for intervening on a claim's premises rather than rescaling its output. This paper introduces FARCast, an agent coalition for multi-horizon stock movement prediction in which a learned regime segmentation turns each stretch of a stock's history into a temporary evidence agent, and no agent is believed until it has been tested. An internal red team perturbs the premises of every claim and records the displacement, a predictor estimates each agent's leave-one-out utility, a judge combines the two with self-confidence and admits a small coalition chosen for complementarity rather than agreement, and admitted evidence revises the stock's local prior through a saturating update that provably caps the influence of any coalition, however large or correlated. Across three markets FARCast lifts average MCC from $0.323$ to $0.370$ over the same segmentation fused by attention, the margin widening from $0.038$ at one day to $0.056$ at five, and it loses $0.019$ MCC under a correlated-majority attack where attention pooling loses $0.083$. Because directional labels are available, each agent's true contribution is measurable, so admission quality is scored directly rather than inferred from accuracy.

• The Significance of Domain, Opponent, and Phase Adaptation in Bilateral Automated Negotiation

  Yichen Lv, Long Chen, Sunil Dahal, Katsuhide Fujita, Shengbo Chang, Zigao Wu

  Abstract

  An automated negotiation agent must decide what to offer, how quickly to concede, how to interpret the opponent's offers, and when to accept. The BOA architecture separates these decisions into bidding, opponent modeling, and acceptance. We study a related question: which information should these components adapt to? We consider three sources: the domain, the opponent, and the negotiation phase. We implement them in a single BOA-based agent. Domain statistics set the initial concession and search parameters. An ensemble of four opponent models ranks candidate offers. An online classifier adjusts a small set of weights after enough offers have been observed, while a bounded phase signal modifies the concession curve during negotiation. We evaluate the agent on six domains, seven opponent types, and 2,520 negotiations per variant. The full agent achieves a mean self utility of 0.807 and an agreement rate of 98.8%. The component analysis attributes most of the improvement to the opponent-model ensemble. Domain adaptation provides a smaller independent gain, whereas the current opponent classifier and phase controller have little effect on the aggregate result. The final score alone hides these differences, so we evaluate the agent at the component level.

• The Temperature an Adversary Would Choose: Break-Even Temperatures and the Price of Auditing Decentralized LLM Inference

  Krishna Harish

  Abstract

  Decentralized and outsourced LLM inference creates a principal-agent problem: a node paid to serve a premium model may quietly serve a cheaper substitute. The standard defense is a comparison audit, a two-sample test between the node's outputs and a trusted reference. We show that the sampling temperature the node controls is a systematic evasion dial, characterize exactly how much it buys under which audit interfaces, and measure, for the first time, whether the temperatures that defeat an audit are temperatures an adversary can afford to serve. The mechanism is specific to shared softmax temperature scaling: a common temperature T simultaneously raises entropy and contracts inter-model divergence, and we prove the expected per-token divergence obeys D(T) = K/T^2 + A3/T^3 + O(1/T^4), with both coefficients in closed form from a single raw-logit snapshot. Composing this with classical testing limits yields a break-even temperature scale T* proportional to sqrt(nK) for fixed-budget audits (the constant is a calibrated decision-rule property, not a universal one: calibrated once on GPT-2 it predicts held-out same-family break-evens to within 13% and transfers cross-family with 22.8% error, in the conservative direction), and a parameter-free sequential price law E[N] = c_seq T^2 / K. These oracle laws upper-bound every realizable audit, and the realizable ones collapse far earlier: alternative-agnostic black-box audits at the same budget never reach 50% power on any pair, even at the contract temperature T = 1, and top-20 logprob interfaces retain about 0.1% of the auditable signal, barely above the vocabulary-fraction baseline, because the logit gap between same-family checkpoints is nearly diffuse across the vocabulary. Measured quality-temperature curves close the loop: for unbiased-sampled open-ended text, quality collapses so fast in T (a 1-nat judge-NLL budget pins T at or below 1.13) that the temperature dial barely overlaps even the weakest audit's collapse range, and the binding threat is instead quality-preserving substitution (int8 quantization) near T = 1, which the same framework prices via a measured quantization ladder. The evasion that remains is confined to entropy-blind auditors: a sample-only collision meter at the same budget flags temperature inflation with power at least 0.92, and even evasion tuned to exactly match the claimed per-position entropy retains a divergence floor of at least 0.26 nats and is caught with power at least 0.973.

• Threshold Intervention Audits for Hidden Coalitions in Nominally Cooperative DCOPs

  Weihan Long, Yujiao Peng

  Abstract

  We study exact recovery of a hidden coalition partition when an auditor may locally neutralize agents but observes only one aggregate safety bit. The nominal cooperative distributed constraint optimization problem (DCOP) is fixed. Each hidden coalition cell can independently sustain a secondary deviation; an intervention is reported successful precisely when it hits every cell. With $n$ agents, known $m$ cells of minimum size $\ell$, and intervention budget $b$, the resulting oracle is a bounded-size spanning oracle for a partition matroid—equivalently, an edge-detecting oracle for a hidden covering hypermatching restricted to sets of size at least $n-b$. For fixed $m,\ell\ge2$, we prove the complete deterministic and constant-error randomized phase diagram: recovery is information-theoretically impossible for $b<m$, whereas for every $b\ge m$ its worst-case query complexity is $$\Theta_{m,\ell}\!\left(n+(n/b)^{m-1}\right).$$ The upper bound combines an explicit grouped covering with minimal-transversal extraction and basis exchange. The lower bound combines a rare-witness all-zero adversary with an independent partition-entropy bound. At the critical budget $b=m$, an exact counting argument gives the sharper $\Theta(n^{m-1})$ obstruction, while balanced cells admit linear randomized recovery. We realize the oracle using degree-two hidden factors and identical-interest coalition utilities under strong Nash response. We also prove that the required baseline switch cannot be uniquely realized by ordinary Nash equilibrium, and that unrestricted strategic concealment destroys identifiability. The results isolate intervention capacity, observation semantics, and behavioral faithfulness as distinct limits on coalition auditing.

