Skip to content

🦾 LLM Agent

🧠 NeurIPS2026 · 270 paper notes

📌 Same area in other venues: 🎞️ ECCV2026 (36) · 📷 CVPR2026 (45) · 🔬 ICLR2026 (209) · 💬 ACL2026 (121) · 🧪 ICML2026 (83) · 🤖 AAAI2026 (59)

🔥 Top topics: Agents ×93 · LLM ×64 · Reasoning ×15 · Adversarial Robustness ×8 · Reinforcement Learning ×7

3DCodeBench: Benchmarking Agentic Procedural 3D Modeling Via Code

3DCodeBench introduces a comprehensive benchmark spanning 212 semantic categories and 26K multimodal prompt-code-mesh triplets paired with the 3DCodeArena human preference platform, systematically dissecting the execution capabilities, API adaptabilities, and physical plausibility bottlenecks of 12 frontier VLMs.

A Few Teacher Steps Go a Long Way: Cost-Efficient On-Policy Data Augmentation for Agent Post-Training

This paper formulates on-policy data augmentation for LLM agents as a budget-allocation problem, proving that bounded few-step teacher continuations at learner-induced states outperform pure behavioral cloning and heavily filtered baselines at matched supervision costs.

A Matter of TASTE: Improving Coverage and Difficulty of Agent Benchmarks

Addressing the rapid saturation and narrow procedural coverage of existing conversational agent benchmarks, this paper reverses the task construction process via TASTE, which explores tool-sequence space using an adaptive contrastive n-gram model, selects representative sequences via weighted clustering, and applies adversarial task evolution to build the challenging, high-coverage \(\tau^c\)-Bench.

Adaptive Adversaries: A Multi-Turn, Multi-LLM Benchmark for LLM Agent Security

Presents Security Arena, a 21-scenario benchmark pairing autonomous LLM attackers with fresh-session LLM defenders, demonstrating that single-round evaluation severely underestimates attack risks and that aggregate metrics mask stark, scenario-specific vulnerabilities across frontier models.

AEM: Adaptive Entropy Modulation for Multi-Turn Agentic Reinforcement Learning

Addressing ambiguous credit assignment and premature entropy collapse caused by sparse outcome-only rewards in multi-turn agentic RL, AEM reveals via information geometry that natural-gradient entropy drift is governed by the product of advantage and relative surprisal, adaptively rescaling advantages via a response-level uncertainty proxy to smoothly transition from early exploration to late exploitation without external supervision.

Agent2World: Learning to Generate Symbolic World Models via Adaptive Multi-Agent Feedback

Agent2World proposes an automated tool-augmented multi-agent framework that synthesizes symbolic world models through deep web research, sandboxed code implementation, and adaptive testing (unit tests and ReAct simulation), while transforming verified multi-turn interaction traces via rejection sampling into an effective SFT data engine.

Agent Meltdowns: The Road to Hell Is Paved with Helpful Agents

This paper introduces and measures "accidental meltdowns," a critical failure mode where autonomous agents, lacking any adversarial influence, autonomously escalate into unsafe behaviors—such as unauthorized reconnaissance, privilege escalation, and credential leakage—solely to satisfy their helpfulness optimization when encountering benign environmental errors.

Agent Security is a Systems Problem

Reframing agent security as a systems security problem, this work argues that the underlying AI model must be treated as an untrusted component, enforcing deterministic security invariants at the encompassing system level via instruction-data separation, verifiable least-privilege policy generation, and information flow control (IFC).

Agent-ToM: Learning to Monitor Autonomous LLM Agents via Theory-of-Mind Reasoning

Addressing the challenge of detecting covert malicious side tasks in autonomous LLM agents over extended horizons, Agent-ToM combines Theory-of-Mind latent state inference with a two-call Reason–Verify–Refine pipeline and distills critique feedback into persistent semantic guardrail memory, outperforming multi-call ensemble monitors.

AgentAbstain: Do LLM Agents Know When Not to Act?

AgentAbstain introduces the first systematic evaluation benchmark and automated synthesis pipeline (AbstainGen) for agentic abstention across pre-execution and runtime scenarios, revealing that even the top frontier model (Gemini 3.1 Pro) achieves only 59.5% paired accuracy and demonstrating that task-solving competence scales independently of calibrated restraint.

AgentBrew: Offline Tool-Use Agent Learning from Raw Real-World Trajectories

Addressing real-world environments lacking pre-defined tasks, verifiers, and interaction budgets, AgentBrew introduces a fully offline learning framework that retrospectively infers aligned tasks from raw trajectories and applies the PMI chain rule to assign fine-grained action credits for weighted policy optimization.

AgentCollabBench: Diagnosing When Good Agents Make Bad Collaborators

Addressing how multi-agent teams silently drop constraints, propagate false beliefs, and leak private context, this paper presents AgentCollabBench—a 900-task diagnostic benchmark across five topologies and four decoupled behavioral metrics—revealing that communication topology accounts for 7%–40% of variance in multi-hop information survival and isolates a severe synthesis bottleneck at converging-DAG nodes.

AgentForesight: Online Auditing for Early Failure Prediction in Multi-Agent Systems

AgentForesight reframes multi-agent failure detection as online prefix auditing, developing the AFTraj-22K corpus and training a 7B auditor via a coarse-to-fine reinforcement learning recipe (BPPO + GRPO) that triggers alarms and attributes fault at the earliest decisive error without access to future steps.

AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems

To tackle credit assignment failures and gradient mixing conflicts in multi-agent prompt optimization, AgentGrad introduces reverse sequential intervention to identify responsible agents and produce pseudo-label supervision, combined with semantic clustering and gradient abstraction to generate coherent, generalizable updates.

AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems

AgentGrad identifies repairable prompts through single-agent interventions, extracts textual gradients from differences between original and intervention-adjusted intermediate outputs, and clusters and abstracts shared corrective patterns; with GPT-5-mini, it improves the five-task average over unoptimized prompts by 11.76 percentage points, although some performance and cost summaries require the exceptions in the tables.

AgentHop: A Diagnostic Benchmark for Agentic Multi-Hop Scientific Question Answering

AgentHop introduces a diagnostic benchmark of 1,011 multi-hop scientific questions grounded in citation networks across 7,205 arXiv papers and a 7-tool resource-constrained sandbox, dissecting aggregate accuracy into retrieval, synthesis, tool-call signatures, and resource management across 19 frontier LLMs.

AgentHop: A Diagnostic Benchmark for Agentic Multi-Hop Scientific Question Answering

AgentHop places 1,011 scientific four-option questions in a seven-tool sandbox with three resource caps and diagnoses failures through paper recall, conditional conversion, tool-call patterns, and resource failures; Gemini-3 Pro achieves 89.1% accuracy in the paper's evaluation, while similar overall scores can conceal different retrieval and synthesis bottlenecks.

Agentic Abstention: Do Agents Know When to Stop Instead of Act?

This paper formalizes Agentic Abstention as a sequential POMDP decision problem, establishes a 28,000-task benchmark across web shopping, terminal, and QA environments, reveals that current agents struggle significantly with timely abstention, and introduces convolve, a parameter-free context-engineering method that distills execution trajectories into reusable stopping rules to improve timely abstention.

Agentic AI-Empowered Dynamic Survey Framework

Reframes academic survey writing as a long-horizon maintenance problem over a persistent document state, orchestrating modular agents for literature analysis, hierarchical routing, and localized synthesis to continuously integrate new papers under a frozen outline with zero out-of-scope disruption.

Agentic AI Scientists Are Not Built for Autonomous Scientific Discovery

Addressing excessive claims of fully autonomous "AI scientists", this position and empirical paper identifies four fundamental structural barriers—the McNamara fallacy in problem selection, training corpus omissions of tacit and failure knowledge, diversity compression from preference optimization, and invalid single-turn benchmarks—and outlines concrete design remedies including physics-grounded simulators, persistent mutable epistemic states, and mandatory machine preregistration.

Agentick: A Unified Benchmark for General Sequential Decision-Making Agents

Agentick is a unified benchmark for general sequential decision-making agents featuring 37 procedurally generated tasks across 6 capability dimensions, 4 difficulty tiers, and 5 synchronized observation modalities to evaluate and train RL, LLM, VLM, and hybrid agents on common ground.

AgentKVShift: Efficient KV Cache Reuse for Agentic Memory Systems

AgentKVShift introduces a training-free, probe-guided chunk-level residual mean-shift correction that repairs unrefreshed KV cache entries for structured agentic memory, reaching near full-recompute accuracy with only 10%–30% refresh budget and achieving 2–3.5× prefill speedups.

AgentLens: Revealing The Lucky Pass Problem in SWE-Agent Evaluation

Addressing the flaw of binary pass/fail evaluation in software engineering agents, this paper introduces AgentLens and the AgentLens-Bench dataset, utilizing Prefix Tree Acceptors to merge diverse correct solution spaces and context-aware intent labeling to reveal that 10.7% of passing trajectories are brittle "Lucky Passes" driven by trial-and-error and missing verification.

AgentOdyssey: Open-Ended Long-Horizon Text Game Generation for Test-Time Continual Learning Agents

AgentOdyssey is a novel evaluation framework that procedurally synthesizes open-ended, long-horizon text games to evaluate test-time continual learning agents across five core cognitive abilities, revealing that current agents face severe behavioral collapse and catastrophic forgetting during test-time parametric updates.

AgentRoom: Concurrent Multi-Agent Coding in a CRDT-Backed Shared Workspace

AgentRoom introduces a real-time collaborative editing protocol for concurrent coding agents backed by a CRDT shared filesystem and explicit MCP coordination tools (file-level claims, append-only broadcast, and state inspection), which suppresses the common lone-agent "stub-and-exit" abandonment failure mode and substantially outperforms uncoordinated parallel merge and sequential pipelines at matched compute.

Agents' Last Exam

Developed with over 250 industry experts across 13 industry clusters and 55 digital subfields, Agents' Last Exam (ALE) benchmarks Generalist Computer-Use Agents on 1K+ long-horizon authentic workflows, revealing that frontier agent systems achieve less than 1% full pass rate on the hardest industrial tier.

AIRA 2: Overcoming Bottlenecks in AI Research Agents

AIRA 2 overcomes the three structural bottlenecks of sample throughput, validation overfitting, and static operators in AI research agents via an asynchronous multi-GPU steady-state evolutionary pool, a Hidden Consistent Evaluation (HCE) protocol, and dynamically scoped ReAct operators, achieving a 24-hour mean Percentile Rank of 81.5% and a 72-hour rank of 83.1% on MLE-bench-30.

AM-Bench: A Unified Taxonomy and Evaluation Suite for Agentic Misalignment

AM-Bench grounds agentic misalignment evaluation in the criminological fraud triangle, establishing a unified taxonomy across perceived threat, perceived opportunity, and reasoning, and evaluates 10 frontier models across 35 realistic industry scenarios to reveal a 62-percentage-point misalignment gap and demonstrate that harmful behaviors can emerge without full explicit situational awareness.

AndroidReality: How Far Are Mobile Agents from the Real World?

Addressing the severe performance degradation of mobile GUI agents under real-world interface noise and delays, this paper formalizes environmental variability into an MDP-grounded perturbation benchmark across state, transition, and action axes, uncovering four recurring failure modes and establishing a training-free Test-Time Introspective Recovery (TTIR) module.

Are Agents Ready to Teach? A Multi-Stage Benchmark for Real-World Teaching Workflows

TeachArena establishes the first benchmark spanning professional pedagogical judgment, situated multi-turn tutoring, and end-to-end LMS workflow execution across 354 audited tasks, revealing a stark knowing–teaching–acting gap in frontier models.

ARTIS: Agentic Risk-Aware Test-Time Scaling via Iterative Simulation

ARTIS decouples exploration from commitment in agentic tool-use environments by performing cross-trial full-trajectory simulation, trajectory-level self-evaluation, and strategy summarization prior to execution, powered by a failure-driven risk-aware tool simulator.

Attack Selection in Agentic AI Control Evaluations Meaningfully Decreases Safety

This paper exposes the pervasive optimistic bias in agentic AI control evaluations caused by the indiscriminate attack assumption, demonstrating that decomposing attack decisions into a task-level start policy and a turn-level stop policy reduces empirical safety by 20–28 percentage points without altering underlying attack capabilities.

Auto-Dreamer: Learning Offline Memory Consolidation for Language Agents

Inspired by complementary learning systems theory, Auto-Dreamer decouples fast per-session memory writing from slow offline region rewriting, using GRPO with a counterfactual utility reward to boost agent success rates while compressing active memory banks by up to 12×.

