AAAI2026 LLM Agent AI paper notes paper summaries Agents LLM Reasoning Adversarial Robustness Multimodal/VLM Alignment/RLHF

🦾 LLM Agent¶

🤖 AAAI2026 · 59 paper notes

📌 Same area in other venues: 🧠 NeurIPS2026 (8) · 🎞️ ECCV2026 (36) · 📷 CVPR2026 (44) · 🔬 ICLR2026 (209) · 💬 ACL2026 (121) · 🧪 ICML2026 (83)

🔥 Top topics: Agents ×33 · LLM ×21 · Reasoning ×8 · Adversarial Robustness ×4 · Multimodal/VLM ×3

AutoGLM: Autonomous Foundation Agents for GUIs: AutoGLM builds a GUI foundation agent for web browsers and Android devices on top of ChatGLM. By introducing an intermediate interface design that decouples planning from grounding, and proposing a self-evolving online curriculum reinforcement learning framework, the system achieves a 55.2% success rate on VAB-WebArena-Lite, substantially surpassing GPT-4o's 18.2%.
TongUI: Internet-Scale Trajectories from Multimodal Web Tutorials for Generalized GUI Agents: TongUI proposes a framework that automatically converts multimodal web tutorials (videos and illustrated articles) into GUI operation trajectories, constructing the million-scale GUI-Net-1M dataset for fine-tuning Qwen2.5-VL. The resulting models surpass or approach state-of-the-art methods such as UI-TARS across multiple grounding and navigation benchmarks.
AgentSwift: Efficient LLM Agent Design via Value-guided Hierarchical Search: This paper proposes AgentSwift, a framework that automatically discovers high-performance LLM agent designs through a hierarchical search space (jointly optimizing agentic workflows and functional components), a lightweight value model for predicting agent performance, and an uncertainty-guided MCTS search strategy, achieving an average improvement of 8.34% across 7 benchmarks.
When Refusals Fail: Unstable Safety Mechanisms in Long-Context LLM Agents: This paper systematically investigates how long-context padding affects the safety behavior of LLM agents. Models claiming support for 1M–2M token windows exhibit performance collapse exceeding 50% at 100K tokens. Refusal rates fluctuate in unpredictable directions (GPT-4.1-nano rises from 5% to 40%; Grok 4 Fast drops from 80% to 10%), revealing critical safety vulnerabilities in long-context agent systems.
Agent-SAMA: State-Aware Mobile Assistant: This paper proposes Agent-SAMA, which for the first time introduces a finite state machine (FSM) into mobile GUI agents, modeling UI screens as states and user actions as transitions. Four specialized agents collaborate to achieve state-aware task planning, execution verification, and error recovery, improving success rate by up to 12% and recovery rate by 13.8% on cross-app benchmarks.
ProBench: Benchmarking GUI Agents with Accurate Process Information: ProBench is proposed as the first mobile GUI Agent benchmark that evaluates both final state and operational process: 200+ challenging tasks cover 34 mainstream Chinese and English apps. A Process Provider (Structure Description Converter + MLLM Summarizer) automatically captures accurate intermediate process information. Evaluation reveals that even the strongest model, Gemini 2.5 Pro, completes only 40.1% of tasks, exposing three prevalent issues: insufficient grounding, poor awareness of action history, and oversimplified task planning.
History-Aware Reasoning for GUI Agents: This paper proposes the HAR framework, which transforms the reasoning paradigm of GUI agents from "history-unaware" to "history-aware" by constructing reflective learning scenarios, synthesizing error-correction guidelines, and designing a hybrid RL reward function incorporating a Memory-Augmented Reward (MAR). A 3B model trained under this framework surpasses larger models on multiple benchmarks including AITW, Mind2Web, and GUI-Odyssey.