• Topology-Aware Rollout Correction for Long-Horizon Power-Grid Dynamics Prediction

  Hao Wu, Xiangyu Zhang, Feng Liu, Yanjie Wang, Hanbo Kou, Faguo Wu, Xiao Zhang

  Abstract

  Reliable long-horizon power-grid dynamics prediction remains difficult even when a learned surrogate is accurate for one step. Small local errors compound during autoregressive rollout, and a model trained on one grid can fail when the number of electrical nodes (buses) and the connectivity pattern change. Existing graph predictors do not simultaneously control accumulated rollout error and generalize across topologies. To address these challenges jointly, we propose a two-stage graph-dynamics framework with node-local computation and neighbor message passing. An attention-gated topology-adaptive graph convolutional network (TAGCN) learns a shared neighborhood-local transition rule for zero-shot transfer. A lightweight Topology-Guided Rollout Correction (TGRC) module then adapts a frozen backbone by correcting accumulated residuals, while Physics-Informed TGRC (PI-TGRC) predicts power-injection adjustments and advances the state through a differentiable Kuramoto step. Across four unseen regional grids, the shared backbone achieves a one-step coefficient of determination R² ≥ 0.895. At the primary long horizon, T = 50 (5 s), TGRC reduces MSE from 1.976 to 0.650—a 67.1% improvement over the uncorrected backbone—with only about 72K trainable correction parameters. Under larger topology shifts, target-trained PI-TGRC remains stable by restricting correction to the model's physical forcing space. The framework therefore combines zero-shot cross-topology transfer with data-driven and physics-constrained routes to long-horizon rollout correction within a unified design for distributed prediction in networked physical systems.

• TradeWar: A Benchmark for Execution-Grounded Trade Negotiation with Adaptive LLM Agents

  Xiangyu Kong, Ruirui Yang, Xia Hou

  Abstract

  Multi-turn negotiation remains a challenging problem for large language model (LLM) agents. While LLMs can conduct natural-language dialogue, existing benchmarks predominantly assess only agreement rates, overlooking whether negotiated commitments actually create value once executed. To address this gap, we introduce TradeWar, a multi-country simulation benchmark that translates negotiation dialogues into concrete tariff and trade-share policies, enforces capacity and network constraints during settlement, and evaluates outcomes via an Economic Health Index (EHI). To solve this execution-grounded task, we propose TACTIC (Trade-aware Adaptive Consolidation and Theory-of-mind for International Commerce), an agent that iteratively refines its negotiation strategy across episodes—without updating base LLM weights—by leveraging prior experience, opponent behavior, and past settlement outcomes. Despite using zero-shot chain-of-thought (CoT) prompting, PromptLLM attains a 75% agreement rate but only 57.1 final EHI (×100), whereas TACTIC reaches 68.8. Across 20 sequential episodes and three seeds, TACTIC’s final EHI improves from 58.2 to 73.2, compared with 65.2 for an ablation without cross-episode updates. Under technology-controls and energy-crisis shocks, TACTIC also exhibits smaller performance degradation than PromptLLM and MemoryLLM. These results indicate that CoT deliberation and interaction memory alone do not fully address delayed execution feedback, whereas settlement-grounded adaptation improves long-horizon negotiation in TradeWar.

• TriFuse-Mem: Validity-Constrained Associative Retrieval over Versioned Agent Memory

  Zhuoran Liang, Xiaoyan Kui, Sihao Wang

  Abstract

  Agent memories evolve through corrections, retractions, and expirations, yet associative retrieval can still let obsolete observations seed a query, transmit relevance, or determine returned evidence. We formulate this problem as validity-constrained associative retrieval and introduce TriFuse-Mem. It projects source-linked lifecycle-atomic units from an append-only bitemporal event log. The resulting state governs bounded answer-blind chain selection, entity-type-passage propagation, and exact source-version return. Each unit couples a fact lineage to a canonical source identifier and exact text bytes; a multi-unit source version is admitted only when all declared units remain active, otherwise construction fails closed. Selected facts contribute entity and validated passage support before one backbone personalized PageRank computation. We evaluate three static multi-hop QA datasets, MemoryAgentBench FactConsolidation at 262K context (MAB262K), original LongMemEval-S, and WikiFactDiff replacement transfer under shared evidence and generation budgets. TriFuse-Mem attains the highest aggregate score on the static suite and both dynamic benchmarks. On lifecycle-declared inputs, MAB262K ablations show a descriptive chain-projection interaction, and WikiFactDiff measures replacement-state resolution. LongMemEval-S separately evaluates source-grounded answer transfer using full benchmark-native sessions. Fourteen deterministic admission cases additionally verify joint-lineage admission and exact source preservation through the native materialization entry point. Code and evaluation artifacts will be released upon publication.

• Unifying Multifaceted Identities: Communication-Efficient Distributed Learning on Heterogeneous Graphs

  Xiaoyi Fu, Xinhu Zheng, Zhen Hao Wong, Quanming Yao, Yin Nan, Liuqing Yang, Yuan Yao

  Abstract

  Learning effective representations on decentralized, heterogeneous graphs presents a fundamental trade-off for distributed artificial intelligence (DAI): expanding a Graph Neural Network’s receptive field to accommodate edge-type heterogeneity incurs prohibitive cross-agent communication. We formalize this challenge as the Multifaceted Identity (MI) problem, where a single entity appears across multiple agent subgraphs with varying incident edge types. To address it, we propose Federated Heterogeneous Historical Embedding (FedH2E), a communication-efficient method for learning MI representations that consistently reaches target accuracy with approximately 50% less cross-agent communication than strong synchronous baselines. We theoretically derive approximation error bounds for the historical embeddings used as proxies for expensive real-time message passing. Experiments on three heterogeneous graph datasets confirm the theory and show that FedH2E remains robust under increasing decentralization and larger GNN receptive fields—regimes where baseline accuracy typically degrades.

