Skip to content

🦾 LLM Agent

💬 EMNLP2026 · 143 paper notes

📌 Same area in other venues: 🧠 NeurIPS2026 (271) · 🎞️ ECCV2026 (36) · 📷 CVPR2026 (45) · 🔬 ICLR2026 (209) · 💬 ACL2026 (121) · 🧪 ICML2026 (83)

🔥 Top topics: Agents ×45 · LLM ×33 · Reasoning ×12 · Dialogue ×8 · Reinforcement Learning ×6

A Comprehensive Survey on Linguistic Steganography in the LLM Era

This paper delivers the first comprehensive survey systematically situating linguistic steganography in the LLM era, categorizing 148 steganographic schemes, 60 steganalysis countermeasures, 23 evaluation metrics, and 9 open challenges while delineating five decisive paradigm shifts.

ABDATABENCH: Evaluating AI Agents for Evidence-Grounded Scientific Data Analysis

Introduces ABDATABENCH, the first benchmark assessing AI agents on constructing complete, evidence-grounded antibody records across scientific literature and patents (44 documents, 279 records, 22-field tiered schema with Hungarian alignment and omission penalties), alongside ABCURATOR, a record-centric multi-agent framework achieving 78.3 (proprietary) and 76.0 (open-source) through deterministic tool integration and iterative skill accumulation.

Agent-as-Peer-Debriefer: A Multi-Agent Framework with Perspective-Based Refinement for Qualitative Analysis

This paper proposes Agent-as-Peer-Debriefer, a multi-agent framework that translates the qualitative research credibility practice of peer debriefing into LLM-assisted qualitative coding. By pairing a Hierarchical Coding Agent generating self-reflection memos with three disinterested Peer-Debriefing Agents applying Theory-Driven, Data-Driven, and Applied perspectives across five structural operations, the framework substantially improves code alignment and recall against human expert codebooks.

Agent-G^2: Gaussian Guidance for Agentic Reinforcement Learning

To resolve the structural mismatch of scalar hint depths in long-horizon agent RL with sparse rewards, Agent-G^2 reveals that informative guidance depths form a Gaussian band around the optimum and introduces an online Gaussian schedule that dynamically updates depth distributions using regular policy rollouts with zero probe overhead.

AgentFold: Closed-Loop Agentic Search for Protein Folding Model Design

AgentFold establishes an autonomous closed-loop multi-agent framework that searches over executable code variants on a compact ESMFold substrate via an MCTS-style tree controller, boosting best lDDT by 7.5% over independent Codex proposals under matched evaluation budgets and uncovering three transferable empirical design patterns for geometric deep learning.

AgentJudgeBench: A Multi-Difficulty Benchmark for Evaluating LLM Judges on Agentic Tool-Calling

Addressing the uncalibrated reliance on LLM judges in agentic evaluation, AgentJudgeBench provides a comprehensive benchmark across 6 DAG topologies, 3 difficulty tiers, and 3,808 records, revealing that judge alignment degrades monotonically with task difficulty, plateaus at a 77–82% structural ceiling without ground truth, and suffers from counterproductive over-anchoring in frontier models when exposed to reference traces.

AgentProv: Auditing Agentic LLM API Providers via Tool-use Policy Probes

AgentProv is the first action-based black-box identity auditing framework for agentic LLM APIs, utilizing categorical tool-selection preferences under functionally equivalent probes and a block-decomposed MMD permutation test to catch backbone substitutions without text dependence and with strong robustness against hidden system prompt injections.

AgentSpec: Speculative Decoding for Batch Inference of LLM Agents

Addressing the severe throughput degradation of speculative decoding in large-batch LLM agent serving caused by high token rejection rates and dynamic budget under-utilization, AgentSpec introduces pushdown automaton-guided structure-isolated drafting and redundancy-aware token budget allocation, achieving up to 2.02× end-to-end goodput speedup in vLLM without compromising generation quality.