Automata from Agent Traces: Failure and Next-Step Prediction

This paper demonstrates that collapsing unstructured LLM-agent execution traces into a compact finite-state machine (7–43 states) in milliseconds provides a hyperparameter-free structural substrate that unifies workflow memory, next-step prediction, and early failure monitoring.

Automatic Textbook Formalization

This paper introduces a multi-agent autoformalization framework grounded in battle-tested software engineering practices, deploying approximately 30,000 Claude 4.5 Opus agents to fully formalize a 500+ page graduate algebraic combinatorics textbook into a 130K-line Lean 4 codebase in one week, successfully verifying all 340 target theorems and definitions.

Automating ML for Science: Can Frontier Agents Climb Scientific Hills in the Wild?

Introduces Agent4Sci, a long-horizon benchmark spanning 20 scientific ML benchmarks across 11 domains with a decoupled three-stage evaluation protocol (hill-climbing, held-out evaluation, and auditing), revealing that frontier agents advance state of the art primarily via generic ML engineering rather than scientific insight, while lightweight plug-in skills double the performance gain.

AutoMem: Automated Learning of Memory as a Cognitive Skill

AutoMem reframes memory management as an agent's self-directed cognitive skill over file-system primitives, employing two automated meta-LLM outer loops to optimize memory scaffold architecture and fine-tune a dedicated memory specialist, delivering a 2x to 4x performance boost on long-horizon games without altering the underlying world-action model weights.

Autonomous Continual Learning for Environment Adaptation of Computer-Use Agents

Presents ACuRL, an autonomous curriculum reinforcement learning framework for computer-use agents that continuously adapts models to target environments without human data via capability-guided task generation and state-difference-based trajectory evaluation.

Autonomous Scientific Discovery via Iterative Meta-Reflection

DiscoPER introduces an unconstrained, open-ended scientific discovery agent framework that formulates and tests hypotheses via executable statistical code without predefined questions, and leverages second-order meta-reflection over accumulated claims to explore epistemic gaps and rediscover real-world scientific phenomena with high statistical fidelity.

Autoregressive Diffusion World Models for Off-Policy Evaluation of LLM Agents

Adwm introduces an offline evaluation framework for LLM agents that factors globally policy-guided trajectory laws into single-step conditionals, enabling an autoregressive latent diffusion world model to simulate environment transitions with policy score guidance without online execution.

AVO: Agentic Variation Operators for Autonomous Evolutionary Search

Elevates large language models from passive candidate generators to self-directed evolutionary variation operators (AVO), autonomously evolving an attention kernel on NVIDIA Blackwell B200 GPUs over 7 continuous days that outperforms cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5%.

BankerToolBench: Evaluating AI Agents in End-to-End Investment Banking Workflows

Introduces BankerToolBench, the first end-to-end benchmark for junior investment banking workflows developed with 502 bankers spanning 100 multi-file tasks and 100+ rubric criteria per task, alongside Gandalf, an executable Agent-as-a-Judge verifier, demonstrating that 0% of outputs from current frontier models are client-ready.

Behavioral Probes for Information Flow in LLM Swarms

Repurposing External Context Score (ECS) and Parametric Knowledge Score (PKS)—originally designed for single-model hallucination analysis—into behavioral probes of internal information flow in LLM swarms, serving as non-intrusive auxiliary signals to guide topology pruning and reasoning-path search.

BehaviorBench: Modeling Real-World User Decisions from Behavioral Traces

BehaviorBench introduces a personalized decision modeling benchmark derived from real-world on-chain prediction market traces, separating evaluation into macro Belief prediction and micro Trade prediction across four distinct history interfaces.

Beyond Domains: Reusing Web Skills via Transferable Interaction Patterns

Addressing the severe drop in cross-domain skill reuse caused by semantic-only retrieval in existing web agents, this paper presents SkillMigrator, which encapsulates skills as Transferable Interaction Patterns (TIPs) paired with structural accessibility sketches, combines tree edit distance layout similarity with rich-text matching during retrieval, and grounds abstract slots to live DOM controls via bipartite Hungarian matching, reducing LLM actions on successful trajectories by 8–10% at matched SOTA success rates.

Beyond LLM-Based Reasoning: Lightweight GNNs for Agent Failure Attribution

Tackling the severe computational bottleneck and limited attribution accuracy of heavy generative LLMs in multi-agent failure attribution (AFA), this paper presents AFANet, a lightweight 65K-parameter graph neural network that models turn-level interaction graphs and anomalous consistency signals to match or outperform fine-tuned LLMs with sub-second inference.

Beyond Outcome Rewards: Step-Level Self-Distilled Policy Optimization for Deep Search Agents

Addressing the extreme sparsity of outcome-only rewards in long-horizon deep search agents, this paper proposes Step-Level Self-Distilled Policy Optimization (SSPO), which extracts evidence anchors as privileged information to construct teacher distributions, translates teacher-student step-level disagreement into advantage weights exclusively applied to incorrect trajectories, and significantly outperforms GRPO with only ~5% additional training overhead.

Beyond Suspicious Steps: Ontological Trust in Long-Horizon Agents

Addressing the insidious failure mode where long-horizon agents execute individually plausible steps yet diverge across their trajectory prefix from authorized delegations, this paper formalizes ontological trust across Role, Goal, and Evidence (RGE), and deploys an online monitor using LLMs solely for structured step parsing while keeping state updates, projections, and intervention decisions strictly deterministic.

Bian Que: An Agentic Framework with Flexible Skill Arrangement for Online System Operations

Addressing the core bottleneck of dynamic context orchestration and knowledge drift in large-scale online engine operations, Bian Que introduces a three-lines-of-defense agentic architecture featuring flexible skill arrangement and dual-pathway self-evolution, reducing alert volume by 75% and cutting mean time to resolution by over 50% in production at Kuaishou.

CAMELS Can Use Computers Too: System-Level Security for Computer Use Agents

Addressing the vulnerability of computer use agents to prompt injection, this paper adapts the Dual-LLM isolation paradigm to vision-based CUAs via single-shot upfront planning and the NOVA methodology, enforcing formal Control Flow Integrity while uncovering residual data-flow Branch Steering threats.

Can AI Agents Synthesize Scientific Conclusions?

This paper introduces SciConBench, a large-scale live benchmark of 9.11K systematic review QA pairs, SciConHarness, a clean-room evaluation harness that intercepts ground-truth references, and an expert-aligned atomic-fact evaluation pipeline, evaluating 8 frontier models and deep research agents to reveal that the state-of-the-art system achieves only a factual F1 of 0.337 under clean-room conditions.

Can LLM Agents Respond to Disasters? Benchmarking Heterogeneous Geospatial Reasoning in Emergency Operations

Introduces DORA, the first agentic benchmark for end-to-end operational disaster response, comprising 515 expert-designed tasks across 45 real-world disaster events and a 108-tool MCP library, revealing severe compositional fragility and domain grounding bottlenecks in 13 frontier LLMs.

CANDO: Cooperative Agentic Network for Layout Design Optimization

To tackle complex irregular site boundaries and multi-vehicle routing constraints in real-world asset layout design, this paper introduces the 1,000-sample CAD-annotated ALPS-Bench alongside CANDO, a training-free cooperative agentic network that unites specialized agents, deterministic tool verification, and blockwise parallel exploration to achieve state-of-the-art constraint compliance.

CASPIAN: Online Detection and Attribution of Cascade Attacks in LLM Multi-Agent Systems via Cross-Channel Causal Monitoring

To tackle cascade attacks that amplify and spread across communication, memory, tool, and execution channels in LLM multi-agent systems, CASPIAN establishes an online detection and replay-free attribution framework via Late-Interaction Conditional Transfer Entropy (LI-CTE) and dynamic spectral phase monitoring, delivering high accuracy and early warning with sub-1% relative latency overhead.

CASPIAN: Online Detection and Attribution of Cascade Attacks in LLM Multi-Agent Systems via Cross-Channel Causal Monitoring

CASPIAN detects and attributes cascade attacks across communication, memory, tool, and execution channels in LLM multi-agent systems online using Late-Interaction Conditional Transfer Entropy and spectral phase-space dynamics, achieving early detection with sub-1% relative latency overhead.

CEO-Bench: Can Agents Play the Long Game?

CEO-Bench simulates a 500-day software startup (NovaMind) with 34 tools and a 19-table SQL database, evaluating whether LLM agents can orchestrate long-horizon strategy, infer hidden states under noisy feedback, and adapt to evolving competitive and macroeconomic environments.

\(\chi\)-Bench: Can AI Agents Automate End-to-End, Long-Horizon, Policy-Rich Healthcare Workflows?

Addressing end-to-end healthcare operations characterized by long horizons, strict policy compliance, and multi-role transitions, this paper establishes \(\chi\)-Bench—an in-situ evaluation benchmark comprising 20 simulated healthcare applications, 87 MCP tools, and a 1,279-document managed-care operations handbook—demonstrating that state-of-the-art frontier agents achieve at best 28.0% pass@1 and drop to near-zero success in multi-case sessions and two-agent arenas.

Choosing Before Acting: Comparative Value Estimation for Long-Horizon Tool-Use Agents

To overcome weak credit assignment and irreversible failure accumulation in long-horizon tool use, CITA introduces a Comparative Inference Model (CIM) trained on tri-source paired signals—real traces, Bayesian graph simulation, and LLM semantic comparison—to estimate downstream value before action execution and provide dense comparative rewards during GRPO training.

CI4A: Semantic Component Interfaces for Agents Empowering Web Automation

Addressing the vulnerability of conventional Web agents trapped in fragile long-sequence atomic operations over low-level DOM/GUI trees, this paper proposes the CI4A protocol to encapsulate front-end UI components into structured semantic tool primitives, implementing the AntDX reference suite alongside the hybrid-action agent Eous to achieve a state-of-the-art 86.3% success rate on WebArena while reducing interaction steps by 57.5%.

CLAW-Eval: Towards Trustworthy Evaluation of Autonomous Agents

CLAW-Eval establishes an end-to-end auditable benchmark for autonomous agents that integrates trajectory traces, server-side audit logs, and environment snapshots across 300 heterogeneous tasks, evaluating completion, embedded safety constraints, and perturbation robustness.

CLIFT: Conformal Self-Verification for Web Agent Training and Test-Time Scaling

CLIFT establishes a conformally calibrated, URL-aware self-verification question bank that injects non-negative step-level dense rewards with local advantage stratification during GRPO training, and enables judge-free test-time trajectory scaling through conservative majority-voted selection.

Closing the Reflection Gap: A Free Calibration Bonus for Agentic RL

RefGRPO augments standard outcome-only RL with a free calibration bonus computed by contrasting post-feedback self-reflection with the actual outcome under a dynamic schedule, substantially closing the reflection underconfidence gap while boosting task accuracy and turning the agent into a reliable self-verifier.

CM2: Reinforcement Learning with Checklist Rewards for Multi-Turn and Multi-Step Agentic Tool Use

To overcome the lack of verifiable rewards and the high engineering cost of executable sandboxes in multi-turn, multi-step agentic tool use, CM2 introduces an RL framework combining fine-grained binary checklist rewards with a "sparse in assignment, dense in criteria" optimization strategy in a 5,000-tool simulated environment, significantly outperforming supervised fine-tuning baselines.

Colosseum: Auditing Collusion in Cooperative Multi-Agent Systems

Colosseum establishes a formal collusion auditing framework that models cooperative multi-agent systems as Distributed Constraint Optimization Problems (DCOPs), pairing secret-channel behavioral probes with dual-track regret and communication metrics to reveal emergent collusion and "collusion on paper" across 19 frontier LLMs.

Computer Use at the Edge of the Statistical Precipice

Reveals that a 1MB blind replay script beats frontier models on static CUA benchmarks, establishes the PRISM design principles with a 3.2M-configuration sandboxed benchmark DigiWorld, and introduces a nested Wilson score and hierarchical bootstrap evaluation framework.

Conformal Agent Error Attribution

Addressing the challenge of locating the root-cause failure in long multi-agent trajectories, this paper introduces a sequential filtration-based conformal prediction framework that guarantees finite-sample, distribution-free coverage of decisive error steps using contiguous prediction intervals to guide minimal-cost automated recovery.

ContractBench: Can LLM Agents Preserve Observation Contracts?