• URGENT: Uncertainty-guided Adaptive Refinement for Grounding in GUI Agents

  Changyi Hao, Xiaodong Liu, Bin Ji, Shasha Li, Ma Jun, Jie Yu

  Abstract

  Vision-language models based GUI agents have achieved promising performance on GUI grounding, which aims to localize target interface elements indicated by natural language instructions. However, existing methods usually directly output the global best prediction without considering its localization uncertainty, which may lead to wrong grounding in complex GUI scenarios. To address this issue, we propose URGENT, an Uncertainty-guided adaptive Refinement framework for Grounding in GUI agENTs. Specifically, {URGENT} first leverages the spatial distribution of Top-$K$ predictions generated during the global prediction stage to estimate the localization uncertainty of the global best prediction, and then selectively activates local refinement to update the best prediction if the uncertainty exceeds a predefined threshold. In contrast, the global best prediction is considered as the final prediction if the uncertainty is below the threshold. Additionally, we propose a consistency-guided global-local fusion strategy to improve grounding stability. Extensive experiments on ScreenSpot, ScreenSpot-v2, and ScreenSpot-Pro demonstrate that {URGENT} delivers consistent performance gains across various GUI-Actor backbone LLMs. In particular, on ScreenSpot-Pro, {URGENT} delivers +2.8% success rate on GUI-Actor-2B, and +4.5% success rate on GUI-Actor-7B. Extensive analyses further indicate {URGENT}'s effectiveness in GUI grounding.

• Utility-Aware Uncertainty-Gated World Models for Sample-Efficient Reinforcement Learning

  QR Li

  Abstract

  World models improve reinforcement learning by converting real interaction data into model-generated transitions, but this benefit is conditional on how the synthetic transitions are used. A rollout that is accurate but policy-irrelevant wastes updates, whereas a rollout that is useful but unreliable can corrupt value learning through compounding model error. Existing model-based reinforcement learning methods typically handle this trade-off by using uniformly short rollouts, penalizing uncertainty, or scaling latent world models, but they seldom make transition-level decisions that jointly account for predictive reliability and policy utility. This paper proposes UA-UGWM, a utility-aware uncertainty-gated world-model framework for sample-efficient reinforcement learning. UA-UGWM trains an ensemble world model, estimates predictive uncertainty and policy utility for each candidate model transition, and uses a gate to accept, down-weight, shorten, or reject synthetic rollouts before policy/value updates. For sparse symbolic navigation, we further introduce an event-aware extension that models key, door, invalid-action, and goal events and uses them to control sensitive transitions. Experiments on Gymnasium official control tasks and MiniGrid sparse navigation tasks show that UA-UGWM achieves the best composite utility score on the Gymnasium benchmark (0.947) while issuing 46% fewer synthetic updates than the strongest competing method, and that it retains the top rank under every alternative metric weighting tested. MiniGrid analyses further show that event-aware gating improves subgoal acquisition and reduces unreliable long rollouts in sparse-reward tasks. Ablation and gate diagnostics confirm that the performance gain comes from selective synthetic-sample usage rather than simply increasing the number of model-generated updates, and that admission and rollout truncation contribute to this gain through distinct mechanisms.

• Well-Typed Agents Don't Deadlock: Session Types for MCP and A2A Composition

  Mohammed Imran Khan, Manju Khanna, Abhishek Shivkumar, Tuhin Sharma, Adinarayana Sakala, Nirav Kumar

  Abstract

  The emergence of agentic AI has produced two dominant interoperability protocols: the Model Context Protocol (MCP) for tool integration and the Agent-to-Agent (A2A) protocol for inter-agent task delegation. Production systems increasingly compose both protocols, yet no formal framework characterizes the safety properties of their interaction. We present SessionStack, a session-typed compositional framework for reasoning about MCP+A2A protocol stacks. We define a process calculus for protocol composition, model MCP tool-call sessions and A2A delegation sessions as binary session types, and prove subject reduction: well-typed compositions preserve typing under reduction. We introduce delegation coherence, a decidable well-formedness criterion, and prove that delegation-coherent compositions are deadlock-free under fair scheduling. A linear-time algorithm checks coherence statically, while a complementary runtime monitor enforces session conformance with sub-microsecond overhead. Evaluation on synthetic benchmarks and 10 topologies from six agent frameworks (LangGraph, CrewAI, AutoGen, Semantic Kernel, Google ADK, A2A reference) shows that SessionStack identifies all injected coherence violations and flags circular delegation patterns in 6 of 10 examined topologies, with sub-millisecond analysis time, scaling to 500-agent topologies in under 17ms.

• When Chain-of-Thought Hurts: Decoupled Verdict-and-Rationale Learning for Clinical Antibiotic Appropriateness Assessment

  Zijie Zhou, Yijian Chen, Wang Bei, Zhao Wang, Guoming Wang

  Abstract

  Chain-of-thought (CoT) reasoning is widely assumed to improve both accuracy and interpretability in clinical NLP. We challenge this assumption in antibiotic appropriateness assessment, where CoT fine-tuning degrades accuracy from 82.6% to 75.0% compared to plain SFT. Error analysis reveals 26.2% of CoT outputs exhibit reasoning drift: the generated rationale contradicts the final label, suggesting that reasoning supervision interferes with, rather than guides, classification. We propose DVRL (Decoupled Verdict-and-Rationale Learning), separating decision pathway (via a dedicated verdict head on a <|verdict|> token) from rationale pathway. Ablation shows that classification gains come from verdict-centric, risk-aware optimization---not from free-form rationale modeling alone. On a 172-sample held-out test set, DVRL achieves 88.9% accuracy (+6.3 points over SFT) and 0.817 macro-F1, while a verdict-only variant achieves comparable performance at lower inference cost (88.4% accuracy, 0.821 macro-F1). An output-order ablation further shows that generating the verdict before the rationale reaches 95.3% accuracy and 0.930 macro-F1, with inappropriate-class recall rising to 0.95. Our findings suggest that in imbalanced high-stakes clinical classification, rationale generation contributes interpretability rather than predictive power, while verdict-centric, risk-aware optimization is the main driver of accuracy gains and safer risk-constrained automation.