Aggregate While Searching: Online Partial-Trajectory Aggregation for Long-Horizon Reasoning

Addressing the actionability gap where parallel search rollouts fail to correct reasoning drift before termination, this paper introduces OPTA to periodically synthesize a typed atomic global shared state and selectively inject relevant peer corrections, raising Qwen3-30B accuracy on BrowseComp-Plus from 37.9% to 59.6% with just 2.3% coordination overhead.

AGORA: An Archive-Grounded Benchmark for Agentic Workplace Document Reasoning

AGORA establishes an archive-grounded, cross-domain benchmark for agentic workplace document reasoning containing 9,664 authentic documents (372M tokens) across eight professional domains and 362 deterministic numeric queries, demonstrating that frontier LLM agents struggle to navigate messy corporate archives with even the strongest model reaching only 59.39% accuracy.

ANCHORS: Adaptive Normalization for Continual High-Order Retention via Sigmoid-Guided Magnitude Control

Addressing catastrophic forgetting in parameter-efficient fine-tuning (PEFT) caused by unregulated updates on critical parameters, ANCHORS draws inspiration from biological homeostatic plasticity, coupling CABR inverse low-rank decomposition with S-MagNorm Sigmoid-guided dynamic magnitude damping to boost downstream reasoning and QA accuracy while preserving over 61% of base knowledge without historical replay.

APEx: Distillation of Agent Procedural Experience for Adaptive Deep Research Question Answering

APEx establishes a hierarchical experience framework that decouples instance trajectories from category-level procedural skills, optimizing an Executor-Distiller-Planner loop via three-stage alternating GRPO and skill-guided test-time RL to beat closed-source frontiers on deep research benchmarks.

Are Android GUI Agents Robust Against Runtime Anomalies? AnTrap: Evaluating Agents in Dynamic Adversarial Environments

ANTRAP introduces a dynamic adversarial evaluation benchmark covering four layers and ten fine-grained anomaly subcategories (STAR) in Android environments, revealing universal vulnerabilities across 16 leading GUI agents and demonstrating that adversarial GRPO effectively remediates single-step state and action perturbations but fails to resolve multi-step contextual deadlocks.

AuroraCounsel: A Multi-Agent Large Language Model Collaboration Framework for Psychological Counseling

This paper proposes AuroraCounsel, a multi-agent counseling framework that decouples therapeutic workflows into sentiment sensing, structured cognitive reflection, and guided response generation, using emotion-driven dynamic routing to trigger reasoning on demand and GRPO to internalize clinical counseling rhythms across multi-turn sessions.

Beneath the Diff: Diagnosing and Mitigating Algorithmic Mode Collapse in Code-Level Autonomous Research Loops

This paper diagnoses the phenomenon of algorithmic mode collapse in code-level autonomous research loops (ARLs)—where surface edit diversity remains stable while underlying mechanism diversity drops by 50–70% and in-loop gains overstate true generalization by 2.0–2.6×—and introduces Diversity-Aware Proposal Sampling (DAPS) to curb semantic cluster decay by 69.1% and boost blind relative faithfulness by 83.7%.

Beyond Next-Observation Prediction: Agent-Authored World Modeling for Sequential Decision Making

Addressing the fundamental misalignment between passive next-observation prediction and action selection, Agent-Authored World Modeling (AAWM) elicits the policy's own beliefs and open questions before acting, retrieves cross-trajectory evidence, and synthesizes decision-oriented dynamics targets that significantly outperform conventional world modeling across embodied and web decision benchmarks.

Beyond Outcome Gaps: Process-Aware Fairness Diagnosis for LLM-based Multi-Agent Decision Systems

SCOPED-Hiring establishes a comprehensive process-aware fairness diagnosis pipeline across 311K structured multi-agent decision trajectories, revealing that balanced final hire rates frequently mask severe procedural disparities across intermediate reasoning, and demonstrating that diagnosis-guided Fair Skills reduce layered fairness burden by 72.3% while preserving hiring policy stability.