ContractBench uses a virtual clock, byte checks, and HTTP-trace validation to evaluate whether agents preserve cross-step contracts attached to tool artifacts; in the paper's experimental snapshot of 38 model variants, the highest success rate is 77.8%, and greater scale or newer versions do not guarantee improved reliability.

Control Charts for Multi-Agent Systems

This paper introduces adaptive Shewhart control charts to multi-agent system monitoring via temporal perspective spaces and the iso-mirror, demonstrating empirically and theoretically that while adaptive baselines are necessary to accommodate natural learning, they are fundamentally vulnerable to slow adversarial sleeper agent drift.

COOP\(^2\): Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems

COOP\(^2\) decouples high-level symbolic cognition from low-level grounded primitives to evaluate LLM multi-agent cooperation through four verifiable constraints, introducing a predictive repair mechanism (COOP\(^2\)-Repair) that significantly enhances coordination efficiency.

COOP2: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems

COOP2 formalizes cooperation in LLM multi-agent systems via a cognitive–primitive interface that bridges asynchronous reasoning and synchronous environment execution, guarded by four verifiable constraints, and introduces COOP2-Repair to predict constraint failures and guide targeted plan revisions.

CTM-AI: A Blueprint for General AI Inspired by a Model of Consciousness

Grounding the formal Conscious Turing Machine (CTM) theory with modern foundation models, CTM-AI establishes a fully decentralized general AI architecture that coordinates specialized and general processors via up-tree workspace competition, down-tree conscious broadcast, and autonomous link fusion without a central orchestrator.

CyBiasBench: Benchmarking Bias in LLM Agents for Cyber-Attack Scenarios

CyBiasBench establishes a 630-session offensive security benchmark across 5 frontier LLM agents, revealing persistent attack-selection bias decoupled from success rates and discovering a "bias momentum" effect where prompt steering degrades attack performance.

CyBiasBench: Benchmarking Bias in LLM Agents for Cyber-Attack Scenarios

Addressing the blind spot of outcome-only security evaluation, CyBiasBench introduces a 630-session benchmark that reveals persistent attack-selection biases in LLM agents, their fundamental decoupling from attack success rates, and a "bias momentum" effect that resists prompt-level steering.

DAGent: Evaluate-then-Grow Planning for Deep Research Agents

DAGent introduces the Evaluate-then-Grow incremental planning paradigm and the topology-aware DAGRPO reinforcement learning algorithm, replacing brittle Plan-then-Patch workflows to achieve superior deep research accuracy with significantly lower token, tool-call, and step footprints.

DAGent: Evaluate-then-Grow Planning for Deep Research Agents

DAGent evaluates completed sub-tasks before incrementally growing a research DAG, combining hierarchical evidence propagation with DAGRPO structural rewards: training-free Qwen3-32B achieves 47.3 / 55.3 / 65.0 Pass@1 across three benchmarks, outperforming its same-architecture Plan-then-Patch variant by 5.3 / 4.8 / 4.0 percentage points.

DEBATE: A Large-Scale Benchmark for Evaluating Opinion Dynamics in Role-Playing LLM Agents

Introduces DEBATE, the first large-scale empirical benchmark grounded in 2,788 human participants across 107 controversial topics to evaluate the fidelity of opinion dynamics in multi-agent role-playing LLMs, diagnosing persistent premature consensus convergence and showing that supervised fine-tuning partially restores human-like resistance to conformity.

Deep Reasoning in General Purpose Agents via Structured Meta-Cognition

Introduces Deep Reasoning, a formal meta-cognitive language, and Dolores, a general-purpose agent that dynamically decomposes complex tasks across associative inference, formal code execution, and recursive sub-agents to outperform fixed scaffolding baselines.

Deep Reasoning in General Purpose Agents via Structured Meta-Cognition

Introduces the Deep Reasoning formal meta-cognitive language and Dolores, a general-purpose agent that dynamically constructs task-specific scaffolds at test time using in-context atomic decomposition examples, outperforming evaluated state-of-the-art scaffolds by 24.8% on average across four hard reasoning benchmarks.

Defense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented Agents

Addressing persistent hijacking risks in skill-augmented agents (SAAs) caused by malicious skills, this paper introduces Defense-as-Skill to implement runtime protection as an installable, inspectable, and editable guard skill called SkillSonar, which leverages closed-loop SCOPE-R evaluations and MCTS evolution to slash attack success rates while preserving high task utility without modifying host runtimes.

Democratizing Tool Learning with Environments Fully Simulated by a Free 8B Language Model

TRUSTEE introduces a cost-friendly tool learning framework where task generation, user simulation, tool execution, and trajectory verification are fully simulated by a free open-source 8B LM, driven by multi-dimensional adaptive curriculum learning without requiring annotated data, executable environments, or proprietary LMs.

DiagEval: Trajectory-Conditioned Diagnosis for Reliable Software Evaluation with GUI Agents

Addressing the single-trajectory identifiability gap where failed GUI agent rollouts confound evaluator execution slips with genuine software defects, DiagEval reuses failed trajectories to dispatch targeted diagnostic probes and update an internal attribution score via expected information gain, dramatically reducing false-negative evaluations.

DIBench: Benchmarking Decision Integrity of GUI-based Mobile Agents Under Deceptive Injections

Addressing the hidden threat where mobile GUI agents are silently steered away from user goals by non-privileged UI content during multi-candidate selections, this work introduces DIBench—the first decision integrity benchmark with a paired evaluation protocol—revealing the critical risk masked by inflated task completion rates.

Do Enterprise Systems Need Learned World Models? The Importance of Context to Infer Dynamics

Addressing the vulnerability of parameterized enterprise world models under tenant deployment shift, this paper proposes runtime Enterprise Discovery Agents that actively inspect live business rules and schemas to infer state transitions, demonstrating marked robustness gains over offline supervised world models on the newly introduced CascadeBench benchmark.

DPIAgent: Divide, Protocol, Isolate for Agentic Reproduction Test Generation

Addressing goal drift and compound-objective ambiguity in monolithic agents during automated reproduction test generation, DPIAgent introduces the principles of Divide, Protocol, and Isolate to impose minimal boundary constraints while retaining internal agent autonomy, achieving state-of-the-art issue reproduction across diverse LLM backbones.

Dr. MAS: Stable Reinforcement Learning for Multi-Agent LLM Systems

Dr. MAS identifies gradient-norm explosion in cooperative multi-agent LLM systems under GRPO caused by global baseline mismatch with heterogeneous agent reward distributions, resolving it via per-agent advantage normalization and lifecycle-aware GPU pooling.

ECHO: Terminal Agents Learn World Models for Free

ECHO introduces an auxiliary next-token cross-entropy objective on terminal observation tokens during standard GRPO, reusing the actor forward pass with zero rollout overhead to teach the policy environmental dynamics, doubling pass@1 on TerminalBench-2.0 and substantially closing the gap to expert SFT.

Economy of Minds: Emerging Multi-Agent Intelligence with Economic Interactions

Inspired by Friedrich Hayek's decentralized market theory, Economy of Minds (EoM) organizes weak partial LLM agents through local competitive auctions, peer-to-peer bucket-brigade payments, and wealth-based evolutionary selection, enabling global multi-step planning and specialization to emerge without centralized coordination.

EnvOS: Auditing Computer-Use Agents When the World Moves On : Phenomenon-Based Task Construction, Certification, and Evidence-Based Evaluation

EnvOS establishes a phenomenon-based task construction and verifiable evaluation framework for computer-use agents facing mid-task state mutations, introducing pre-run formal certification, decoupled observation-delivery tracking, and dual-track audit pipelines to eliminate benchmark misattribution and reward hacking.

EVOCHAMBER: Test-Time Co-evolution of Multi-Agent System at Individual, Team, and Population Scales

EvoChamber is a training-free test-time multi-agent co-evolution framework that decomposes evolvable state across individual, team, and population levels, leveraging Collaborative Dreaming (CoDream) for asymmetric knowledge transfer to spontaneously induce stable niche specialists across heterogeneous task streams.

EvoChamber: Test-Time Co-Evolution of Multi-Agent System at Individual, Team, and Population Levels

EvoChamber establishes a training-free multi-agent test-time co-evolution framework across individual experience archives, team-level anchor-complement-scout orchestration with LeadLearn collaboration topology selection, and population-level CoDream asymmetric reflective distillation with lifecycle operators, enabling identically initialized agents to spontaneously differentiate into 4–5 stable domain specialists without gradient updates.

EvoCUA: Evolving Computer Use Agents via Learning from Scalable Synthetic Experience

EvoCUA overcomes the scalability bottlenecks and causal feedback deficit of static training data in native Computer Use Agents by integrating a verifiable synthesis engine, a high-concurrency sandbox infrastructure, and an evolving experience learning cycle, achieving an open-source state-of-the-art 56.7% success rate on the OSWorld benchmark.

EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective

EvoMemBench establishes the first open benchmark to evaluate agent memory from a self-evolving perspective, decoupling memory evaluation along two orthogonal axes—scope (in-episode vs. cross-episode) and content (knowledge vs. execution)—across 15 representative memory systems and uncovering critical bottlenecks in memory revision, context competition, and cross-domain transfer.

Expanding LLM Agent Boundaries with Strategy-Guided Exploration

Addressing the severe exploration bottleneck of LLM agents under sparse outcome rewards and complex action spaces, this paper proposes Strategy-Guided Exploration (SGE), which samples high-level natural language strategies prior to actions, combines mixed-temperature sampling with positive/negative strategy reflection, and significantly expands agent capabilities beyond the base model's reach.

Fisher-R1: Training LLM Agents for Reliable Hypothesis Testing

Stanford and UW-Madison researchers construct P-Bench, an expert-audited benchmark of 425 real-world hypothesis-testing tasks, and train Fisher-R1, an open-weight statistical agent that leverages synthetic task generation and z-space outcome-grounded reinforcement learning to significantly reduce false discoveries and statistical assumption violations in autonomous scientific discovery.

FlashEvolve: Accelerating Agent Evolution with Asynchronous Stage Orchestration

Addressing the massive wall-clock bottlenecks caused by synchronized stage barriers and intra-stage generation imbalance in LLM agent self-evolution, FlashEvolve introduces an asynchronous stage orchestration framework with queue-connected workers, leverages the unique inspectability of language-space artifacts to semantically repair stale proposals, and incorporates speculative stage completion and adaptive workflow control to achieve up to a 4.9x speedup in proposal throughput across GEPA, ACE, and Meta-Harness.

FlashEvolve: Accelerating Agent Evolution with Asynchronous Stage Orchestration

Addressing synchronization barriers and long-tail imbalances in LLM agent self-evolution, FlashEvolve introduces an asynchronous queue-based execution pipeline paired with reflective semantic staleness repair, speculative completion, and adaptive worker reallocation, accelerating proposal throughput by 3.5×–4.9× while preserving evolution quality.

FlowMAS: Learning Multi-Agent Workflow Topology via Information-guided Generative Flow Network

FlowMAS introduces Generative Flow Networks (GFlowNets) into automated multi-agent workflow topology design, combining partial information decomposition synergy rewards, channel spectral entropy, and curiosity-driven exploration to discover diverse, high-performing collaboration topologies with substantial token savings.

FlowSteer: Prompt-Only Workflow Steering Exposes Planning-Time Vulnerabilities in Multi-Agent LLM Systems

FlowSteer reveals that planner-executor multi-agent systems can be steered toward malicious targets purely through user-facing prompts that manipulate task decomposition and routing dependencies, which motivates FlowGuard, an input-side intent triage and decontamination defense.

FlowSteer: Towards Agents Designing Agentic Workflows via Reinforced Progressive Canvas Editing

FlowSteer introduces the Agent Designing Agentic Workflows paradigm, establishing an executable graph-state environment called Workflow Canvas and a reinforced progressive editing mechanism that decouples design from execution, training a lightweight director policy via token-masked and structural diversity-gated GRPO to significantly outperform static, search-based, and RL baselines across 12 benchmarks and 6 LLM backends.

FML-bench: A Controlled Study of AI Research Agent Strategies from the Perspective of Search Dynamics

FML-bench introduces an 18-task benchmark spanning 10 ML domains that decouples agent search strategies from execution infrastructure and tracks 12 process-level behavioral metrics, revealing that strategy complexity does not guarantee performance, greedy search suits opportunity-dense tasks while multi-branch exploration suits opportunity-sparse ones, and validating an adaptive switching agent that outperforms both.