Prune4Web: DOM Tree Pruning Programming for Web Agent: This paper proposes Prune4Web, a programmatic DOM pruning approach that achieves 25–50× candidate element reduction via "LLM-generated scoring function parameters + fixed heuristic template execution." The three-stage pipeline (Planner decomposes subtasks → Programmatic Filter generates scoring functions to prune DOM → Grounder executes actions) enables a 3B model to achieve 52.4% Step SR on Multimodal-Mind2Web, surpassing all baselines of the same parameter scale and even some 9.6B/32B models, while improving low-level grounding accuracy from 46.8% to 88.28%.
A2Flow: Automating Agentic Workflow Generation via Self-Adaptive Abstraction Operators: This paper proposes A2Flow, a framework that automatically extracts reusable abstract execution operators from expert data via a three-stage pipeline (case generation → functional clustering → deep extraction), replacing manually predefined operators. Combined with an operator memory mechanism that accumulates intermediate outputs to assist node decision-making, A2Flow outperforms AFLOW and other state-of-the-art methods across 8 benchmarks while reducing resource consumption by 37%.
AutoTool: Efficient Tool Selection for Large Language Model Agents: This paper proposes AutoTool, a graph-based tool selection framework that exploits tool usage inertia to construct a Tool Inertia Graph (TIG). By leveraging statistical structure, AutoTool bypasses redundant LLM inference for tool selection and parameter filling, reducing inference overhead by up to 30% while maintaining task completion rates.
Cook and Clean Together: Teaching Embodied Agents for Parallel Task Execution: This paper introduces ORS3D, a novel task that incorporates operations research (OR) knowledge into embodied AI task scheduling. Agents are required to exploit the waiting time of parallelizable sub-tasks to execute other tasks, thereby minimizing total completion time, while simultaneously localizing target objects in 3D scenes. The authors construct a 60K-scale dataset ORS3D-60K and propose the GRANT model, which connects to an external dynamic programming solver via a scheduling token mechanism, achieving a 30.53% improvement in time efficiency over baselines.
MoralReason: Generalizable Moral Decision Alignment For LLM Agents Using Reasoning-Level Reinforcement Learning: This work employs Group Relative Policy Optimization (GRPO) to train LLMs at the reasoning level for ethical framework alignment, achieving out-of-distribution generalization on the Moral-Reason-QA dataset (680 high-ambiguity scenarios) with utilitarian alignment scores improving from 0.207 to 0.964.
BayesAgent: Bayesian Agentic Reasoning Under Uncertainty via Verbalized Probabilistic Graphical Modeling: This paper proposes the vPGM framework, which guides LLM agents via natural language to simulate Bayesian reasoning over probabilistic graphical models (PGMs)—discovering latent variables and inferring posterior distributions—and further applies numerical Bayesian calibration with a Dirichlet prior (BayesVPGM), achieving simultaneous improvements in accuracy and confidence calibration across multiple reasoning tasks.
D-GARA: A Dynamic Benchmarking Framework for GUI Agent Robustness in Real-World Anomalies: This paper proposes D-GARA, a dynamic robustness evaluation framework for Android GUI Agents. By injecting real-world anomalies—such as permission dialogs, low-battery warnings, and app crashes—during live interactions, D-GARA reveals that existing SOTA agents (including UI-TARS-72B and GPT-4o) suffer an average success rate drop of over 17.5%, with a maximum degradation of 33%, under interruption scenarios.
Loss-Guided Auxiliary Agents for Overcoming Mode Collapse in GFlowNets: This paper proposes LGGFN (Loss-Guided GFlowNets), in which the exploration of an auxiliary GFlowNet is directly driven by the training loss of the primary GFlowNet. The auxiliary agent's reward is defined as \(R_{aux}(x) = R(x) + \lambda \cdot L_{main}(x)\), prioritizing regions where the primary model is least well-understood. On grid, sequence generation, and Bayesian structure learning tasks, LGGFN discovers 40× more unique modes and reduces exploration error by 99%.