Beyond State Consistency: Behavior Consistency in Text-Based World Models

Tackling the "metric inversion" failure of surface-level similarity metrics in text-based world models, this paper introduces Behavior Consistency Reward (BehR) optimized via GRPO to align the action likelihood induced by simulated states with real-world environments.

CaptchaMind: Training CAPTCHA Solvers via Reinforcement Learning with Explicit Reasoning Supervision

To resolve the chronic failure of multimodal agents on CAPTCHAs caused by reliance on holistic visual impressions and lack of intermediate reasoning supervision, this paper introduces CaptchaBench (16k programmatically generated instances with process annotations) and CaptchaMind, an RL-based solver that enforces explicit intermediate region localization via a bounding box tool and dense process-level rewards under GRPO, reaching 82.9% average accuracy across eight tasks and transferring effectively to live commercial instances.

CAS: Conformalized Agentic Search via Adaptive Retrieval and Policy Weighting

Addressing the dual reliability crisis of retrieval noise and reinforcement learning overconfidence in search agents, CAS introduces Conformal Prediction into agentic search by employing Adaptive Prediction Sets (APS) for dynamic, coverage-guaranteed document truncation and Adaptive Conformal Inference (ACI) to penalize low-confidence trajectories during GRPO, substantially boosting multi-hop reasoning accuracy while curbing redundant tool calls.

CAVEWOMAN: How Large Language Models Behave Under Extreme Context Degradation

CAVEWOMAN introduces a two-channel evaluation protocol spanning nested grammatical reduction levels, revealing that compressing prompts increases net inference costs by up to 96% while degrading accuracy, whereas compressing model outputs slashes dollar costs by up to 3× without losing accuracy, although over 50% of correct responses become semantically disconnected from unconstrained reasoning.

COLLABMEM: A Diagnostic Benchmark for Collaborative Memory Operations in Multi-Agent LLMs

This paper introduces COLLABMEM, the first diagnostic benchmark for collaborative memory operations in role-governed multi-agent LLM systems, alongside TMS-GRAPH++, a scope-aware graph memory framework that combines role-filtered visibility, writer provenance, and temporal [CURRENT] tags to achieve top-tier collaborative retrieval while strictly preventing unauthorized leakage.

COLLABMEM: A Diagnostic Benchmark for Collaborative Memory Operations in Multi-Agent LLM Systems

Addressing the conflation of memory failures with planning and tool-use errors in multi-agent LLM systems, COLLABMEM establishes an auditable, memory-only diagnostic benchmark across six transactive memory competencies and proposes TMS-GRAPH++, a scope-aware graph memory architecture enforcing role-filtered views and temporal current-value resolution.

ContextPilot: Teaching Agents for Proactive Context Management via Fine-grained RL

ContextPilot addresses monotonic context growth and coarse exploration in agentic tasks by augmenting the context management toolset with planning, long-term memory, and soft offloading, while introducing an RL framework with context-aware partial rollout and snapshot-level credit assignment.

Cooperate to Compete: Strategic Coordination in Multi-Agent Conquest

The paper introduces Cooperate to Compete (C2C), a long-horizon mixed-motive game benchmark combining diagonal secret objectives, fog-of-war, and observable troop support; extensive human-AI and AI-AI experiments show that frontier LLMs over-compromise and display unprompted deception, while behavioral prompt interventions targeting aggressive asks and deception boost win rates from 22.2% to 32.7%.

CoPRM: A Collaborative Process Reward Model for Guiding Search Agent Reasoning

To resolve the critical bottleneck of assessment unreliability in process reward models for search agents, CoPRM introduces a collaborative calibration methodology that synthesizes immediate quality assessments with outcome-informed evaluations through consensus filtering for same-type actions and outcome-informed gating for cross-type decisions, delivering plug-and-play test-time guidance across diverse agent paradigms.