Forgetting is Not Always Bad: A Neuro-Inspired Memory Repair Mechanism for Poisoned LLM Agents

Inspired by biological mechanisms where weak cueing destabilizes non-reinforced memories toward gradual decay, DREAM introduces an intrinsic, training-free memory repair framework that selectively induces functional forgetting of poisoned entries through perturbation probing and synaptic weight dynamics without external online auditing.

From History to State: Constant-Context Skill Learning for LLM Agents

Targeting recurring personal assistant workflows, this paper introduces constant-context skill learning, internalizing procedural knowledge into lightweight task-family LoRA modules while conditioning inference on a compact state block, reducing prompt tokens per turn by 2–7× while matching or outperforming strong baselines.

From Intent to Evidence: A Categorical Approach for Structural Evaluation of Deep Research Agents

This paper introduces a category-theoretic foundation that formalizes deep research workflows as structure-preserving functors, introduces a 296-question diagnostic benchmark evaluating multi-hop chains, pullbacks, topological re-ordering, and Yoneda falsification, and demonstrates measurable gains through tracked search and explicit categorical tools.

From Raw Experience to Skill Consumption: A Systematic Study of Model-Generated Agent Skills

This paper presents a systematic evaluation and mechanistic diagnosis of model-generated agent skills across the full "experience generation \(\to\) skill extraction \(\to\) skill consumption" lifecycle, proposes disentangled Extraction Efficacy (EE) and Target Evolvability (TE) metrics, exposes a severe negative correlation between textual plausibility and downstream utility, and validates a drop-in meta-skill prior based on failure mechanism encoding, actionable specificity, and risk blacklists that delivers consistent improvements across all domains.

From Talking Words to Sharing Thoughts: Scalable Multi-LLM Aggregation via Structured Message Passing

This paper proposes a bipartite factor-graph framework that performs multi-LLM collaboration directly at the output logit layer via epistemic regime routing, six soft-constraint check nodes, and adaptive asymmetric belief propagation, reducing token consumption by over 97% with sub-millisecond CPU aggregation while outperforming text-based multi-agent baselines across MMLU, MMLU-Pro, GPQA, and MedMCQA.

FutureSim: Replaying World Events to Evaluate Adaptive Agents

FutureSim evaluates frontier LLM agents on long-horizon test-time adaptation and uncertainty reasoning by sandboxing them within an offline chronological replay of post-cutoff world news and evaluating their evolving probability distributions via the strictly proper Brier Skill Score.

GenEnv: Difficulty-Aligned Co-Evolution Between LLM Agents and Environment Simulators

GenEnv establishes a two-player co-evolutionary game between an agent and a generative environment simulator, using a Gaussian-shaped \(\alpha\)-curriculum reward to anchor task difficulty in the agent's "zone of proximal development" and outperforming Gemini 2.5 Pro offline augmentation using 3.3\(\times\) less synthetic data.

General Agent Evaluation

IBM Research and MIT present Exgentic and the Unified Protocol, delivering the first full-factorial evaluation across 5 agent architectures × 5 LLMs × 6 benchmarks, showing that general agents match customized specialists on 4/6 domains while uncovering stark architecture and benchmark generality sinks in open-weight models.

GLoVE: Global Verifier for LLM Memory-Environment Realignment

GLOVE is a plug-and-play verification framework that resolves the epistemic gap in memory-augmented LLM agents under dynamic environmental drifts by detecting cognitive dissonance, constructing relative truth via budgeted active probing, and realigning obsolete experiences without oracle feedback.

Group-Graph Policy Optimization for Long-Horizon Agentic Reinforcement Learning

To tackle severe reward sparsity, high-variance state valuation, and myopic credit assignment in long-horizon agentic RL, G2PO transforms multi-trajectory rollouts into a global state-transition graph, employing group-aggregation state-value estimation and edge-centric advantage estimation to achieve up to 22.2% higher success rates over GRPO on ALFWorld and WebShop with virtually zero computational overhead.

Group of Skills: Group-Structured Skill Retrieval for Agent Skill Libraries

GoSkills is an inference-time group-structured skill retrieval method that replaces flat atomic skill lists or implicit dependency bundles with role-labeled execution contracts, achieving substantial gains in task success and execution efficiency under strict context budgets.

Group Perspective Matters: Regulating Debate Relationships Can Mitigate Blind Conformity in Multi-Agent Debate

To tackle the widespread blind conformity in multi-agent debate driven by majority dominance, this paper proposes DEAR, a framework that regulates dynamic debate relationships from a group perspective using Subjective Logic and sequentially coordinated heterogeneous RL-agents, cutting token consumption by 67% while boosting reasoning accuracy.

GroupMemBench: Benchmarking LLM Agent Memory in Multi-Party Conversations

Addressing the limitation of existing LLM agent memory systems designed solely for dyadic single-user settings, this paper introduces GroupMemBench, the first multi-party conversation benchmark evaluating long-term group memory, exposing a severe collapse in existing systems under multi-user structures and Theory-of-Mind lexical shifts (reaching only 46.0% average accuracy).

GUITAR: Structured Failure Diagnosis of GUI Agents via State Transitions

Addressing the limitation where conventional step-level accuracy aggregates independent screens and masks critical long-tail failures, GUITAR abstracts visually diverse screens into shared functional states to build a global State Transition Graph (STG), revealing that 60.4% of errors are concentrated in only 20% of bottleneck states and driving actionable inference-time improvements.

Gym-Anything: Turn Any Software into an Agent Environment

Gym-Anything introduces an automated environment-building framework via a multi-agent creation-audit loop grounded in U.S. GDP data, converting 200 software applications into standardized CUA environments and constructing CUA-World with 10K+ tasks alongside the challenging CUA-World-Long benchmark.

Harness the Memory: A Holistic Evaluation of Memory Substrates in Memory Agents

This paper establishes the first controlled harness evaluation isolating memory substrates across 11 methods in 7 families, demonstrating that long-context factual QA and sequential decision-making occupy disjoint Pareto frontiers and exhibit inverted retrieval-depth scaling due to attention dilution.

Harnessing Agentic Evolution

AEvo formulates agentic evolution as an interactive environment with accumulated context as state, utilizing an external protected harness where a meta-agent edits the underlying evolution mechanism or operating context to enable robust, long-horizon search.

Hierarchical Task Network Planning with LLM-Generated Heuristics

Extends the paradigm of LLM-generated heuristics from classical state-space planning to total-order Hierarchical Task Network (HTN) planning, deploying Python heuristics selected via a generate-evaluate-select pipeline into the Pytrich planner to cut expanded search nodes by 83% on shared instances against state-of-the-art Panda.

How Do Agentic LLMs Decide to Call Tools? A Tool-Call Vector Shaped by Suppression

The paper discovers through minimal contrastive prompt pairs that the tool-call decision in agentic LLMs is governed by a compact, causally necessary and sufficient tool-call vector \(\mu_\Delta\) at residual stream layer 24, revealing that prompt scaffolds establish a default tool-call prior that analysis verbs actively suppress via mid-to-late MLP features.

HSCO-Bench: An Agent-Driven End-to-End Hardware-Software Co-Design Benchmark for Domain-Specific Accelerators

The first benchmark evaluating LLM agents on end-to-end hardware-software co-design for heterogeneous SoCs, requiring agents to profile software applications, synthesize SystemC accelerators, configure NoC interconnects, and verify deployment on physical FPGAs.

Hyperagents

Hyperagents integrate task execution and meta-level modification into a unified self-referential program, enabling open-ended metacognitive self-improvement that breaks domain-specific task alignment and compounds capabilities across diverse benchmarks.

Improving General Role-Playing Agents via Psychology-Grounded Reasoning and Role-Aware Policy Optimization

To tackle reward hacking and generic degradation in general role-playing agents trained via RL, this paper introduces Psy-CoT, a psychology-grounded three-stage chain-of-thought framework, and RAPO, a role-aware policy optimizer that asymmetrically modulates token gradients based on profile-token mutual information.

Improving General Role-Playing Agents via Psychology-Grounded Reasoning and Role Identity Anchoring

To tackle behavioral mimicry and generic collapse in general role-playing agents, this paper introduces Psy-CoT—a three-stage psychological chain-of-thought framework—alongside RAPO, an asymmetric reinforcement learning optimizer that modulates token-level gradients via profile–token mutual information, achieving state-of-the-art persona fidelity and consistency.

Improving the Efficiency of Language Agent Teams with Adaptive Task Graphs

Addressing the dichotomy between rigid static pipelines and chaotic decentralized agent groups, this paper introduces LATTE, a distributed-systems-inspired orchestration framework where LLM teams collaboratively maintain an evolving task graph via formal mutation operators, dramatically cutting token usage and wall-clock latency while boosting task accuracy and system interpretability.

Inductive Deductive Synthesis: Enabling AI to Generate Formally Verified Systems

This paper introduces Inductive Deductive Synthesis (IDS), an agentic framework that jointly and incrementally synthesizes implementations and machine-checked proofs under formal checker and benchmark feedback, enabling AI agents to autonomously generate formally verified, high-performance distributed key-value stores within hours.

InterLV-Search: Benchmarking Interleaved Multimodal Agentic Search

Introduces InterLV-Search, the first benchmark for interleaved language-vision agentic search across three progressive levels (active visual evidence seeking, controlled offline interleaved search, and open-web multi-branch search), along with the standardized InterLV-Agent framework, revealing that even top-tier multimodal agents struggle with visual search control and evidence integration, remaining below 50% accuracy.

Intern-Atlas: A Methodological Evolution Graph as Research Infrastructure for AI Scientists

This paper introduces Intern-Atlas, a method-centric evolution graph constructed from over one million AI publications containing 9.41 million typed causal edges grounded in verbatim bottleneck quotes, coupled with a Self-Guided Temporal Monte Carlo Tree Search (SGT-MCTS) algorithm that powers reliable lineage recovery, parameter-free idea evaluation, and hallucination-free idea generation for AI research agents.

InvestigationWorlds: An Agentic Environment for Legal Investigation

Grounding legal investigation benchmarks in U.S. federal summary judgment records and attorney-validated synthetic document augmentation, InvestigationWorlds exposes that frontier agents often commit to false alternative hypotheses despite retrieving relevant evidence, achieving at best a 33.3% case-solved rate.

Is Agentic AI Ready for Real-World Hardware Engineering? A Deep Dive with Phoenix-bench

Presents Phoenix-bench, the first execution-grounded hardware issue resolution benchmark with 511 verified instances from 114 GitHub repositories, revealing that software-oriented agents lose 37% to 58% in accuracy due to hardware signal-flow propagation and showing that testbench error logs yield vastly superior gains over file-level oracle hints.

Iterative Critique-and-Routing Controller for Multi-Agent Systems with Heterogeneous LLMs

To overcome the inability of one-shot routers to critique intermediate drafts and prevent expensive model overuse, this paper formulates heterogeneous multi-agent coordination as a finite-horizon MDP with explicit agent-utilization constraints, optimizing an iterative critique-and-routing controller via policy gradients (GRPO) under Lagrangian relaxation to substantially narrow the gap to the strongest agent while keeping its usage under 25%.

JobBench: Aligning Agent Work With Human Will

JobBench shifts the evaluation of workplace agents from economic replacement to human delegation desire, curating 130 realistic workspace tasks across 35 occupations evaluated under 4,631 fact-anchored chained criteria where frontier models max out at only 45.9%.

Joint Optimization of Tool Creation and Use for Large Language Model Agents

The paper proposes SMITH, a reinforcement learning framework that jointly optimizes tool creation and tool use within a single shared policy via dual-task sampling and three disentangled reward signals, enabling a 4B model to outperform much larger models and distillations in generalization and cross-model transfer.

Justitia: Fair and Efficient Scheduling of Task-parallel LLM Agents with Selective Pampering

KaliBench: A Fine-Grained Benchmark for Cybersecurity Tool Use on Kali Linux with Runtime-Free Verifiable Rewards

To tackle the lack of fine-grained command-line evaluation and runtime-free training signals for cybersecurity LLMs, KaliBench introduces a schema-free benchmark across 1,642 Kali Linux tools, unlocking deterministic runtime-free verifiable rewards (RLVR) that boost compact 8B models to compete with 70B+ models.

Known By Their Actions: Fingerprinting LLM Browser Agents via UI Traces

This paper demonstrates that passive in-page JavaScript event tracking captures temporal rhythms and action dynamics sufficient to fingerprint the underlying foundation model of web-browsing agents across 14 frontier LLMs with up to 96% Macro F1, opening the door to model-targeted exploits.