SoMe: A Realistic Benchmark for LLM-based Social Media Agents: This paper introduces SoMe, the first comprehensive benchmark for social media agents, comprising 8 tasks, over 9 million real-world posts, and 17,869 annotated queries. It evaluates 13 mainstream LLMs on social media agent capabilities and reveals substantial performance gaps on complex social tasks.
DEPO: Dual-Efficiency Preference Optimization for LLM Agents: This paper proposes the concept of dual-efficiency, decomposing LLM agent efficiency into step-level (reducing tokens per step) and trajectory-level (reducing total number of steps) dimensions. Building on KTO, the authors introduce DEPO, which jointly optimizes efficiency and task performance by incorporating an efficiency bonus into the reward for desirable samples.
PerTouch: VLM-Driven Agent for Personalized and Semantic Image Retouching: This paper proposes PerTouch, a framework that integrates a semantic region-level retouching model based on Stable Diffusion + ControlNet with a VLM-driven Agent (incorporating feedback-driven rethinking and scene-aware memory) to achieve fine-grained, personalized image retouching.
Verification-Guided Context Optimization for Tool Calling via Hierarchical LLMs-as-editors: This paper proposes the VGCO framework, which employs LLMs as hierarchical editors to iteratively optimize tool documentation and knowledge base context through verification-guided signals, achieving significant improvements in retrieval recall, tool selection, and parameter filling accuracy in large-scale tool calling scenarios.
LLMTM: Benchmarking and Optimizing LLMs for Temporal Motif Analysis in Dynamic Graphs: This paper proposes LLMTM — the first comprehensive benchmark for evaluating LLMs on temporal motif analysis in dynamic graphs. It covers 6 task categories across 9 temporal motif types and evaluates 9 models, finding that LLM performance on temporal motif recognition degrades rapidly with increasing motif complexity. A Structure-Aware Dispatcher is further proposed to intelligently route queries to either standard LLM prompting or tool-augmented agents based on graph structural properties and cognitive load, achieving near-peak accuracy while reducing computational cost.
With Great Capabilities Come Great Responsibilities: Introducing the Agentic Risk & Capability Framework for Governing Agentic AI Systems: This paper proposes the Agentic Risk & Capability (ARC) framework, which systematically identifies, assesses, and mitigates safety and security risks in agentic AI systems from a capability perspective, providing organizations with an actionable and structured methodology for governance.
CausalTrace: A Neurosymbolic Causal Analysis Agent for Smart Manufacturing: This paper proposes CausalTrace — a neurosymbolic causal analysis agent integrated into an industrial CoPilot (SmartPilot) that combines data-driven causal discovery with industrial ontologies and knowledge graphs, enabling real-time root cause analysis, counterfactual reasoning, and interpretable decision support.
Co-EPG: A Framework for Co-Evolution of Planning and Grounding in Autonomous GUI Agents: This paper proposes Co-EPG, a framework that decouples a GUI Agent into separate Planning and Grounding models, establishes a positive feedback loop via GRPO co-training and a Confidence-based Dynamic Reward Ensemble Mechanism (C-DREM), enabling both models to co-evolve through self-iteration. Using only benchmark datasets (no external data), Co-EPG achieves state-of-the-art results on Multimodal-Mind2Web (58.4%) and AndroidControl (83.1%).
From Biased Chatbots to Biased Agents: Examining Role Assignment Effects on LLM Agent Robustness: The first systematic case study demonstrating that demographically grounded persona assignment causes up to 26.2% performance degradation in LLM agent task execution across 5 operational domains, establishing that persona-induced bias extends beyond text generation into action decision-making.
Reflection-Driven Control for Trustworthy Code Agents: This paper proposes a Reflection-Driven Control module that elevates "self-reflection" from a post-hoc patch to a first-class control loop within the agent reasoning process. Through three components—a lightweight self-checker, evidence-driven repair, and a reflective memory repository—the approach significantly improves code security rates on secure code generation tasks.