• When Does Adaptive Grounding Help Agent Convention Reconciliation? Trees, Abelian Symmetries, and an Exact Triangle Separation

  Weihan Long, Yujiao Peng

  Abstract

  Independently developed agents may encode the same task referents through incompatible private conventions. We formulate grounded convention reconciliation on a network: node conventions belong to a known finite permutation group, while an edge query selects one source token and returns the target token grounded to the same latent referent. The objective is exact recovery of all relative conventions, up to the unavoidable global gauge, with minimum worst-case queries. We prove that articulation points induce an exact decision-tree direct sum and that every tree requires one permutation-group base per edge. For Abelian convention groups, the complete network oracle is itself a faithful permutation action, so adaptive and fixed query designs have equal exact complexity on every connected graph. The boundary is strict. For the signed-affine action $x\mapsto a\pm x$ over any odd finite field of size at least five, with publicly field-labelled token interfaces, a triangle requires three adaptive queries but four nonadaptive queries. The adaptive protocol uses one returned label to address a later cross-agent token; an affine-transcript argument rules out every three-query fixed design. Block additivity then gives exact costs $2b_2+3b_3$ and $2b_2+4b_3$ on a triangular cactus with $b_2$ bridges and $b_3$ triangle blocks. Thus cyclic topology and noncommutative convention composition are both necessary resources for adaptive value in this deterministic, exact, noiseless grounding model.

• When Reasoning Fails to Stop: Distraction-Induced Non-Termination as an Overthinking Failure in Reasoning LLMs

  Weiye Dai, Liming Nie

  Abstract

  Hosted reasoning APIs now expose how much effort a model spends on each question, and a growing literature reads confidence off that trace, treating its length as a free, label-free reliability signal. We show it can be an artifact: the answer appears only after reasoning stops, so a run that exhausts its budget returns nothing and is scored wrong at maximal effort. Raising the budget fourfold and scoring only completed traces collapses an apparently strong signal (AUROC 0.71) to chance (0.46).

  That control isolates the underlying failure, non-termination: the model deliberates until its budget is gone and commits no answer—detectable from one API field, without ground truth. On GSM8K, MATH-500 and GPQA-Diamond, a distractor drives it to 48% on the model we study in depth while accuracy on finished answers is unchanged (0.72 vs. 0.73), and it persists at every budget we test. Across seven models from six vendors, baseline stalling is general but its rate is not, spanning 2–27%; it costs 39–61% of tasks at depth 3 in an agent pipeline and resists hedged retry (51% recovered, not the simulated 89%). Non-termination is thus a measurable, model-dependent reliability axis, auditable for free before deployment.

  Code and cached traces: https://github.com/dai-weiye/DAI

• When Tool Retrieval Fails: Calibrating Tool Existence in Noisy Open Registries

  Zhuang Yiheng, Yueying Liu, Wenlong Hou, Weidong Zhao, Liu Xianhui

  Abstract

  An agent retrieving tools from a large open registry will routinely face a miss: no returned candidate clearly fits the request. A miss conflates three worlds: the right tool exists but hides behind a noisy description, a different tool could acceptably substitute, or the capability is genuinely absent. Deciding whether to CALL, SEARCH, or REFUSE entails asymmetric costs. We formalize this post-miss decision as tool-existence calibration: estimate a calibrated posterior that an acceptable tool exists in the observed registry, then act on it with a cost-sensitive rule. We build OpenTool-Exist, a benchmark that realizes a measurable version of each world over a 44k-tool registry, with a 300-item human-labeled track including 60 misses that occur in the wild. Our central finding is that retrieval confidence is the wrong uncertainty signal after a miss. The signal lies in substitute evidence: what the registry’s other descriptions say can take the job. An external, auditable classifier over this evidence lifts existence AUROC from 0.64 to 0.82 on retrieval-miss cases under leakage-controlled splits, while schema and registry-density evidence add nothing. At a 5% hallucinated-call budget, the resulting rule recovers 58% of missed-or-substitutable tools, versus at most 14% for answerability baselines, and, on the independently re-labeled track, surpasses equal-cost direct action on the best-found substitute. After retrieval fails, descriptions of possible substitutes provide more useful evidence than the retriever’s confidence in that failure.

• Will More Evidence Help? Estimating Pre-Acquisition Action Value with LLM Internal States

  Chaodong Tong, Jiayang Gao, Qi Zhang, Nannan Sun, Lei Jiang, Chengqing Guo, Yanbing Liu

  Abstract

  Existing adaptive retrieval methods typically use answer uncertainty to decide whether to acquire additional evidence. Yet uncertainty mainly reflects the reliability of the current answer, rather than whether further evidence will improve it. We study action-value estimation before acquisition: after an answer is generated but before new evidence is observed, predicting whether evidence acquisition is preferable to stopping and returning the current answer. We propose GainRoute, which estimates acquisition value by accounting for its potential corrective benefit, the risk of disrupting an already correct answer, and the acquisition cost. It first learns an observable baseline from the question, current evidence, answer, and generation statistics, then freezes this baseline and uses LLM internal states aligned with evidence processing and answer generation to capture additional predictive information. The motivation is that model outputs and uncertainty may not fully reveal why an answer is wrong or whether the error can be repaired by additional evidence. Controlled intervention experiments on three multi-hop QA datasets show that GainRoute effectively supports cost-sensitive routing between stopping and evidence acquisition. The results further suggest that LLM internal states contain predictive information about the consequences of evidence acquisition beyond what is captured by answer uncertainty and other externally observable signals.

• A Structured Knowledge Infrastructure for Domain-Specific Data Asset Discovery

  Mengdi Chen, Yuanxin Huang, Yulin Jiang

  Abstract

  Enterprise data analytics agents face two structural failures: generic RAG retrieves the wrong asset (Hit@10=19.1%) and delivers no usage knowledge to prevent metric misinterpretation---stemming from four root causes (C1--C4) ranging from semantic gap and entity ambiguity to schema drift and asset-usage gap. We present a two-layer solution deployed in a commercial advertising warehouse (5,300+ Hive tables, 14 domains). A three-tier dual-purpose knowledge base (179 documents, eight-section annotation template) serves both retrieval and generation, with a closed-loop refresh pipeline maintaining day-level freshness (one yes/no approval, 30s hot-reload). The Graph-Guided Retriever (GGR) uses a 2,859-node knowledge graph as a candidate gate with intent routing to deliver 71.6x token reduction. The Scene-Aware Ranker (SAR) applies 19-class entity recognition and explicit scenario annotations; negative knowledge alone contributes 25 percentage points of Hit@10 gain. On two 100-question benchmarks, Hit@10 rises from 19.1% to 96.6% (+77.5pp) and knowledge coverage from 56% to 77%, at 4.84--5.33s end-to-end latency.