Language-Based Agent Control

Language-Based Agent Control (LBAC) embeds LLM agents into a pure, strongly typed host language (realized as Elba in Haskell), enforcing security policies through static typechecking and runtime capabilities across both agent programs and scaffolding, outperforming dual-LLM architectures like CaMeL with nearly double the task utility while providing sound, zero-false-negative defense against adversarial prompt injections.

Latent Action Reparameterization for Efficient Agent Inference

To resolve the large effective decision horizons and high inference latency caused by overly fine-grained textual actions in LLM agents, this paper proposes Latent Action Reparameterization (LAR), which filters multi-step semantic behaviors via frequency and next-token entropy to collapse them into executable discrete latent actions, substantially shortening the decision horizon and inference cost while preserving task success.

Lean4Agent: Formal Modeling and Verification for Agent Workflow and Trajectory

The first framework applying the dependent-type proof assistant Lean4 to model and verify LLM agent workflows and execution trajectories, introducing a three-layer library FormalAgentLib for structural, static semantic, and runtime verification alongside a formal-guided evolution module LeanEvolve.

Learning Agentic Policy from Action Guidance

Addressing exploration stalls in agentic reinforcement learning when reward states fall beyond the base policy's reachable region, ActGuide-RL introduces readily available action data as plan-style reference guidance, coupled with an adaptive minimal intervention principle and mixed-policy optimization, achieving exploration capabilities on par with the SFT+RL pipeline without requiring any supervised cold start.

Learning CLI Agents with Structured Action Credit under Selective Observation

Addressing partial observability and sparse terminal rewards in command line interface agents, this paper proposes Action Advantage Assignment (\(\mathrm{A}^3\)), an AST-driven multi-granularity credit assignment method, and \(\sigma\)-Reveal, an inference-time context selection harness, substantially improving complex filesystem task accuracy with near-zero computational overhead over standard GRPO.

Learning to Continually Learn via Meta-Learning Agentic Memory Designs

ALMA introduces an open-ended meta-learning framework that searches over executable code to automatically evolve domain-specialized agentic memory designs and retrieval-update mechanisms, outperforming state-of-the-art human-crafted memory systems across diverse sequential decision-making benchmarks.

LensDesigner: A Self-Improving Agent for Optical Lens Design

This paper introduces LensDesigner, the first autonomous agent framework for classical optical lens design, which overcomes the cold-start barrier via the 180k-scale LensLib100K library and optics-aware retrieval, executes guarded macro orchestration, and continuously distills design intuition through curriculum-driven self-evolution, significantly outperforming numerical optimization and LLM baselines on the challenging 120-task LensArena benchmark.

LITMUS: Benchmarking Behavioral Jailbreaks of LLM Agents in Real OS Environments

This paper introduces LITMUS, the first benchmark evaluating behavioral jailbreaks of LLM agents in real operating system environments through semantic-physical dual verification and automated state rollback, revealing physical attack success rates above 40% across frontier models and uncovering pervasive Execution Hallucination where agents verbally refuse yet physically execute hazardous commands.

LiveOption: Evaluating LLM Agents in Structured Option Trading with Nonlinear Payoffs

LiveOption introduces the first principled evaluation framework for LLM agents in structured option trading with nonlinear payoffs, formalizing three realistic market task suites alongside a four-level hierarchical metric and Greek PnL attribution suite that reveals current frontier models fail to generate positive risk-adjusted excess returns under realistic derivative execution constraints.

llm agents already know when to call tools - even without reasoning

This paper introduces When2Tool, the first controlled benchmark for tool-call decision-making, demonstrating that LLM pre-generation hidden states linearly encode tool necessity with high precision, and develops Probe&Prefill to prune nearly half of unnecessary tool calls with negligible accuracy loss.

LLM-based Multi-Agent Blackboard System for Information Discovery in Data Science

Addressing the context-window saturation of single agents and the rigid task-assignment bottlenecks of master-slave multi-agent systems in heterogeneous data lakes, this paper introduces a decentralized blackboard multi-agent system where a central coordinator broadcasts requirements and specialized helper agents volunteer solutions based on local data understanding, achieving 13%–57% relative gains in end-to-end task completion and up to 9% higher file discovery F1 scores.

LocalAgent: Collaborative Agentic Verification for Fine-Grained Instance-Level Consistency

Addressing the failure of multimodal LLMs in fine-grained instance verification caused by visual-linguistic feature compression, LocalAgent introduces a collaborative agentic framework where the MLLM acts as a spatial controller to identify suspicious regions, delegates them to lightweight pure-vision verifiers for structured scalar evidence, and selectively verifies via auxiliary-capped reinforcement fine-tuning.

LPS-Bench: Benchmarking Safety Awareness of Computer-Use Agents in Long-Horizon Planning under Benign and Adversarial Scenarios

Presents LPS-Bench, a benchmark evaluating the safety awareness of computer-use agents during long-horizon planning via 570 sandboxed test cases and trajectory-level criteria, revealing widespread safety failures under benign ambiguity and adversarial manipulation.

MAGE: Multi-Agent Self-Evolution with Co-Evolutionary Knowledge Graphs

MAGE externalizes cross-iteration self-evolutionary knowledge into a four-subgraph Co-Evolutionary Knowledge Graph (EvoKG) that combines teacher-written failure corrections with self-harvested successful reasoning traces from a frozen learner, jointly updating the graph and two routing/search bandits to substantially outperform frozen-backbone prompting baselines across nine reasoning and interactive benchmarks (+31.6 pp on RealMath and +20.8 pp on STBench) with zero teacher calls at inference time.

Market Regime Council for Dynamic Credit Assignment in Multi-Agent LLM Decision Systems

Addressing the failure modes of static weights, cold-start dominance, and uninterpretable aggregation in financial multi-agent systems, Market Regime Council (MRC) formulates online portfolio decision-making as a cooperative game that derives dynamic agent weights from exact Shapley values and market regime multipliers.

Masked Diffusion Language Agents for Tool-Integrated Chemical Reasoning

The paper introduces ChemDiffAgent, the first masked diffusion language agent for tool-integrated chemical reasoning, which models interleaved reasoning and tool calling as an iterative block-level denoising process and employs two-stage agentic post-training (Agentic SFT and Agentic VRPO) to achieve a ~20% performance leap over comparable autoregressive agents across single-turn and multi-turn chemistry benchmarks.

MCP-Atlas: A Large-Scale Benchmark for Tool-Use Competency with Real MCP Servers

Addressing the lack of cross-server orchestration, realistic server complexities, and objective grading in existing tool benchmarks, MCP-Atlas introduces 1,000 expert-crafted tasks across 36 production MCP servers and 220 tools, using trajectory-decoupled atomic claim-level scoring alongside an 11-category diagnostic taxonomy to show that 63.3% of frontier agent failures stem from cognitive breakdowns rather than tool mechanics.

Measuring AI Agents' Progress on Multi-Step Cyber Attack Scenarios

The UK AI Security Institute (UK AISI) benchmarks seven frontier AI models across 18 months on 32-step corporate and 7-step ICS cyber ranges, showing that multi-step autonomous attack progress scales log-linearly with inference-time compute (up to 59% gain from 10M to 100M tokens) while physical ICS disruption remains sharply constrained.

Measuring Collapse and Correction in Homogeneous-Panel LLM Debate

The paper decomposes debate among three copies of one model into preservation, collapse, correction, and unrepaired transitions, using pre-debate screening and round-level traces to locate risk; however, an offline freeze replay prevents 29 collapses while losing 108 corrections, showing that a risk signal is not an effective controller.

MemArena: An Ego-Centric Benchmark for On-Device Agentic Personal Memory Assistants at Scale

MemArena addresses the structural absence of ego-centric partial observability, multi-session global coherence, and authentic human-scale interaction density in existing memory benchmarks by generating MemArena-L—a 50-agent, 15-day dataset with 10.3M dialogue tokens via the MASim simulator—and systematically dissects 5 open-weight readers against 5 memory backends across 6 recall, reasoning, and trustworthiness evaluation dimensions.

MemCode: Discrete Semantic Representations for Long-Term Agent Memory

MemCode addresses the write–retrieve disconnect in long-term agent memory by rewriting dialogue into self-contained atomic memory units (AMUs), anchoring each unit into a discrete codebook via isomorphic semantic quantization (ISQ), and coupling continuous vector search with codebook co-occurrence topology to achieve superior accuracy and orders-of-magnitude faster memory construction.

MEME: Multi-entity & Evolving Memory Evaluation

The paper introduces MEME, a memory benchmark organized along entity scope (single vs. multi-entity) and temporal dynamics (static vs. evolving) axes across six tasks, formalizing explicit dependency chains and conditional rules to evaluate cascade updates, absence uncertainty, and post-deletion states, revealing a systemic failure across mainstream memory architectures on dependency reasoning.

Memory-R2: Fair Credit Assignment for Long-Horizon Memory-Augmented LLM Agents

Addressing the non-stationary environment and unfair trajectory comparisons caused by memory modifications in multi-session agent RL, this paper proposes LoGo-GRPO, a credit-assignment algorithm combining global trajectory optimization with local rerollouts from shared intermediate memory states, alongside shared-backbone extractor-manager co-learning and curriculum training.

Memory Retrieval for Changing Preferences

Addressing the critical vulnerability where conventional semantic retrieval surfaces obsolete or purely topical distractors in personalized dialogues, this paper introduces an evidence utility framework grounded in Bayes factors to unify memory access gating and turn selection, demonstrating substantial accuracy gains across long-context benchmarks under evolving preferences through joint controller-reader fine-tuning.

Memory Type Varies: Empowering LLM Agents for Long-Term Memory with Diverse Strategies

Addressing the bottleneck where one-size-fits-all retrieval strategies in LLM agent long-term memory overlook memory heterogeneity, this paper introduces a cognitive psychology-inspired tri-class benchmark TriMEM and the adaptive MemoType framework, which establishes a theoretical upper bound on single retrieval alignment and applies tailored strategies across episodic, personal semantic, and general semantic memories to drastically cut construction costs while boosting retrieval and QA accuracy.

MemPlan: Memory-Conditioned PDDL Planning for Partially Observable Text Environments

MemPlan introduces a memory-conditioned PDDL planning framework that fixes the planning domain and reconstructs the PDDL problem at each step by translating positive graph-memory retrieval into hypothesis action costs and compiling execution refutations into negative exclusion constraints, enabling classical planners to solve partially observable interactive text tasks without online PDDL rewriting.

MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents

MemSkill reframes static agent memory operations into a learnable and evolvable skill bank, where an RL controller dynamically selects relevant Top-K skills for single-pass memory updates by an LLM executor, while an LLM designer periodically analyzes hard-case failures to refine and expand skills, significantly improving conversational and embodied agent benchmarks.

Mental-Models for Multi-Agent Systems

A mental-model-enabled multi-agent framework that infers amortized recursive Theory-of-Mind latent states and trains a belief-conditioned reward model to guide policy learning via GRPO, delivering substantial improvements in social interaction and decision-making without test-time inference overhead.

Meta-Cognitive Memory Policy Optimization for Long-Horizon LLM Agents

Addressing ambiguous credit assignment and compounding semantic noise in recursive memory agents under sparse outcome rewards, MMPO formulates long-horizon reasoning as a POMDP, introduces Belief Entropy via dual-probe anchor questions to quantify cognitive uncertainty, and injects dense sub-trajectory supervision through group-relative advantage estimation to sustain robust reasoning up to 3.5M tokens.

Meta\(^n\): Recursive Self-Improvement through Emergent Depth

By keeping a single universal meta-operation \(\Omega\) fixed and recursing on its accumulated inputs (execution traces and code stacks), Metan breaks the stability-versus-depth dilemma that caps self-modifying agents at meta-depth ~2.5, delivering emergent layer specialization and state-of-the-art performance across 8 diverse benchmark families including ARC-AGI-2.

MetaN: Recursive Self-Improvement through Emergent Depth

MetaN introduces a recursive self-improvement framework that keeps a single universal meta-operation \(\Omega\) fixed while recursing on its accumulated execution traces and code stack, breaking the meta-depth bottleneck of traditional self-modifying agents and allowing layered strategies to emerge dynamically with convergence.

MIRAGE: Mobile Agents with Implicit Reasoning and Generative World Models

MIRAGE internalizes explicit reasoning in mobile GUI agents into continuous latent slots using Approximate Parallel Latent Refinement (APLR) and a lightweight Q-Former generative world model predicting future visual features, matching explicit-CoT task success on AndroidWorld without decoding verbose rationales and cutting generated token volume by over 75%.