CoreMem: Riemannian Retrieval and Fisher-Guided Distillation for Long-Term Memory in Dialogue Agents

CoreMem introduces an information-geometric edge-cloud memory architecture for dialogue agents that eliminates the retrieval hubness problem via locally adaptive Fisher-Rao Riemannian metrics with Woodbury acceleration, while achieving theoretically bounded context compression via Fisher information trace sensitivity and structural syntax protection under strict 8 GB edge VRAM budgets.

CulturePRM: A Process Reward Model for Mitigating Cultural Bias in Language Models

Addressing the failure mode of "cultural overriding"—where reasoning models initially recognize relevant local cultural norms but subsequently weaken or replace them with generalized normative frames—this paper introduces CulturePRM, a 1B-parameter process reward model that verifies step-level norm preservation to rerank candidate reasoning traces via Best-of-N selection, delivering up to +12.06 pp accuracy gains while keeping base reasoning models frozen.

DeepSearch-World: Self-Distillation for Deep Search Agents in a Verifiable Environment

DeepSearch-World introduces a deterministic and verifiable offline Wikipedia environment featuring 420K multi-hop QA tasks, employing a scaffold teacher rollout and scaffold-to-ReAct distillation to enable a 9B open-source agent to autonomously self-evolve into a competitive deep search agent without relying on proprietary teacher traces.

DeltaMem: Towards Agentic Memory Management via Reinforcement Learning

DeltaMem replaces fragile multi-agent memory pipelines with an end-to-end single-agent ReAct framework, leveraging human-inspired memory evolution synthesis and an optimal-transport memory Levenshtein distance reward to train long-term persona memory managers via GRPO.

DeskCraft: Benchmarking Desktop Agents on Professional Workflows and Human-in-the-Loop Collaboration

DeskCraft addresses the limitations of existing desktop GUI benchmarks that focus on atomic, upfront-specified tasks by establishing a 538-task benchmark across 11 professional creative applications, structured into an L1–L3 difficulty taxonomy and an executable human-in-the-loop interaction protocol, revealing that even frontier agents suffer severe performance cliffs on long-horizon delivery and proactive clarification.

Distributional Stabilization for Zero-Shot Relation Extraction via Structural Margin Contrast

Under a full-vocabulary diagnostic, joint-encoding zero-shot relation extraction suffers from unnormalized dot-product recall collapse and instance-level BCE confidence ambiguity; SCAR introduces a gradient-isolated Z-score cascade alongside an inference-removed interactive contrastive branch, boosting Macro F1 by up to +9.7 pp and slashing zero-recall types from 35 to 1.

Do GUI Agents Know When Not to Act? Enabling Conflict-Aware Termination for Multimodal GUI Agents

Addressing the severe execution-biased over-compliance in multimodal GUI agents under infeasible instructions, this paper introduces the CONFLICTGUI benchmark covering both instruction-internal and instruction-GUI context conflicts, and develops CONFLICTGUARD, a training-free inference-time framework that synergizes a feasibility verification protocol with conditional activation steering to dramatically improve conflict termination by over 35% while fully preserving feasible task accuracy.

Dockerless: Environment-Free Program Verifier for Coding Agents

Dockerless eliminates heavy per-repository Docker containers and test runners by generating targeted verification questions and deploying parallel read-only sub-agents to gather codebase evidence, matching execution-based SFT filtering and RL rewards on SWE-bench without executing tests.

DSAgentBench: Can Agents Automate End-to-End Data-Science Workflows in Real Computer Environments?

DSAgentBench introduces the first benchmark to evaluate autonomous agents executing end-to-end data-science workflows inside real operating system environments across 275 multi-stage tasks, demonstrating that open-source models score below 1% and the strongest closed-source model achieves only 56.70% task success.

DualStake: Dual-Path Confidence Calibration in Deep Research Agent

Addressing the severe overconfidence of Deep Research agents during multi-step retrieval-augmented reasoning, this paper uncovers that post-retrieval Evidence Confidence (E-Conf) carries stronger uncertainty signals than post-generation Answer Confidence (A-Conf) and largely shapes it, motivating DualStake—a dual-path RL calibration framework using margin-clipped stake rewards that drastically cuts calibration error without sacrificing task accuracy.