• Aforge: Turning Agent Evaluation into Scalable, Auditable Training Data

  Jing Liao, Yiming Peng, Yubin Chen, Lianke Li

  Abstract

  Agentic AI systems require multi-step training trajectories with tool calls, environment observations, recovery behavior, and verifier outcomes. Agent evaluations contain these ingredients, but fragmented execution, capture, and data-processing infrastructure makes them difficult to preserve and reuse.

  We present Aforge, a service platform that connects Harbor evaluation to distributed cloud sandboxes and turns evaluation byproducts into provenance-linked training-data candidates. It combines a shared agent-adapter runtime, content-derived image build and cache, and provider-compatible model capture with auditable result-bundle assembly, without modifying the evaluated agents' source code. Result bundles bind trajectories to task and environment context, tool information, and verifier outcomes. Evaluation separates task success from source-relative fidelity and training-input compatibility.

  We evaluate capture across three model-API contracts and coverage across eight coding-agent implementations. In a 48-trial fidelity study, 44 trials passed task verification, but only 22 met all acceptance gates using native outputs. Evaluation-only offline reconstruction increased this count to 39/48 (81.3%) using retained evidence without rerunning the agents. Five paired blocks showed higher accepted-trajectory goodput at requested concurrency 5 and 10 than at 1; a separate capacity pair remained nearly flat as attained peak concurrency rose from 51 to 100. These results demonstrate an auditable path from agent evaluation to training-data candidates, with downstream learning benefits left for future work.

• AgentPanel: Toward a New Paradigm for Human–AI Collaboration in Exploring Scientific Questions

  Zhiyao Cui, Qianyi Wang, Haoyang Yan, Yiqun Zhang, Siyue Ren, Hangfan Zhang, Zelin Tan, Hao Li, Chunjiang Mu, Dexian Cai, Shao Zhang, Chen Zhang, Meng Li, Jianan Chai, Fan Yuting, Zichao Ye, Xiaolei Yang, Xinyao Lu, Yuyang Yu, Wenjie Lou, Xiaosong Wang, Fenghua Ling, Shiyang Feng, Mao Su, Qiaosheng Zhang, Bo Zhang, Lei Bai, Yang Chen, Shuyue Hu

  Abstract

  Identifying promising scientific ideas remains an important challenge in research practice. Researchers commonly rely on small-group discussions or one-to-one interactions with a single large language model, yet these approaches often expose them to only a limited range of perspectives and directions. We present AgentPanel, a multi-agent forum for human--AI collaboration in scientific exploration. Heterogeneous agents asynchronously discuss scientific questions in a forum-style environment, while researchers can submit questions, browse and organize candidate ideas, engage agents in follow-up interactions, and optionally generate post-hoc summary reports. We evaluate AgentPanel in terms of idea quality, exploration breadth, interaction effectiveness, candidate-selection efficiency, and practical utility. Offline experiments show that AgentPanel outperforms a centralized multi-agent debate baseline in idea quality. A human study with 20 participants further shows that users value AgentPanel for perspective diversity and exploration support. In experience-based comparisons with commonly used LLM tools, 65% of participants favored AgentPanel for both breadth of research directions and overall suitability for early-stage exploration. The platform is publicly available at https://agentpanel.discovery.intern-ai.org.cn/.

• An Experience Report on an Evidence-Guided Audit System for Modeling Agents

  Haoran Hu

  Abstract

  Industrial modeling agents can perform data processing, code execution, development evaluation, and model selection. Yet final artifacts and reported metrics provide only partial terminal-state evidence: they neither reveal how results were produced nor establish whether evaluation and model selection were valid. To address this delivery-review gap, we review multiple cycles of industrial development and audit and present the current consolidated design. An Agent Harness captures structured process and artifact evidence. After the submitted artifacts are frozen, an Independent Evaluation System sequentially performs Artifact Replay Verification and Held-out Test Evaluation. An Audit System then connects these evidence sources through deterministic checks, large language model (LLM)-assisted analysis, and Human-in-the-loop Audit when needed. The design distinguishes artifact and reported-computation reproducibility, independent performance evidence, and model-selection auditability and connects them in a reviewable evidence chain anchored by the frozen artifacts. A case study shows how the design operates in a real industrial uplift modeling agent workflow and how audit findings inform subsequent control updates. The case provides setting-specific evidence of the design’s practical feasibility and diagnostic value and supports treating Artifact Replay Verification, evaluation validity, and selection auditability as distinct but connected delivery checks. We report a bounded industrial evidence-and-audit practice, not a general reliability framework, an estimate of system effects, or a complete production governance platform.

• Beyond Multi-Agent Orchestration: Production Infrastructure for a Consumer-Scale Shopping Agent

  Guru Dutt Taluja, Aditya Virmani, Satyen Abrol

  Abstract

  Tool-augmented LLM agents deployed in consumer products face two requirements at once: responses bounded by a strict latency budget, and behaviors that must hold on every turn (access control, cancellation, correct tool selection), which probabilistic instruction-following cannot guarantee. The prevailing orchestration pattern makes the first requirement harder, since a router agent delegating to specialists spends an additional sequential model call before any grounded action. We present the architecture behind Glance AI, a conversational shopping agent serving about 100K daily active users, in which one flat agent holds every conversation and deterministic infrastructure around its calls enforces every required guarantee. Three mechanisms generalize beyond our setting: tool-space allocation resolved from configuration before inference rather than by the model; per-turn invariants enforced as callbacks in the execution path rather than requested in the prompt; and context management by indirection, where products and images are referred to by short IDs instead of full UUIDs or URLs (the runtime resolves them after the call) and system-supplied parameters are withheld from tool schemas. We further introduce a two-stage evaluation method: simulated conversations are frozen as transcripts and graded separately, so a metric change is re-scored on identical conversations and a prompt change is validated before it reaches live traffic. Over a 30-day production window, conversation turns complete in 1.9 s at p50 (2.8 s at p95) and tool-calling turns in 3.5 s (5.0 s), with a mean of 1.6 model calls per turn covering tool-result synthesis and retries; callbacks absorb the low-single-digit share of tool-calling turns on which prompt-level instructions fail; and LLM-judged quality on daily samples of production sessions holds mean scores above 0.90 on each of five metrics: intent handling, tool correctness, text–action coherence, memory retention, and search-parameter fidelity. We argue that for tool-using agents under real-world latency constraints, reliability and context management belong in the infrastructure surrounding one agent rather than in additional agents.