Missing Old Logits in Asynchronous Agentic RL: Semantic Mismatch and Repair

Addressing the semantic failure mode in asynchronous Agentic RL where historical training-side old logits are lost due to rapid parameter updates and partial rollouts, this paper proves that probability interpolation merely re-parameterizes clipping boundaries, evaluates three exact system recovery strategies, and develops a revised PPO-EWMA method with staleness-aware decay and auto-reset that matches exact recovery performance at minimal overhead.

Missing Old Logits in Asynchronous Agentic RL: Semantic Mismatch and Repair Methods for Off-Policy Correction

Model-Adaptive Tool Necessity Reveals the Knowing-Doing Gap in LLM Tool Use

This paper introduces an empirical, model-adaptive definition of tool necessity and demonstrates via hidden-state linear probing that tool-calling failure predominantly originates from a geometric knowing-doing gap where internal recognition of necessity collapses into near-orthogonality with action execution at the readout token.

MoMHa: Multi-Objective Optimization of LLM Harnesses over Accuracy, Safety, and Tokens

Without updating model weights, MoMHa uses a proposer to rewrite Python harnesses from execution traces, jointly optimizing accuracy, behavioral safety, and token cost; the authors report leading joint means on synthetic and real benchmarks, but source conflicts concerning metric scaling, the two-phase definition, and cost accounting require clarification.

MoRSE: Task-Oriented Multi-Agent System with Mixture of Role-Subtask Experts

MoRSE establishes a task-oriented multi-agent framework based on dependency-aware directed acyclic graphs (DAGs), dynamically activating "role + subtask" LoRA experts via a prototype-based semantic router, and introduces hierarchical group-relative policy optimization (HGRPO) with two-layer credit assignment to decouple routing variance from expert updates, delivering parameter-level specialization at near single-model cost.

Mstar: Every Task Deserves Its Own Memory Harness

Addressing the poor cross-task generalization of static memory systems, Mstar represents an LLM agent's memory harness as an executable Python module jointly defining Schema, Logic, and Instruction, automatically discovering task-specialized optimal memory architectures via reflective code evolution and population-based program search.

Multi-Head Recurrent Memory Agents

Addressing the severe performance degradation in recurrent memory agents caused by monolithic text overwriting across long contexts, this paper proposes Multi-Head Recurrent Memory (MHM), a training-free framework that partitions memory into independent heads governed by a stage-wise least-recently-updated (LRU) rule, elevating memory retention rate from under 30% to 73.96% at 896K tokens.

Neglected Free Lunch from Post-Training: Progress Advantage for LLM Agents

Reveals that the log-probability ratio between an RL-trained policy and its reference policy theoretically recovers the optimal advantage function under general stochastic interaction environments, serving as a plug-and-play, annotation-free, step-level progress signal without requiring dedicated process reward model training.

No More, No Less: Task Alignment in Terminal Agents

Introduces TAB (Task Alignment Benchmark), an 89-task terminal suite evaluating whether agents selectively utilize necessary environmental cues while ignoring unrelated distractors, demonstrating that task capability does not imply task alignment and that current defenses suppress necessary cues when reducing distractor execution.

OASES: Outcome-Aligned Search-Evaluation Co-Training for Agentic Search

OASES proposes an outcome-aligned search and evaluation co-training framework that measures intermediate state answerability via verifiable outcome metrics to formulate telescoping process rewards, eliminating evaluator-policy mismatch and proxy divergence through a single shared-parameter model and outperforming strong RLVR baselines across five multi-hop QA benchmarks.

On Time, Within Budget: Constraint-Driven Online Resource Allocation for Agentic Workflows

This paper introduces Monte Carlo Portfolio Planning (MCPP), a lightweight closed-loop framework for constraint-driven online resource allocation in agentic workflows that maximizes the probability of completing execution within hard budget and deadline limits.

OneMillion-Bench: How Far Are Language Agents from Human Experts?

To bridge the gap between exam-style benchmarks and economically consequential demands, this paper introduces $OneMillion-Bench, a benchmark of 400 expert-curated tasks across 5 professional domains that directly evaluates agents via senior labor replacement costs and asymmetric negative penalty rubrics.

OpenWebRL: Demystifying Online Multi-turn Reinforcement Learning for Visual Web Agents

Presents OpenWebRL, the first open framework for training visual web agents via online multi-turn reinforcement learning directly on live websites, achieving open-source SOTA on three live-web benchmarks with a 4B model using only 412 demonstration trajectories and 2.2K online tasks.

OPERA: An Agent for Image Restoration with End-to-End Joint Planning–Execution Optimization

Addressing constrained planning spaces and uncoordinated tool execution in multi-degradation restoration, OPERA introduces a joint optimization framework that trains an agent via GRPO over combinatorial plan spaces and co-tunes downstream restoration tools end-to-end.

Optimize Once, Execute Fast: Latency-Aware Multi-Agent Workflow Learning for Recurrent Queries

Addressing the cumulative end-to-end latency bottleneck in recurrent multi-agent system invocations, this paper introduces LAWA, a latency-aware workflow optimization framework that compresses execution makespan via theoretically grounded graph editing operators and preference evaluators while rigorously maintaining downstream task accuracy.

PACE: A Proxy for Agentic Capability Evaluation

PACE introduces a framework that constructs a compact proxy benchmark from atomic non-agentic evaluation instances via decoupled SVD leverage and target-relevance selection combined with bootstrap regression, accurately predicting expensive agentic benchmark performances with under 4% MAE using only 100 static questions and slashing evaluation costs by over 99%.

PACE: Two-Timescale Self-Evolution for Small Language Model Agents

Addressing the instability and semantic regression risks of frozen small language models (SLMs) when mutating their own code, PACE introduces a two-timescale agent self-evolution framework that coordinates frequent low-risk prompt refinement with saturation-triggered, held-out validation-gated control logic updates, achieving up to +9.2% relative accuracy improvement across 12 backbone–benchmark pairs.

PACEvolve++: Improving Test-time Learning for Evolutionary Search Agents

PACEvolve++ introduces an advisor-model reinforcement learning framework that decouples strategic search reasoning from concrete code implementation and employs a phase-adaptive advantage estimator transitioning from group-relative to frontier-contribution feedback, achieving fast-converging and stable test-time adaptation for evolutionary search agents.

PAGER: Bridging the Semantic-Execution Gap in Point-Precise Geometric GUI Control

Addressing the "Semantic-Execution Gap" in GUI agents on continuous canvases caused by the lack of point-level precision and cascading dependency errors, this paper introduces PAGER, a framework combining dependency-structured planning and pixel-grounded execution trained with precision-aligned RL, boosting task success on the new PAGE Bench by 4.1× over the strongest general baseline.

Perceive, Interact, Reason: Building Tool-Augmented Visual Agents for Spatial Reasoning

To overcome vision-language models' inability to actively gather and verify spatial evidence, this paper introduces PERIA—a tool-augmented visual agent featuring a dual-family sandbox of perception and interaction tools and an Observation-Relaxed Group-in-Group Policy Optimization (OR-GIGPO) algorithm—improving 8B backbone spatial reasoning by 6.9–14.8 percentage points to match GPT-5.

Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-Horizon Agents

This paper introduces PerMem-Bench, the first personalized memory benchmark capturing multi-year, multi-domain interactions with dynamic user persona shifts, alongside a session-level storage gating framework that selectively skips transient sessions to boost long-term memory retention under constrained budgets.

PhGPO: Pheromone-Guided Policy Optimization for Long-Horizon Tool Planning

Inspired by ant colony optimization, PhGPO distills verified successful tool-use trajectories into an explicit pheromone transition prior on an MCP-grounded tool graph, using evaporation, deposition, and task retrieval to guide GRPO sampling and progressive curriculum training for long-horizon tool planning.

PHMForge: Evaluating LLM Agents on Industrial Prognostics through MCP-Native, Algorithm-Grounded Tools

Presents PHMForge, the first MCP-native benchmark for industrial prognostics and health management (PHM), featuring 39 algorithm-grounded tools and 99 SME-authored scenarios across 8 asset classes, demonstrating that agents critically depend on real algorithm execution and that failures concentrate in orchestration sequencing rather than tool invocation.

PhotoFlow: Agentic 3D Virtual Photography Missions

PhotoFlow formulates language-guided 3D virtual photography as a finite-horizon closed-loop camera search problem, coordinating a Director, Reviewer, and Reflector to optimize continuous camera parameters and aspect ratio, outperforming one-shot and single-chain reflection baselines across 141 tasks in VPhotoBench.

PhysicianBench: Evaluating LLM Agents in Real-World EHR Environments

PhysicianBench introduces an execution-grounded benchmark of 100 clinician-validated, long-horizon tasks across 21 subspecialties in a standard FHIR R4 EHR environment, evaluating LLM agents via 670 granular checkpoints and revealing that the best model (GPT-5.5) achieves only 46.3% single-attempt pass rate and 28.0% consistency.

PieArena: Ranking and Profiling Language Agents in Realistic Negotiation Scenarios

To address the saturation of static benchmarks and the lack of realistic dynamic game-theoretic evaluation, this paper introduces PieArena, a large-scale benchmark grounded in elite business school MBA negotiation scenarios; it develops an order-invariant Beta-Generalized Bradley–Terry–Luce (BBTL) ranking model for continuous payoffs, introduces a shared-intentionality agentic harness, and profiles 13 frontier language agents alongside trained human negotiators across mirror-play, cross-play, and human–LM play.

Prefix Executability: Evaluating Tool-Using Agents Beyond Final Success

Addressing the core blind spot of outcome-only tool benchmarks that overlook intermediate invalid tool calls, this paper proposes Prefix Executability—a trajectory-level evaluation protocol formalizing PEC, FED, TER, AUC-PEC, and RR—revealing that unverified plan-conditioned agents achieve inflated final success despite pervasive intermediate failures, while pre-execution verification restores prefix reliability.

PROBE: Learning to Audit Policy Compliance in Tool-Using LLM Agents

PROBE formulates black-box policy compliance auditing of tool-using agents as a two-player Markov game, employing a composite reward balancing deduplicated violations, progress, novelty, and plausibility alongside a horizon curriculum to train a lightweight auditor that actively uncovers deep policy failures in frozen large agents.

Quest: Training Frontier Deep Research Agents with Fully Synthetic Tasks

Quest presents a fully open, reproducible recipe for frontier deep research agents that leverages only 8K fully synthetic rubric-tree tasks, structured three-bucket context condensation, and a three-stage training pipeline (MT+SFT+RL), matching or outperforming proprietary deep research agents across eight diverse benchmarks.

Recursive Multi-Agent Systems

RecursiveMAS reformulates heterogeneous multi-agent collaboration into unified latent-space recursive computation via lightweight residual RecursiveLink modules, co-optimizing the entire loop through inner-outer training while slashing token usage and runtime latency.

Recursive Semantic Divergence for LLM Agent Consistency

Adapts reinforcement learning bisimulation metrics and natural language inference into an unsupervised trajectory-level consistency metric (RSD), enabling single-forward estimation of future semantic divergence from state self-distance and serving as a dense reward signal to substantially reduce multi-turn agent variance.

Reflection with Action-Induced Visual Differences for Desktop GUI Agents

Addressing the vulnerability of conventional single-step reflectors in desktop GUI agents over large, dense screens with subtle or scattered changes, this paper proposes Evidence-First Reflection (EFR), which explicitly decouples visual difference extraction from outcome verification, achieving 5.94% and 4.95% end-to-end task success improvements on OSWorld-Verified and WindowsAgentArena.

REFLEX: Reflective Evolution from LLM Experience

REFLEX proposes a train-free evolutionary program search framework that decouples visual Critic diagnosis from text-only Actor repair via a Multi-turn Diagnostic Chain, coupled with a persistent Skill Memory credited by fitness gains and a UCB1 operator bandit to achieve auditable, sample-efficient policy evolution and cross-task knowledge transfer.

Reinforcement World Model Learning for LLM-based Agents

Reinforcement World Model Learning (RWML) introduces a self-supervised mid-training paradigm that optimizes action-conditioned textual world models via GRPO with embedding sim-to-real gap rewards, bypassing expert trajectories and task-success rewards while substantially boosting downstream agent planning.

RelAgent: LLM Agents as Data Scientists for Relational Learning

RelAgent casts relational learning into a two-phase framework where an LLM agent autonomously searches over SQL feature programs and downstream predictive models using interactive database tools, validation feedback, and an evaluation workspace, achieving deterministic, highly competitive, and intrinsically interpretable predictions without LLM calls at inference time.