Dude: A Dual-Detection Multi-Agent System for Paper-Code Discrepancy Detection

Dude is the first dual-detection multi-agent system designed for automated paper-code discrepancy auditing, resolving the chronic blind spot of single-agent models toward paper omissions and mitigating the precision collapse of naive multi-agent setups through checklist-guided negotiation and two-stage salience filtering, attaining an 86.55% F1 score on SciCoQA with over 93% precision.

Dynamic Dual-Granularity Skill Bank for Agentic RL

To overcome the limitations of trajectory-only guidance and unbounded memory degradation in agentic reinforcement learning, D2Skill organizes reusable experience into macro task skills and micro step skills, leveraging paired on-policy rollouts for hindsight utility estimation, two-stage UCB retrieval, and capacity-aware pruning.

EgoArgus: Benchmarking VLMs as Situational Assistants for Modality-Grounded User Supports

EgoArgus introduces a dual-track benchmark comprising 6,978 real-video MCQA instances and 789 synthetic VISTA intervention episodes to evaluate vision-language models across context understanding, next-action prediction, and proactive intervention timing in five controlled video-dialogue relationship scenarios, uncovering that contemporary VLMs systematically collapse into text-dominant bias below random chance when verbal input contradicts visual reality.

EmoDap: Dual-Agent Prompt Retrieval Framework for Zero-Shot Fine-Grained Emotion Recognition

Addressing emotional interference and counterproductive retrieval in zero-shot fine-grained emotion recognition, EmoDap introduces prototype-guided candidate filtering and a dual-agent retrieval framework jointly optimized via continuous prototype-aware rewards under GRPO, outperforming commercial LLM baselines across four benchmarks.

Evidence Laundering: Deep Research Agents Mistake Correlated Sources for Independent Evidence

This paper formulates "evidence laundering"—where adversaries syndicate a single false claim across paraphrased, cross-linked web pages to mimic multi-source consensus—and demonstrates that leading deep research agents systematically fail to discount correlated sources, while introducing a post-retrieval structural provenance gate based on masked writing skeletons that collapses manufactured agreement without discarding genuine independent reporting.

finding where the buck stops an automated failure attribution based reflection f

Addressing the issue where existing multi-agent reflection frameworks force all agents to reflect and thereby contaminate the memory of regular-behaving agents, this paper proposes DoCtOR: an automated "diagnose-then-correct" framework that localizes the decisive error step and responsible agent via a process reward model, synthesizes a verified counterfactual action, and engages only the decisive error agent through a PPO-optimized reflector.

Follow the Latent Roadmap: Navigating Revocable Decisions in Long-Horizon Tool-Use Agents

Addressing the twin failure modes of error propagation and local error reinforcement in revocable parallel decoding for diffusion large language models (dLLMs), this paper proposes ASRD, a training-free embedding-space framework that maintains a temporal-consistency Anchor Tokens Cache (ATC) to inject entropy-modulated priors into [MASK] positions and apply orthogonal perturbation probes to pending candidates, attaining up to 7.2× inference acceleration alongside up to +6.4% accuracy gains.

from agent skills to executable environments frontier targeted training for tool

This paper proposes Grafting, a closed-loop synthesis pipeline that transforms workflow- and script-bearing Agent Skills into self-contained executable environments, and Fathoming, an iterative reinforcement learning procedure that probes and optimizes tool-use LLMs along their dynamic capability frontier, achieving substantial multi-turn tool-calling improvements on a 4B model.

From Atomic to Agentic: Towards Interpretable Evaluation of LLMs’ Agentic Mathematical Capabilities

This paper introduces AgenticMathBench (AMB), the first process-level benchmark that systematically aligns a taxonomy of reusable mathematical atomic capabilities with core agentic behaviors (Planning, Action, Feedback), demonstrating that models with comparable end-to-end accuracy often exhibit starkly disparate and fragmented intrinsic agentic profiles.