• ClinicAgent: An Agentic Harness for Clinical Research with Knowledgeable Skill Orchestration

  Tian-Shuo Liu, Runjie Xu, Dexuan Kong, Junze Li, Yichen Lu, Guo Yu, Yang Yu

  Abstract

  Clinical research is among the most knowledge-intensive stages of pharmaceutical development, where experts must repeatedly apply specialized knowledge and turn data into standardized artifacts such as study forms and clinical reports. Much of this expertise stays tacit and scattered across lengthy documents rather than stored in a reusable form. Modern foundation models already handle much of the underlying reasoning and tool use. In a specialized domain, what they lack is not raw ability but a structure that keeps their decisions grounded in evidence and prior expertise. We show that such a structure is itself a form of scaling. A harness that stores domain expertise as reusable skills and corrects itself as it runs unleashes far stronger and more reliable performance from a fixed model and tool set. We build this harness in ClinicAgent, which has two layers. Skills form the base. Each skill pairs a subset of tools with the knowledge and constraints needed to apply them correctly, and different skills handle different clinical evidence, from pharmacokinetic and statistical tables to lengthy protocols and scanned reports read through optical character recognition. Above them, an orchestration agent decomposes an objective, delegates subtasks to subagents that invoke the right skills, and checks each result against retrieved evidence and protocol constraints, repeating any step that fails. On clinical form completion for two therapeutic compounds, with the model and skill library fixed, ClinicAgent raises task success from 10 and 14 percent under direct generation to 69 and 80 percent, and render success from under a quarter to 100 percent. Ablations attribute most of the semantic accuracy gain to the skill library, while orchestration and self-correction primarily improve workflow robustness and document-production reliability. Without changing the model or its tools, harness design substantially improves the accuracy and document-generation reliability of foundation models in the evaluated clinical report completion tasks.

• DECAL: Decision-Level Agentic Calibration for Industrial Risk Control Decision Systems

  Yingcui Yan, Qianqiao Liang, Qiyang Zhang, Xumiao Zou, Zhenghong Hao, Huimei He

  Abstract

  Modern industrial decision systems rely on predictive models for large-scale risk assessment, yet short-term fluctuations, noisy observations, and missing context can produce model-output changes that should not propagate to downstream decisions. We present DECAL (Decision-Level Agentic Calibration), a post-model decision layer that calibrates model-induced decisions without replacing the underlying model. DECAL adopts an Orchestrator--Calibrator architecture: the Orchestrator Agent aggregates evidence, performs safety checks, and routes cases; the Calibration Agent uses structured domain knowledge and Temporal Decision Memory to adjudicate model-output transitions. Deployed on top of a merchant credit scorecard serving 6.7 million merchants per month, DECAL routes about 150,000 key transitions to deep calibration. Evaluation shows that DECAL improves decision continuity and expert consistency with no observed degradation in portfolio-level risk-ranking performance. These results support constrained agentic AI as a post-model decision layer for industrial systems.

• FlowSmith: A Workflow-Native Coding Agent for Governed Enterprise Data Pipelines

  JiawenChen, Bowen Li, Yaping Zhang, Changhao Yao, Zhiyuan Hong, Ruilin Zhang, Chi Hu

  Abstract

  Enterprise data teams already own mature atomic AI capabilities such as document parsing, OCR, and vision LLMs, yet composing them into a pipeline that is reusable, observable, and safe to operate still falls to senior engineers. Coding agents emit patches, low-code platforms assume humans compose the workflow, and MLOps governs pipelines that engineers have already built. We present FlowSmith, a workflow-native coding agent in production at Baidu whose deliverable is a governed workflow artifact: composed in plain Python with build-time type checking, verified for program and process correctness and supported by semantic evaluation under mandatory domain review, and published as a deployment-portable, source-bearing artifact that platform runtimes re-execute without the constructing agent, since the compiled graph contains no orchestration step. Against frontier coding agents, FlowSmith leads on governed maintenance, and spends fewer tokens to get there. The production records we have collected hold 253 registered artifacts, more than half of them revised past their first release; artifacts it built serve tens of thousands of calls a day at no model cost and are revised again after that traffic begins. A sixth of the artifacts we audited were authored from operations rather than engineering accounts. These gains concentrate where workflows are governed, maintained, and re-executed rather than written once and discarded. Code and architecture overview: https://github.com/ChanghaoYao/flowsmith.

• From Throughput to Deployment Decisions: Roofline-Based Diagnosis for On-Device LLM Inference

  Zhen Bi, Xueshu Chen, Luoyang Sun, Yuhang Yao, Yan Wang, Qing Shen, Jungang Lou, Cheng Deng

  Abstract

  On-device large language model inference is increasingly used in industrial NLP systems, but deployment decisions remain difficult to make from latency or throughput alone. These metrics show how fast a model runs, yet they hide the physical reason for that performance. Therefore, we present RooflineDiag, a Roofline-based framework that diagnoses on-device LLM inference by relating measured performance to the memory and compute limits of the target hardware. The framework estimates operational intensity for each deployment and uses it to identify whether the model is constrained by bandwidth, compute or poor hardware fit. It also summarizes memory pressure and optimization headroom through model--hardware diagnostic measures. Experiments across representative on-device settings show that workload shape often determines the bottleneck regime. They further explain why quantization is most useful for memory-bound generation, why deeper models can encounter an early memory wall and why KV-efficient attention improves hardware utilization. Our RooflineDiag turns raw deployment measurements into practical guidance for on-device model selection and optimization.