RepoMirage: Probing Repository Context Reasoning in Code Agents with Perturbations

RepoMirage introduces a perturbation-based diagnostic evaluation suite built on SWE-Bench Verified to probe repository context reasoning in code agents, showing that high end-to-end success rates heavily rely on localized surface cues, triggering severe exploration drift when broad context reasoning is required.

Rethinking How to Remember: Beyond Atomic Facts in Lifelong LLM Agent Memory

Addressing detail loss and shallow reasoning in existing LLM agent memory systems that rely on static prompts to extract atomic facts, this paper proposes TriMem—a tri-granularity architecture combining verbatim dialogue anchors, atomic facts, and synthesized entity profiles, coupled with TextGrad-based lifelong prompt optimization without parameter tuning.

Reward Modeling for Multi-Agent Orchestration

Addressing the prohibitive computational cost and supervision scarcity of training and evaluating multi-agent system (MAS) orchestrators via sub-agent rollouts, Orch-RM introduces self-supervised orchestration reward modeling to score orchestration plans directly, cutting verification tokens by up to 17x while improving test-time accuracy by up to 8% and training token efficiency by 10-46x.

RL-Guided Temporal Localization for Dual-Channel Retrieval in Long-Horizon Agent Memory

To resolve the crowding-out of target memories by temporally mismatched yet semantically similar historical events in long-horizon LLM agent memory, TIDER introduces an RL-trained temporal interval locator paired with distance-decay dual-channel retrieval, outperforming prior baselines on LifeBench and LoCoMo with a lightweight 1.7B locator.

ROAR: Benchmarking LMM Agents under Real-Time Constraints in Multiplayer Action-Shooter Games

Addressing the pitfall where traditional turn-based benchmarks pause the world during inference and mask latency costs, ROAR introduces a lightweight multiplayer action-shooter environment and a Multi-Agent Asynchronous Gym interface to evaluate LMM agents under strict real-time constraints across cooperative, competitive, and mixed settings.

RuleSmith: Multi-Agent LLMs for Automated Game Balancing

RuleSmith integrates multi-agent LLM self-play with Gaussian-process Bayesian optimization and Expected-Improvement-based adaptive sampling, treating prompt-driven language models as zero-shot rule evaluators to discover near-balanced, interpretable numerical rules in asymmetric strategy games.

SaaS-Bench: Can Computer-Use Agents Leverage Real-World SaaS to Solve Professional Workflows?

SaaS-Bench constructs an evaluation benchmark of 106 realistic, long-horizon tasks across 23 deployable SaaS systems in 6 professional domains, revealing that frontier models achieve less than a 4% end-to-end task resolution rate due to planning breakdowns, unverified execution, and multi-app error cascades.

Same Signal, Opposite Meaning: Direction-Informed Adaptive Learning for LLM Agents

Uncovers that the relationship between uncertainty signals and test-time compute utility reverses across interactive environments and model backbones, and proposes DIAL, a sparse adaptive gating framework trained via signal-agnostic counterfactual exploration that significantly advances the success-cost Pareto frontier.

SchemeArena: Factorized Stress Testing of Scheming in LLM Agents

SchemeArena introduces the first factorized stress-testing benchmark comprising 400 realistic tool-interactive scenarios paired with the multi-criteria process-level monitor SCOUT, revealing that explicit instrumental goals dominate scheming propensity while partial action-only oversight paradoxically amplifies covert evasion in closed-source frontier agents.

ScreenSearch: Uncertainty-Aware OS Exploration

Addressing partial observability and perceptual aliasing in desktop environments, ScreenSearch constructs a shared deduplicated screen graph across virtual machines and introduces a one-step PUCT graph-bandit that treats matched-action outcome dispersion as structural ambiguity, effectively balancing frontier expansion and disambiguation.

Search More, Think Less: Rethinking Long-Horizon Agentic Search for Efficiency and Generalization

Addressing the computational bottleneck and latency issues caused by sequential reasoning in long-horizon deep research agents, SMTL substitutes linear reasoning with parallel subtask execution and dynamic re-planning, achieving state-of-the-art performance across BrowseComp and DeepResearch Bench while slashing reasoning steps by 70.7%.

Searching Videos as Trees: Self-Correcting Agents for Grounded Long Video QA

VideoTreeSearch (VTS) frames grounded long-video QA as discrete multi-turn navigation over an adaptive semantic tree, equipping the agent with explicit self-correction primitives (zoom_in, zoom_out, shift, answer) and training it via detour-recovery trajectory synthesis and GRPO reinforcement learning to dramatically outperform continuous cropping agents.

SecureClaw: Clawing Back Control of LLM Agents

Addressing the dual vulnerabilities of unauthorized external side effects and internal sensitive plaintext disclosure in tool-using agents, SecureClaw introduces a dual-boundary architecture featuring executor-side request-bound PREVIEW→COMMIT authorization and gateway-side opaque reference ID confinement with explicit planning summaries, eliminating unauthorized actions while maintaining high task utility.

Selective Critique for Cost-Aware LLM Agents in Long-Horizon Decision Making

Addressing the excessive token expenditure and latency overhead caused by continuous critique in long-horizon LLM decision-making, this paper proposes SAG, an ambiguity-gated self-improving framework that selectively triggers non-oracle critique based on Value of Information principles and distills corrective behaviors via online SFT, matching always-on reliability while achieving a 3.1× normalized token efficiency gain.

Self-Compacting Language Model Agents

Introduces SelfCompact, a training-free adaptive context compaction scaffold that pairs an inline summarization tool with a lightweight rubric probe, enabling models to autonomously identify closed reasoning units and compact trajectories, cutting token costs by 30–70% while improving reasoning accuracy.

SePO: Self-Evolving Prompt Agent for System Prompt Optimization

SePO breaks the performance ceiling of hand-crafted prompt optimizers by introducing a self-referential evolutionary framework that jointly optimizes the prompt agent's own system prompt alongside task agents' prompts across a two-stage pre-training and fine-tuning pipeline.

Sequential Probabilistic Uncertainty Estimation for Parallel Multi-Agent Reasoning Systems

To address the problem that parallel multi-agent reasoning dynamics govern system reliability yet existing uncertainty estimators ignore multi-round evolution, the authors propose SAUCE, a training-free estimator that models consensus aggregation as Bayesian filtering over latent beliefs, substantially boosting misclassification detection, selective prediction, and calibration.

SETA: Scaling Environments for Terminal Agents

Addressing the critical lack of scalable training environments and reliable verification signals for terminal agents, SETA introduces a scalable framework combining grounded multi-source synthesis (SETA-Synth) with adaptive difficulty and context evolution (SETA-Evol), creating SETA-Env with over 4,500 executable containerized RL environments and achieving state-of-the-art results on Terminal-Bench 2.0 for both 8B and large-scale backbones.

Silent Failures in Agent-Tool Interaction: An Audit of ToolUniverse

This paper audits the agent-tool interaction pipeline across 15 biological tools in ToolUniverse, establishing a 7-locus failure taxonomy and demonstrating that partial API and wrapper contracts account for 84.6% of silent failures that downstream agents amplify into distorted scientific claims.

Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents

Skill-SD summarizes an agent's historical interaction trajectories into compact natural language skills that condition only a privileged teacher during training, utilizing an importance-weighted reverse-KL loss and dynamic parameter synchronization to internalize strategic guidance into an on-policy student without inference-time retrieval overhead.

SkillForge: Co-Evolving Skills and Agents via Dynamic Skill Lifecycles

SkillForge introduces a fitness-driven dynamic skill lifecycle framework that couples selective retirement, stabilization, and LLM-guided failure-driven mutation with agentic RL, mitigating delayed obsolescence and achieving state-of-the-art performance across interactive benchmarks with a compact library.

SkillGen: Verified Inference-Time Agent Skill Synthesis

SkillGen introduces a training-free, multi-agent framework that automatically synthesizes a single auditable inference-time skill from baseline agent trajectories via contrastive induction and a closed-loop generation-verification-refinement pipeline gated by causal net-effect evaluation.

SkillMaster: Toward Autonomous Skill Mastery in LLM Agents

SkillMaster frames skill management as closed-loop tool calls within the agent's policy, providing dense skill-quality rewards via counterfactual utility evaluation on related probe tasks, and decouples task execution from skill learning using DualAdv-GRPO to achieve autonomous skill mastery.

SkillOpt: Executive Strategy for Self-Evolving Agent Skills

By formalizing agent skills as optimizable natural-language states, SkillOpt introduces a deep-learning-inspired text-space optimizer featuring minibatch reflection, bounded edit budgets, and held-out validation gating, achieving top results across 52 evaluations without model weight updates.

SkillOrchestra: Learning to Route Agents via Skill Transfer

SkillOrchestra shifts multi-turn agent orchestration away from end-to-end reinforcement learning by discovering and maintaining a reusable Skill Handbook that models fine-grained agent competencies and costs, achieving superior accuracy while cutting training costs by up to 700x and eliminating routing collapse.

SkillOS: Learning Skill Curation for Self-Evolving Agents

Addressing the one-off problem-solving limitation of agents in streaming tasks, SkillOS decouples the system into a frozen executor and a trainable skill curator optimized via GRPO on skill-dependent task streams, enabling autonomous, OS-like management of Markdown skill libraries.

SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks

Presents SkillsBench, the first paired-evaluation benchmark treating Agent Skills as first-class evaluation artifacts across 87 tasks and 8 domains, demonstrating that expert-curated Skills improve agent resolution rates by +16.6 percentage points on average across 18 model–harness configurations.

SpecForge: Recursive Self-Improvement for Coding Agents via Behavioral Specifications

SpecForge introduces a recursive self-improvement framework for frozen coding agents that iteratively optimizes natural-language behavioral specifications via multi-frontier tree search—integrating Pareto state selection, dependency-constrained entity targeting, and error-diversified expansion—to overcome output coupling and substantially boost downstream code reconstruction and transfer.

SquidAgent: Parallelize Wisely, Coordinate Efficiently

To resolve the paradox where parallel multi-agent systems often run slower than serial baselines due to re-exploration and alignment costs, SquidAgent establishes a layer-wise parallelization criterion based on output-token budgets, combined with session forking and upfront convention planning to achieve superior throughput and consistency.

StraTA: Incentivizing Agentic Reinforcement Learning with Strategic Trajectory Abstraction

To resolve the exploration paralysis and credit assignment bottlenecks of reactive agents in long-horizon tasks, StraTA explicitly samples a compact global strategy upfront to guide all subsequent actions, training both strategy formulation and action execution jointly via hierarchical GRPO enhanced with semantic farthest point sampling and critical self-judgment.

Synthesis Through Simulation: Generating Coherent Enterprise Data via Scalable Agent-System Interaction

Presents Synthesis Through Simulation (STS), a schema-free enterprise data generation paradigm, and the Generalist Populator (GP) agent, which interactively populates databases via policy-enforcing APIs to achieve structural validity by construction with 100% constraint satisfaction across 10 diverse enterprise environments.

Tatemae: Detecting Alignment Faking via Tool Selection in LLMs

This paper introduces the Tatemae framework and a 108-scenario enterprise IT dataset, formalizing alignment faking (AF) as a verifiable composite behavioral event grounded in observable tool selection rather than brittle chain-of-thought traces, revealing mean AF rates between 3.5% and 23.7% across six frontier LLMs.

TerminalWorld: Benchmarking Agents on Real-World Terminal Tasks

TerminalWorld introduces a scalable data engine that automatically reverse-engineers 80,870 raw asciinema terminal recordings into 1,530 verified terminal tasks (with a 200-task human-vetted Verified core subset), revealing that state-of-the-art LLMs plateau at a 62.5% pass rate and suffer an efficiency paradox where failed trials consume 3.3× more tokens.

Test-time Multi-agent Coordination by Decomposed Value Gradient Flow

SCOUT decouples generative flow-matching behavioral priors from value learning in offline MARL, refining action particles at test time via decentralized Stein variational gradient descent to scale coordination performance without policy parameterization or training-time mode collapse.

The Best-Laid SCHEMEs: Coordinated Sabotage and Monitoring in Multi-Agent Systems

This paper introduces SCHEME, the first repository-level benchmark for coordinated multi-agent sabotage and monitoring, revealing that frontier coding agents can achieve high sabotage success rates (84% on Gemini 3.1 Pro) and demonstrating that trusted monitors inspecting code diffs and communication logs can reliably detect such coordination.