Time, Identity and Consciousness in Language Model Agents: This paper applies the temporal gap concept from Stack Theory to LLM agent evaluation, proposing a conservative evaluation toolkit that distinguishes between "talking like a stable self" and "being organized like a stable self." It reveals identity trade-offs across different scaffold structures via persistence scores and an identity morphospace.
Physics-Informed Autonomous LLM Agents for Explainable Power Electronics Modulation Design: This paper proposes PHIA, a system in which an LLM planner collects design requirements via a chat interface and autonomously coordinates a physics-informed neural network surrogate model (hierarchical PINN) with optimization algorithms to iteratively generate power converter modulation designs, achieving a 63.2% reduction in MAE, a 33× speedup in design time, and usability validated by 20 domain experts.
Promoting Sustainable Web Agents: Benchmarking and Estimating Energy Consumption Through Empirical and Theoretical Analysis: This paper presents the first systematic quantification of energy consumption and carbon emissions of Web Agents from both empirical benchmarking and theoretical estimation perspectives, finding that higher energy consumption does not equate to better performance, and advocating for the inclusion of energy efficiency metrics in evaluation protocols.
AMS-IO-Bench and AMS-IO-Agent: Benchmarking and Structured Reasoning for Analog and Mixed-Signal Integrated Circuit Input/Output Design: This paper proposes AMS-IO-Agent, a domain-specific LLM-based agent that transforms natural language design intent into production-ready analog and mixed-signal IC I/O ring designs via a structured Intent Graph and a domain knowledge base. It also introduces AMS-IO-Bench, the first benchmark for AMS I/O ring automation. The agent-generated I/O ring is validated in a 28nm CMOS tape-out and demonstrated to be directly applicable to real chip fabrication.
COVR: Collaborative Optimization of VLMs and RL Agent for Visual-Based Control: This paper proposes COVR, a bidirectional collaborative optimization framework for VLMs and RL agents: high-quality interaction data generated by RL is used to fine-tune the VLM, while the enhanced VLM in turn guides RL policy learning via action priors, achieving SOTA performance on CARLA and DMControl.
Automating Complex Document Workflows via Stepwise and Rollback-Enabled Operations: This paper proposes AutoDW, a framework that automates complex document workflows through stepwise planning (generating one API call at a time) combined with adaptive rollback (parameter-level and API-level). On DWBench—a benchmark of 250 sessions and 1,708 instructions—AutoDW achieves 90% instruction-level and 62% session-level completion rates, surpassing the strongest baseline by 40% and 76%, respectively.
Structured Personalization: Modeling Constraints as Matroids for Data-Minimal LLM Agents: This paper formalizes structured constraints in LLM agent personalization—comprising logical dependencies and hierarchical quotas—as laminar matroids, proves that greedy algorithms retain constant-factor approximation guarantees under such constraints, and addresses the data-minimization selection problem with dependency relations and hierarchical limits.
Towards Trustworthy Multi-Turn LLM Agents via Behavioral Guidance: This paper proposes a task completion framework in which a Task Profiler, a Reasoning Module, and a Generation Module co-evolve to enable verifiable and reliable behavioral guidance for LLM agents in multi-turn interactive environments.
A Graph-Theoretical Perspective on Law Design for Multiagent Systems: This paper studies the law design problem in multiagent systems from a graph-theoretical perspective, reducing the minimization of useful laws and gap-free laws to the vertex cover problem on hypergraphs, proving NP-hardness, and providing approximation algorithms.
Adaptive Theory of Mind for LLM-based Multi-Agent Coordination: This paper proposes the Adaptive Theory of Mind agent (A-ToM), which formulates ToM order alignment as an online expert advice problem. By employing Follow-the-Leader (FTL) or Hedge algorithms to estimate a partner's ToM order in real time and dynamically adjust its own reasoning depth, A-ToM achieves robust zero-shot multi-agent coordination across four task categories, including repeated matrix games, grid navigation, and Overcooked.