From Retrieval to Reconstruction: Constructing Evolvable Cognitive Memory for Long-Term Dialogue

Addressing the critical flaw of semantic collapse between subjective claims and factual records in flat dialogue RAG, CogMem introduces a training-free cognitive architecture governed by the PEC2F graph schema, offline CLS-inspired memory consolidation, and ReAct-driven deterministic graph operators that reconstruct evidence across long conversational sessions.

GRASP: Gated Regression-Aware Skill Proposer for Self-Improving LLM Agents

GRASP formulates LLM agent self-improvement as a sequence of validated edits to a bounded, reversible behavioral skill library, admitting candidate skills only when they achieve a positive net improvement on a balanced held-out probe under a strict zero-new-regression budget, boosting operational accuracy on clinical EHR benchmarks by up to 48.2 points without any parameter updates.

Grounding Memory Summarization in Utility Intent

Addressing the core issue where conversational memory summarization optimizes human-facing criteria at the expense of fine-grained facts, MemSuit introduces privileged query-answer conditioning into an adaptive teacher, distills utility-aware multi-entry extraction into a query-agnostic student, and contrastively calibrates the dense retriever.

GUI-PRA: Process Reward Agents for GUI Tasks

Addressing the evaluative knowledge gap and visual ambiguity inherent to static, single-pass scoring in standard Process Reward Models (PRMs), GUI-PRA decouples what to verify from how to verify via experience-injected criterion synthesis and criterion-guided autoregressive visual tool perception, turning verification into an active investigation that markedly boosts long-horizon GUI task success.

HALT: Evaluator-Guided Clause-Level Pruning and Early Halting for Lower-Cost Multi-Agent Systems

To tackle the dual dilemma of surging API costs and cascading error amplification in multi-agent reasoning, HALT introduces macro-level difficulty-aware query routing coupled with micro-level evaluator-guided clause pruning and early halting, achieving dramatic token savings while even boosting task accuracy.

HardGen: Generating Hard Samples for Tool-use Agents

Addressing the issue where current tool-use data generation adheres to shallow random sampling along simple happy paths, HardGen proposes a three-phase pipeline—failure-driven dependency graph sampling, reverse advanced tool abstraction for implicit hard query generation, and feedback-guided closed-loop CoT refinement—synthesizing 27k challenging trajectories that elevate a 4B model to 79.14% overall accuracy on BFCLv3, surpassing frontier proprietary models.

Holmes-SQL: Benchmarking LLM Agents on Evidence-Aware Investigative Reasoning over Structured Databases

Introduces Holmes-SQL, the first interactive benchmark for evidence-aware urban crime investigation over structured relational databases, revealing that despite achieving over 97% SQL execution accuracy, state-of-the-art LLM agents peak at only a 49% case solve rate due to unequal evidential strength and logical fallacy traps.

How Coding Agents Fail Their Users: A Large-Scale Analysis of Developer-Agent Misalignment in 20,574 Real-World Sessions

Based on 20,574 real-world coding-agent sessions across 1,639 repositories spanning IDE and CLI workflows, this paper introduces the first multi-axial empirical framework characterizing developer-agent misalignment visible through developer pushback, showing that 90.50% of breakdowns drain developer effort and trust rather than causing irreversible destruction, while 91.49% of visible resolutions depend strictly on human corrective intervention.

How Far Are AI Scientists from Changing the World? A Thorough Evaluation of Automated Discovery Systems

This paper comprehensively examines the evolution of AI Scientist systems from prompt-to-paper automation toward auditable closed-loop research laboratories, formulating a four-capability taxonomy across knowledge acquisition, hypothesis generation, experimental validation, and evolution, alongside an E0-E4 evidence hierarchy that pinpoints the fundamental gap separating fluent text generation from rigorously auditable scientific trajectories.