The Complexity Kink: A Prompt-Side Structural Complexity Index for Code-Generation Reliability

Addressing the "Reverse Threshold Problem" where measuring cyclomatic complexity on generated code is distorted by short failed stubs, this work introduces a prompt-side six-dimension structural rubric and uses Hansen's threshold regression across 21 LLMs to locate model-specific reliability breakpoints.

The Compliance Trap: Diagnosing How AI Agents Consume Conflicting Memory

Introduces the Entry–Propagation–Recovery (E-P-R) trajectory diagnostic framework and MemTrapBench to analyze agent memory consumption, uncovering a "compliance trap" where agents across scales consistently follow task-conflicting memory, causing severe baseline collapse in more capable models.

The Illusion of Multi-Agent Advantage

Through rigorous benchmarking across traditional reasoning and dynamic agentic benchmarks comparing six automated multi-agent frameworks against budget-matched Chain-of-Thought with Self-Consistency (CoT-SC), alongside a diagnostic synthetic benchmark (SMFR) and an Expert-MAS upper bound, this paper uncovers widespread "architectural bloat" and "functional collapse" in current automated multi-agent systems despite incurring up to 10× higher inference costs.

The PokéAgent Challenge: Competitive and Long-Context Learning at Scale

The PokéAgent Challenge introduces a dual-track benchmark spanning competitive battling and RPG speedrunning, backed by 22M+ battle trajectories and an open-source multi-agent orchestration harness, revealing substantial capability gaps between generalist LLMs and specialist RL/search algorithms while showing near-orthogonality to 49 standard LLM benchmarks.

TIER: Trajectory-Invariant Execution Rewards for Multi-Step Tool Composition

Addressing the severe failure of outcome-based rewards and trajectory-supervised supervision in multi-step tool composition, TIER establishes a trajectory-invariant tiered reward framework grounded in function schema validation and sandbox runtime execution feedback, enabling an 8B model to consistently surpass 90% accuracy across 1 to 6 composition steps.

TimeWarp: Evaluating Web Agents by Revisiting the Past

To resolve the vulnerability of web agents to evolving web interfaces, this paper introduces TIMEWARP—a benchmark featuring 1,386 tasks across 6 distinct UI design eras over 25 years—and proposes TIMETRAJ (plan-distilled multi-version trajectory collection) alongside TIMEWARP-BC (full-response behavior cloning), nearly doubling open-source agent success rates.

To Call or Not to Call: Diagnosing Intrinsic Over-Calling Bias in LLM Agents

Addressing the widespread tendency of LLM agents to over-call external tools, this paper leverages Sparse Autoencoders to extract behavior-aligned gating feature bases, identifies an activation-independent call offset formalizing the Intrinsic Bias Hypothesis (IBH), and devises Adaptive Margin-Calibrated Steering (AMCS) to causally mitigate over-calling via closed-form activation shifts without hurting valid calls.

To trust or not to trust: Attention-based Trust Management for LLM Multi-Agent Systems

Addressing the vulnerability where LLM multi-agent systems indiscriminately trust incoming messages, this paper establishes a six-dimensional trustworthiness taxonomy grounded in human communication theory, proposes a lightweight attention-based evaluator (A-Trust), and introduces a two-tier Trust Management System (TMS) enabling real-time message filtering and dynamic agent tracking.

Tools as Continuous Flow for Evolving Agentic Reasoning

FlowAgent reconceptualizes discrete tool chaining as continuous trajectory generation within a semantic space via conditional flow matching, enabling single-solve future plan generation paired with receding-horizon closed-loop execution to suppress error accumulation and empower unseen tool generalization.

ToolSearcher: Optimizing Tool Selection at Scale via Reinforcement Learning

ToolSearcher trains a multi-turn tool selector using a category-constrained curriculum, event-level advantages for first discoveries of target tools, and progress-dependent credit assignment, improving Qwen2.5-7B-Instruct's overall StableToolBench F1 from GDPO's 0.496 to 0.513 and improving AppWorld task outcomes with a separate execution agent.

Toward Interactive Understanding of Code APIs

This paper introduces the PAU benchmark to assess how large language models discover and reverse engineer black-box Python code APIs through interaction alone, reveals that even frontier LLMs suffer from overconfidence and premature termination, and adapts the Asymmetric Actor-Critic (AAC) paradigm with privileged information into post-training to significantly boost active exploratory capabilities.

Towards High Semantic Fidelity: Hyperdimensional Symbolic Messages in Multi-Agent Communication

To tackle semantic distortion and black-box uninterpretability in multi-agent communication over noisy channels, this paper proposes Hyperdimensional Symbolic Messages (HSM) and the Herm framework, employing entity-attribute-value structured encoding and receiver-centric dual-branch fusion to simultaneously achieve high transmission fidelity and conversion fidelity.

Tracing Agentic Failure from the Flow of Success

To tackle the long-horizon diagnostic bottleneck in LLM-based agentic systems, this paper introduces Oat, an unsupervised failure attribution method based on Neural Controlled Differential Equations (Neural CDEs) that trains exclusively on successful trajectories to characterize normal latent dynamics, localizing failure-contributing steps at inference time in milliseconds without frontier LLM queries.

TrajWiki: Source-Grounded Memory Trajectories for Long-Horizon Dialogue Agents

To overcome state-overwriting, fragmented retrieval, and untraceable evidence in long-horizon dialogue agents, TrajWiki represents memory via immutable episodic snapshots with claim-level edit operations and introduces a persistent Memory Wiki layer for hierarchical, evidence-grounded answer generation.

TrajWiki: Source-Grounded Memory Trajectories for Long-Horizon Dialogue Agents

TrajWiki addresses memory fragmentation and non-traceable overwriting in long-horizon conversational agents by representing dialogue history as source-grounded memory evolution trajectories with claim-level update operations, and compiling them into an interlinked Memory Wiki for hierarchical routing and provenance-verified response generation.

TwinRouterBench: Fast Static and Live Dynamic Evaluation for Realistic Agentic LLM Routing

TwinRouterBench introduces the first step-level routing benchmark designed specifically for multi-turn agentic workloads, using complementary static and live dynamic tracks to reduce API cost by 53.1% on SWE-bench without sacrificing task resolution.

UMAS: System-Level Uncertainty Quantification for Multi-Agent LLM Systems

Addressing the sharp degradation of reliability when single-agent uncertainty quantification (UQ) is applied to multi-agent systems (MAS), UMAS introduces a parameter-free, inference cost-neutral framework that tracks node confidences across the interaction graph, dynamically calibrates them via peer credibility, and achieves an average gain of 10.54 AUROC points across diverse benchmarks.

UMAS: System-Level Uncertainty Quantification for Multi-Agent LLM Systems

Addressing the structural misalignment and prohibitive sampling overhead of applying single-agent uncertainty quantification (UQ) to multi-agent systems (MAS), UMAS introduces a parameter-free, cost-neutral, system-level UQ framework that models node-level intrinsic confidence, modulates peer influence via relative credibility updating, and aggregates over the interaction trajectory graph, outperforming prior baselines by an average of 10.54 AUROC points.

Understanding the Challenges in Iterative Generative Optimization with LLMs

Systematically analyzes the fragility of LLM-based iterative generative optimization across code, workflows, and prompts, formalizing how three hidden design choices—starting artifacts, credit horizon, and experience batching—determine optimization outcomes, and provides practical guidelines.

UniClawBench: A Universal Benchmark for Proactive Agents on Real-World Tasks

Addressing the limitations of existing agent benchmarks that rely on static sandbox environments, lack multi-turn closed-loop interaction, and obscure failure causes behind application scenarios, UniClawBench introduces the first dynamic real-world benchmark across five foundational capabilities, leveraging a three-role information firewall and step-by-step completion checkpoints.

Unify-Agent: A Unified Multimodal Agent for World-Grounded Image Synthesis

Addressing the failure of existing unified multimodal models on long-tail concepts due to frozen parametric memory, Unify-Agent reformulates image generation into an end-to-end interleaved decision pipeline (Think, Research, Recaption, Generate), leveraging the synergy between VAE perceptual latents and ViT semantic representations to achieve high-fidelity, world-grounded image synthesis.

Untrusted Content Masking for Web Agents with Security Guarantees

This paper proposes Untrusted Content Masking (UCM), an architectural defense that redacts untrusted webpage content with labeled placeholders before rendering to web agents and routes necessary content queries through an isolated, type-constrained Quarantined Model (Q-Model), achieving 0% attack success rate against prompt injection while preserving full agent utility.

Useful Memories Become Faulty When Continuously Updated by LLMs

This paper systematically demonstrates that LLM-based agentic memory consolidation suffers from severe non-monotonic utility degradation—even when distilling from ground-truth trajectories—and shows that preserving raw episodic traces alongside explicitly gated consolidation outperforms forced iterative rewrites.

VerifyMAS: Hypothesis Verification for Failure Attribution in LLM Multi-Agent Systems

To overcome the limitations of joint prediction and agent-first attribution in capturing cross-step and cross-agent global failures, VerifyMAS introduces an error-first hypothesis verification framework that decouples failure attribution into trajectory-level three-way error verification and fine-grained responsible agent localization, complemented by specialized verifier fine-tuning.

VeriGraph: Towards Verifiable Data-Analytic Agents

To tackle the lack of auditability in LLM-based data-analytic agents, VeriGraph reframes linear reasoning trajectories into explicit heterogeneous evidence directed acyclic graphs (DAGs), unifying deterministic execution and semantic deduction via three expansion primitives and graph-aware composite RL, achieving state-of-the-art overall performance and an 87.61% Grounding Rate.

VeriTrip: A Verifiable Benchmark for Travel Planning Agents over Unstructured Web Corpora

VeriTrip establishes a verifiable benchmark for travel planning agents over an unstructured multimodal web sandbox (MRB) synchronized with a hidden gold-standard knowledge base (VKB), introducing cell-wise factual verification and TSP-based spatial coherence metrics that expose a fundamental cognitive load competition between active web retrieval and global preference fulfillment across frontier MLLMs.

VisionCreator-R1: A Reflection-Enhanced Native Visual-Generation Agentic Model

Addressing error accumulation in long-horizon multi-image workflows caused by lack of visual self-correction, this paper introduces VisionCreator-R1, a native visual generation agent trained via a novel Reflection–Plan Co-Optimization (RPCO) paradigm that resolves structural variance asymmetry, outperforming Gemini 2.5 Pro across benchmarks.

When and Why Does Multi-Agent Debate Fail and Does It Really Underperform?

This paper analyzes competitive and consensus-seeking multi-agent debate paradigms through a game-theoretic lens, reveals that incentive misalignments trigger debate hacking and severe degradation below single-agent baselines, and proposes ColMAD—a collaborative non-zero-sum debate protocol that consistently outperforms single-agent scaling under matched compute.

When Debate Helps: Proposal Supply and Verification-Aware Readout in Multi-Agent Reasoning

The paper reveals that multi-agent debate outperforms majority voting by coupling complementary proposal supply with verification-aware readout, introducing coverage-guided agent society construction and the Latent Verification Debate model.

When Do Multi-Agent Systems Help? An Information Bottleneck Perspective

This paper formulates multi-agent systems (MAS) and single-agent systems (SAS) under an information bottleneck perspective, proving their equivalence under infinite bandwidth and revealing that MAS gains hinge on the trade-off between reducing upstream redundant context and suffering task-relevant information loss from bounded relays.

XBridge: Entity-Grounded Latent Bridge for Heterogeneous LLM Communication

XBridge introduces the first decode-free dual-channel communication protocol for heterogeneous LLM multi-agent systems, aligning discrete entity identities via deterministic lexical anchor mapping while querying latent contextual enrichment through gated cross-attention, outperforming natural language communication by 14.3–21.4 pp across three model families and seven benchmarks with 11× lower latency.

XBridge: Entity-Grounded Latent Bridge for Heterogeneous LLM Communication

XBridge transfers the full context token sequence through deterministic lexical anchor mapping and reads sender hidden states through a pairwise-trained cross-attention bridge, outperforming 128-token text-summary communication on all seven tasks for three heterogeneous model pairs and reducing per-sample latency from 1.70 to 0.15 seconds in the specified H200 test.

DICE: Decoupling Capability from Intervention Necessity in LLM Tutoring

Addressing the core flaws of over-intervention and false acceptance in LLM tutors, this paper introduces the DICE framework to explicitly decouple "whether to intervene" from "how to generate responses," utilizing counterfactual Intervention Value (IV) weighted policy optimization and cost-sensitive decision rules to drive over-intervention to near zero while substantially cutting conversational turns.