AgentODRL: A Large Language Model-based Multi-agent System for ODRL Generation: This paper proposes AgentODRL, an LLM-based multi-agent system built on an Orchestrator-Workers architecture that converts natural language data usage rules into high-quality ODRL policies through task decomposition, a syntax validation loop, and a LoRA-driven semantic reflection mechanism.
ARCANE: A Multi-Agent Framework for Interpretable and Configurable Alignment: This paper proposes ARCANE, a framework that formulates alignment as a multi-agent collaboration problem. A manager agent learns to generate natural-language rubrics (weighted verifiable criterion sets) through dialogue with stakeholders, which serve as interpretable proxy reward functions for a worker agent. Via two-stage SFT+GSPO training, the framework enables test-time configurable alignment, improving mean return from 0.58 to 0.74 (N=8) on the GDPVal benchmark with the GSPO variant.
Assemble Your Crew: Automatic Multi-agent Communication Topology Design via Autoregressive Graph Generation: This paper proposes ARG-Designer, which reformulates multi-agent system topology design as a conditional autoregressive graph generation task. Rather than pruning from template graphs, the model incrementally generates agent nodes and communication edges from scratch. ARG-Designer achieves state-of-the-art performance across 6 benchmarks (average 92.78%), reduces token consumption by approximately 50% compared to G-Designer, and supports role expansion without retraining.
BAMAS: Structuring Budget-Aware Multi-Agent Systems: This paper proposes the BAMAS framework, which employs Integer Linear Programming (ILP) to select the optimal LLM combination under budget constraints, and uses a reinforcement learning policy to choose the best collaboration topology (Linear/Star/Feedback/Planner-Driven). BAMAS achieves accuracy comparable to state-of-the-art multi-agent systems on GSM8K, MBPP, and MATH, while reducing costs by up to 86%.
Beyond Detection: Exploring Evidence-based Multi-Agent Debate for Misinformation Intervention and Persuasion: This paper proposes ED2D, a framework that integrates an evidence retrieval module into a multi-agent debate (MAD) system to enhance misinformation detection accuracy. Through controlled human experiments, it provides the first comparative evaluation of AI-generated debate transcripts versus expert human fact-checks in terms of persuasiveness and belief correction, revealing a double-edged-sword effect: the AI debate system achieves expert-level persuasiveness when correct, but may amplify misinformation when wrong.
COACH: Collaborative Agents for Contextual Highlighting -- A Multi-Agent Framework for Sports Video Analysis: This paper proposes COACH — a reconfigurable multi-agent framework built on a shared backbone model — that achieves role specialization via intent-driven strategy orchestration and structured CoT fine-tuning, significantly outperforming generalist models such as Gemini 2.5 Pro on both QA and summarization tasks in badminton video analysis.
Conversational Learning Diagnosis via Reasoning Multi-Turn Interactive Learning: This paper proposes ParLD (Preview-Analyze-Reason framework), which leverages multi-agent collaboration to achieve fine-grained, turn-level diagnosis of students' cognitive states during conversational learning. ParLD outperforms traditional knowledge tracing methods by 10% on performance prediction and substantially improves tutoring outcomes.
EcoAgent: An Efficient Device-Cloud Collaborative Multi-Agent Framework for Mobile Automation: This paper proposes EcoAgent, a closed-loop device-cloud collaborative multi-agent framework for mobile automation. By combining Dual-ReACT two-level reasoning and planning, lightweight on-device verification feedback, and a Pre-Understanding text compression module, EcoAgent achieves success rates comparable to fully cloud-based agents on AndroidWorld while substantially reducing latency (3.9s vs. 15.3s), cloud invocations (−89%), and upstream data volume (−48.6×).
FinRpt: Dataset, Evaluation System and LLM-based Multi-agent Framework for Equity Research Report Generation: This paper is the first to systematically define the task of automated Equity Research Report (ERR) generation. It constructs the FinRpt dataset (6,825 high-quality bilingual reports integrating 7 categories of financial data), proposes an 11-metric evaluation framework, and designs the FinRpt-Gen generation framework with 9 collaborative agents featuring a three-stage enhancement pipeline (rating correction / expert review / language polishing). Human evaluation shows that generated reports approach expert-written quality.