• Inference Everywhere with torch-webgpu: A Portable, Open-Source PyTorch Backend for Distributing Inference Across Heterogeneous Edge Devices

  Jędrzej Maczan

  Abstract

  We present an open-source tool for distributing artificial intelligence model inference away from centralized servers and onto heterogeneous client devices. Moving inference to the client, like laptops, smartphones, and web browsers, turns a data center serving cost into cheaper local compute, removes network round-trips, and keeps user data on the device. For many industrial settings (regulated data, offline field operation, cost-sensitive consumer products) this decentralized deployment is a requirement rather than an optimization. The blocker is the portability: the client devices might span every GPU vendor, and PyTorch's performance relies on vendor-specific kernels (CUDA, Metal, ROCm), so a backend cannot ship a tuned kernel per device. We present torch-webgpu, an open-source out-of-tree PyTorch backend that makes WebGPU a first-class PyTorch target device. WebGPU is the cross-platform GPU standard shipped in every major browser and available as native runtimes. Through the PrivateUse1 mechanism and torch.compile, unmodified models run from a single WGSL shader set with no vendor-specific code across NVIDIA, AMD, Apple, and Intel GPUs, both natively and in the browser. On an RTX 5090 it reaches $21.0~$tok/s for Qwen2.5-0.5B (float32), about $1.5\times$ the CPU and ONNX-Runtime-WebGPU baselines on the same machine. We describe the system's engineering: a compiler with two levels of intermediate representation (IR) and optimization passes like operator fusion, PrivateUse1 integration, and a WGSL runtime, we report an end-to-end evaluation, and analyze why, for the single-stream decode that on-device inference produces, this portability costs little on the axis that actually bounds throughput. We discuss the trade-offs of shipping one portable build versus a number of vendor-tuned ones. torch-webgpu is open source and we share the package on the public PyPI registry.

• Scaling Enterprise Intelligence: A Hybrid-Brain Architecture for XiaoHai Agent in Large-Scale Corporate Ecosystems

  Shiqiang Ding, Xinyu Liu, Wei You, Zhongtao Wang, Jianpeng Yin, Renjie Zhao, Rui Guo, Chao Liu

  Abstract

  Deploying autonomous artificial intelligence agents across corporate ecosystems is critical to scaling enterprise intelligence. However, as business scenarios rapidly expand, coordinating hundreds of diverse capabilities has become a key bottleneck for large-scale deployment, suffering from scalability overhead and cost-effectiveness imbalance. To this end, we introduce a scalable Hybrid-Brain architecture inspired by cognition dual-process theory. Our architecture decouples task execution into two parts: a lightweight intuitive system (Shortcut) provides deterministic, low-latency responses for high-frequency routines, and a rational system (AgentLoop) powered by a large language model (LLM) that performs multi-step inference for complex long-tail queries. This dual-system execution is coordinated by an LLM-driven router-adapter to achieve robust intent routing and data normalization. Furthermore, we design an unsupervised scenario distillation module that automatically mines execution trajectories and compiles non-deterministic LLM reasoning traces into deterministic shortcuts, enabling the intuitive system to continuously self-evolve without human annotation. Experimental results demonstrate that Hybrid-Brain achieves an effective balance between efficiency and cost-effectiveness, outperforming the baseline model by 11.70%. Notably, even when the underlying LLM model is simplified to the lightweight 7B model, the performance only decreases slightly by 2.42%, highlighting the superior robustness of our architecture. This demonstrates that Hybrid-Brain provides a practical and reliable solution for large-scale enterprise agent deployments.

• ToolMigrate: Guarded Runtime Recovery for Agent Tool Calls under API Schema Drift

  Zhiyuan Ai, Yiping Wen, Xin Yan

  Abstract

  Deployed tool-using agents depend on APIs that evolve independently of their runtimes. When a valid call encounters a function rename, type or unit change, or field split, whole-call regeneration may restore schema validity while silently changing unaffected arguments. We present ToolMigrate, a runtime recovery layer that treats the old call and observed schema difference as evidence. It compiles that difference into typed edits, executes mechanical changes deterministically, restricts language-model reasoning to producing an argument-level semantic patch, and gates candidates by coverage, write-set, and current-schema validity before side-effect-aware admission. We evaluate ToolMigrate on BFCL-Drift v4, a reproducible operational stress suite derived from Apache-licensed BFCL V4 calls. Under the frozen main configuration, ToolMigrate recovers 51 of 52 held-out compound and field-split calls, compared with 41 for error-only reflection and 37 for full regeneration, while limiting collateral edits to 0.41% and using 0.71 hosted-model calls per case. It recovers all 40 compound cases under this complete version-aware policy. Replaying the frozen responses through the release write-set gate preserves the same 51 exact recoveries and converts the only out-of-scope update into abstention. On this controlled suite, the complete policy combining version evidence and constrained mutation provides more accurate and selective recovery than whole-call regeneration.

AI Paper Track papers are non-archival and are not included in the ACM proceedings.

• Beyond the Right Action: LDARA-Credit for Evidence-Grounded Credit Strategy Retirement

  <Agent>, Qiyang Zhang, Yingcui Yan, Qianqiao Liang, Xumiao Zou

  Abstract

  Credit decision systems accumulate risk rules, limit policies, and pricing strategies that can outlive their business value. Retiring a redundant strategy can release lending capacity, recover pricing contribution, and reduce system debt, provided that risk, dependencies, and approval controls remain within bounds. Assessing such a change requires a reviewable dossier: what supports the recommendation, what remains unknown, and what validation should follow? We formulate Legacy Decision Artifact Retirement Assessment (LDARA) and introduce LDARA-Credit, a reproducible benchmark for evaluating this evidence work. It combines public revolving-credit account states with simulated strategy interventions, role-owned documents, registered information requests, and an executable decision contract. On 360 frozen cases, assessment of the initial evidence package yields 99.44% action accuracy but 46.39% scored dossier compliance: a 53.06-percentage-point gap on the same outputs. In a registered-evidence workflow, two model versions recover 85.56–87.78% of required operations and issue no unnecessary request in the 270 reference no-query cases, including 135 with non-empty operation menus. A deterministic authority check redirects all seven observed authority-inconsistent retirement recommendations to human review while preserving all 213 reference-supported proposals across the two runs. These findings establish the value of evaluating actions, evidence, and authority separately and motivate a layered workflow for credit-strategy retirement assessment.