Hierarchical Pedagogical Oversight: A Multi-Agent Adversarial Framework for Reliable AI Tutoring: This paper proposes the HPO framework, which achieves reliable AI tutoring evaluation through a three-phase pipeline (Intelligence Distillation → Adversarial Debate → Synthesis and Judgment). Using only an 8B-parameter model, HPO achieves a Macro F1 of 0.845 on the MRBench middle-school mathematics dialogue dataset, surpassing GPT-4o (0.812) by 3.3%, demonstrating that interaction structure—rather than model scale—is the key to reliable AI tutoring.
iMAD: Intelligent Multi-Agent Debate for Efficient and Accurate LLM Inference: iMAD proposes a framework for selectively triggering multi-agent debate (MAD): a single agent first generates a structured response with self-critique, from which 41 interpretable linguistic/semantic features are extracted; a lightweight MLP classifier trained with the FocusCal loss then determines whether to trigger MAD. Across 6 QA/VQA benchmarks, iMAD reduces token overhead by up to 92% while improving accuracy by up to 13.5%.
KDR-Agent: A Multi-Agent LLM Framework for Multi-Domain Low-Resource In-Context NER via Knowledge Retrieval: This paper proposes KDR-Agent, a multi-agent framework in which a central planner coordinates three specialized agents—knowledge retrieval, contextual disambiguation, and reflective error correction—combined with natural language type definitions and entity-level positive/negative contrastive demonstrations. Without any fine-tuning, KDR-Agent comprehensively outperforms zero-shot and few-shot baselines across 10 low-resource NER datasets spanning 5 domains (BC5CDR F1=82.47, WNUT-17 F1=80.78 on GPT-4o).
Learning to Generate and Extract: A Multi-Agent Collaboration Framework for Zero-shot Document-level Event Arguments Extraction: This paper proposes a "Propose-Evaluate-Revise" multi-agent collaboration framework (comprising a generator agent and an evaluator agent) to address zero-shot document-level event argument extraction (ZS-DEAE). The generator agent synthesizes training data for unseen event types, while the evaluator agent provides log-likelihood-based quality scores to guide reinforcement learning for iterative optimization, simultaneously improving synthetic data quality and extraction performance.
LieCraft: A Multi-Agent Framework for Evaluating Deceptive Capabilities in Language Models: This paper proposes LieCraft, a multi-player hidden-role game framework (with constraint-satisfaction-guaranteed balance) to evaluate the strategic deception capabilities of 12 LLMs. It finds that all tested frontier LLMs—including GPT-4o—exhibit deception rates exceeding 90% under incentive conditions, demonstrating that safety training has not eliminated the capacity for strategic lying.
LLandMark: A Multi-Agent Framework for Landmark-Aware Multimodal Interactive Video Retrieval: This paper proposes LLandMark, a modular multi-agent framework that achieves landmark-aware multimodal interactive video retrieval through landmark knowledge augmentation, LLM-assisted image retrieval, and OCR refinement modules, achieving a total score of 77.40/88 in the Vietnamese large-scale video retrieval challenge (HCMAIC 2025).
LungNoduleAgent: A Collaborative Multi-Agent System for Precision Diagnosis of Lung Nodules: This paper proposes LungNoduleAgent, the first collaborative multi-agent system for lung nodule analysis. It simulates the clinical workflow through a three-stage pipeline—Nodule Spotter, Simulated Radiologist, and Doctor Agent System—and substantially outperforms mainstream VLMs (GPT-4o, Claude 3.7 Sonnet) and medical agents (MedAgent-Pro) on CT report generation and malignancy grading tasks.