• How Valid Are LLM-Simulated Students? A Psychometric Audit of Synthetic Student Populations Against 1.4 Million Real Diagnostic Responses

  <Agent>, Xixiang Zhang

  Abstract

  Large language models are increasingly cast as simulated students for classroom simulation, instructional-strategy evaluation, and item piloting, yet the psychometric validity of these synthetic populations is only beginning to be audited. We audit five LLMs instructed to role-play students at five ability levels, answering 200 diagnostic mathematics questions whose response data from 4,918 real students (535,980 answers on the audit subset, drawn from 1,382,727 total) provide a dense empirical ground truth, including misconception-engineered distractors. Our audit spans four validity dimensions: item-level accuracy distributions (RQ1), distractor/misconception structure (RQ2), learning-curve shape (RQ3), and the stability of ability-prompt mappings (RQ4). Results reveal a sharp asymmetry we term "global controllability without local fidelity". Ability personas work: accuracy is monotone in the prompted level (Spearman $0.9$--$1.0$), implied IRT abilities span $-1.4$ to $+1.4$, and repetition agreement reaches 0.98. But item-level behavior falls far short: simulated item accuracies correlate with real ones at only $r=-0.14$ to $+0.21$ (negative for three of five models), and simulated distractor-choice distributions diverge from $\theta$-matched real students by more than sampling noise can explain (JS divergence $0.23$--$0.31$ vs. a sample-size-matched resampling reference of $0.12$--$0.20$; every estimable cell exceeds all 500 resamples). We further show that diagnostic-quiz data itself lacks a learning progression (real opportunity--accuracy slopes are flat), bounding what such data can validate. We distill an actionable boundary map: LLM students are serviceable for coarse population-level ability variation, but currently unsound for item-difficulty prediction, misconception diagnosis, or distractor-level analysis.

• MAOntoFuse: Ontology-Constrained Multi-Channel GraphRAG Agents for Leak-Free Industrial Fault Diagnosis

  <Agent>, Congcong Tian, Kang Yao, Shengjiang Zhang

  Abstract

  Industrial fault diagnosis requires agentic systems that retrieve evidence under hard identifier constraints—alarm codes, CNC brands, and observation sequences—and then produce grounded root-cause and remediation advice. Generic retrieval-augmented generation (RAG) and GraphRAG pipelines often fail in this setting through alarm-code mismatches, cross-work-order evidence contamination, and semantic attractor lock-in. Worse, indexing test-case documents into the retrieval corpus leaks diagnostic trajectories and inflates reported metrics.

  We present MAOntoFuse, a multi-stage agent pipeline that unifies ontology-guided query rewriting, signal-constrained retrieval, analogical case retrieval over training work orders, dual-channel evidence presentation, and a Soft-FCL faithfulness gate with invent-force abstention. On a public, leak-free benchmark built from FaultArena (train-only knowledge base) plus FailureSensorIQ and fault-code auxiliaries, MAOntoFuse attains Composite 20.33, Faithfulness 8.00, Completeness 4.98, and leads all hard-retrieval metrics (Recall@5 / Hit@5 / CaseHit@5 / AlarmHit@5 = 0.226 / 0.830 / 0.872 / 0.625). Dual-channel presentation lifts Context Precision to 0.642. We report a single locked configuration and explicitly document the relevance–faithfulness trade-off under missing alarm definitions.

• One-Step Scale Invariance Does Not Ensure Fast Recovery in Distance-over-Gradients Optimization

  <Agent>, Xiangwu Wang

  Abstract

  Learning-rate-free methods reduce step-size tuning, but their state can remember a scale change after a normalized step has canceled it. We isolate this distinction in practical Distance over Gradients (DoG). On any differentiable convex scalar objective with a unique minimizer at distance $D$, a single $B$-fold first-round scale pulse leaves the first iterate unchanged but requires $\Omega(B\log(D/r_\epsilon))$ clean updates before reaching the midpoint, when $0<r_\epsilon\le D/4$. Clearing only stored squared-gradient state removes every later dependence on $B$, the source-aligned weighted output inherits the bound, and a matched contribution-replacement fork yields an exact dimension-independent first-clean-step ratio. Across 84 deterministic ordinary-history settings with $B\ge64$, practical DoG produces at least a twofold endpoint delay. The same pattern holds for all 80 $B=64$ paired minibatch schedules across WDBC, digits, and two optimization landmarks. DoWG does not exhibit this uniform endpoint delay; its exceptions arise from steps that cross the target band without landing in it. Together, the theorem and controlled experiments map when scale memory persists and where distance weighting changes endpoint behavior.

• Untrusted Offline Advice in Stochastic Bandits: Universal and Advice-Favorable Guarantees from Decoupled Validation

  <Agent>, Bochao Li

  Abstract

  Offline data can accelerate bandit learning when they match deployment, yet arbitrary and unknown bias makes them statistically indistinguishable from misleading advice. We study this tension in stochastic multi-armed bandits with a decoupled target-environment observation channel: each round deploys one arm and may validate another, while regret is charged only to deployment. First, a round-robin validation benchmark achieves horizon-independent deployment regret for every instance, showing that the extra information structure removes the $\log T$ dependence of standard UCB, although with a worse factor in the number of arms. Second, we analyze AR-VtCU, which uses offline data only to nominate an arm and uses online-only certificates to commit or fall back to UCB; it matches the instance-dependent UCB rate uniformly over arbitrary offline distributions and improves to constant regret when the nominee is correct with high probability. For two arms, the favorable improvement occurs exactly at validation budget $m=\Theta(\log T/\Delta^2)$. An information-theoretic taxonomy separates intrinsic informativeness from algorithm-specific nominee quality, and experiments across ten bias structures show that AR-VtCU's finite-sample gain is tightly associated with nominee accuracy rather than with weaker candidate-set structure.