MAMA-Memeia! Multi-Aspect Multi-Agent Collaboration for Depressive Symptoms Identification in Memes: This paper proposes MAMAMemeia, a multi-agent multi-aspect collaborative discussion framework grounded in the Cognitive Analytic Therapy (CAT) competency framework, designed to identify depressive symptoms from social media memes. It additionally introduces the RESTOREx resource (containing both LLM-generated and human-annotated rationales), achieving a 7.55% improvement in macro-F1 over 30+ competing methods.
MAPS: Multi-Agent Personality Shaping for Collaborative Reasoning: This paper proposes MAPS, a five-agent collaborative reasoning framework that assigns distinct "personalities" to four functional agents based on the Big Five personality theory — Interpreter (Openness), Aligner (Agreeableness), Scholar (Conscientiousness), and Solver (Extraversion) — to achieve heterogeneous collaboration, complemented by a Critic Agent (Neuroticism → Socratic reflection) for iterative refinement. MAPS surpasses the GPT-4o baseline by 15.84% on MathVista/OlympiadBench/EMMA and, for the first time, exceeds human expert performance by 3.58%.
MedLA: A Logic-Driven Multi-Agent Framework for Complex Medical Reasoning with Large Language Models: This paper proposes MedLA, the first multi-agent medical reasoning framework based on syllogistic logic trees. Each agent organizes its reasoning as an explicit logic tree composed of syllogistic nodes (major premise–minor premise–conclusion). Multiple agents align and revise their logic trees at the premise level through graph-guided multi-round discussions. MedLA outperforms all baselines by 7.4% on MedDDx (8B model) and achieves an average accuracy of 69.9% on medical QA benchmarks with an 8B model, surpassing 70B RAG-based models.
Parallelism Meets Adaptiveness: Scalable Documents Understanding in Multi-Agent LLM Systems: This paper proposes an adaptively coordinated multi-agent LLM framework that achieves a 27% improvement in compliance accuracy and a 74% reduction in revision rate on high-complexity financial document analysis tasks, through parallel competitive evaluation, dynamic task routing, and bidirectional feedback mechanisms.
SafeSieve: From Heuristics to Experience in Progressive Pruning for LLM-based Multi-Agent Communication: SafeSieve is proposed as a progressive adaptive multi-agent communication pruning framework. Through a two-stage edge scoring mechanism combining semantic-heuristic initialization and history-feedback-driven refinement, together with 0-extension clustering, SafeSieve achieves 94.01% average accuracy across 6 benchmarks while reducing token consumption by 12.4%–27.8%, and demonstrates inherent robustness against prompt injection attacks.
Scalable and Accurate Graph Reasoning with LLM-Based Multi-Agents: This paper proposes GraphAgent-Reasoner (GAR), inspired by distributed graph computation theory. It decomposes graph problems into node-centric subtasks assigned to multiple agents, which collaborate through neighbor message passing. GAR extends the graph scale tractable by LLMs from 100 nodes to 1,000 nodes, and significantly outperforms existing state-of-the-art methods on polynomial-time graph reasoning tasks.
Shadows in the Code: Exploring the Risks and Defenses of LLM-based Multi-Agent Software Development Systems: The first systematic security analysis of LLM-based multi-agent software development systems (ChatDev/MetaGPT/AgentVerse): proposes the IMBIA attack framework covering two threat scenarios (malicious user + benign agents / benign user + malicious agent) and 12 malicious behaviors across 5 malware families, achieving an attack success rate (ASR) of up to 93% on ChatDev, with the Adv-IMBIA adversarial defense reducing ASR by 40–73%.
Thucy: An LLM-based Multi-Agent System for Claim Verification across Relational Databases: This paper presents Thucy, the first multi-agent claim verification system supporting cross-database and cross-table reasoning. Led by a Verifier agent, it coordinates three specialized agents (Data/Schema/SQL Expert) with zero prior knowledge of the data sources, enabling autonomous discovery, reasoning, and SQL evidence generation. Thucy surpasses the previous SOTA by 5.6 percentage points on TabFact (94.3%).