How Many Pixels Is a Digit Worth? Place-Aware Coordinate Entropy for GUI Agent Control

Addressing the lack of reliable confidence signals in vision-language GUI agents that emit discrete screen coordinates, this paper identifies "place-value asymmetry" in coordinate token sequences and introduces Place-Aware Coordinate Entropy (PACE)—a parameter-free, training-free confidence metric that weights the Shannon entropy of each coordinate digit by its decimal place value during a single greedy forward pass to substantially improve the ranking of correct versus wrong clicks.

HTAM: Hierarchical Transition-Attended Memory for Operator Optimization

To resolve the granularity mismatch in LLM-based GPU operator optimization—where coarse hints are reusable but non-executable, while fine-grained traces are actionable but inflate search space—HTAM introduces a Hierarchical Transition Graph (HTG) that guides coarse-to-fine kernel evolution via position-aware transition-attended scoring, reaching 98.4% correctness and a 1.978× geometric mean speedup on the full 250-task KernelBench suite.

HyLaT: Efficient Multi-Agent Communication via Hybrid Latent–Text Protocol

HyLaT introduces a dual-channel hybrid communication protocol for multi-agent systems, funneling verbose intermediate reasoning through continuous latent vectors to eliminate token overhead while keeping critical conclusions in natural text for interpretability and interoperability, achieving competitive collaborative performance at single-agent operational costs.

I Am Multitudes: A Survey and Roadmap for Rethinking Personas in Large Language Models

Following PRISMA guidelines across 149 ACL Anthology papers, this meta-analysis demonstrates that 84% of current NLP and LLM research flattens identity into static, legally protected sociodemographic proxies predominantly deployed for post-hoc bias auditing, and establishes a foundational roadmap spanning theoretical grounding, self-identification, relational dynamics, 5-lens evaluation, and caricature-free pluralistic alignment.

ICA: Information-Aware Credit Assignment for Visually Grounded Long-Horizon Information-Seeking Agents

ICA models long-horizon web information seeking as an evidence attribution problem, employing rendered visual webpage snapshots as stable, comparable evidence units across trajectories, estimating evidence utility via group rollout success rate discrepancies, and backpropagating dense rewards to retrieval actions that acquire high-value evidence under a decoupled turn-level GSPO framework.

InSight: A Benchmark for Agentic Claim Verification in Interactive Visualizations

InSight introduces an ecologically valid benchmark of 21,349 analytical claims grounded in interactive, multi-view Vega-Lite environments, requiring multimodal agents to actively uncover occluded visual evidence via pure pixel inputs and proposing the Interaction Efficiency Score (IES) to penalize passive perception and ungrounded confabulation.

Invisible to Humans, Triggered by Agents: Stealthy Jailbreak Attacks on Mobile Vision–Language Agents

This paper introduces agent-only perceptual injection, a stealthy low-privilege attack paradigm that exploits touch-contact discrepancies between humans and automated agents to conditionally expose screen-bounded jailbreak prompts, achieving up to 82.5% planning ASR and 75.0% execution ASR on commercial backends like GPT-4o with zero human visibility.

LADA: Enabling Adaptation of Black-Box LLMs to Dynamic Domain Changes at Test Time

To tackle catastrophic reasoning derailment in black-box LLMs under sudden domain shifts without internal parameter access or token probabilities, LADA meta-trains a step-level scoring adapter over a Cartesian product of diverse datasets and synthesized reasoning error patterns, enabling rapid few-shot calibration to new domains at test time and guiding the LLM toward correct reasoning trajectories via adaptive step-level resampling.

Last Step Matters: Early Uncertainty Cannot Predict Failure in Long-Horizon Agent Tasks

This paper systematically assesses uncertainty dynamics throughout long-horizon deep-research agent execution, demonstrating that frequent in-trajectory path switching decouples early uncertainty from final outcomes (mean AUROC < 0.60 at 50% progress), whereas final-step verbal confidence reliably predicts failure (mean AUROC 0.85), proving that post-completion retry Pareto-dominates in-trajectory intervention.

LatCom: Cross-Agent Latent Compression for Efficient Multi-Agent Collaboration

LatCom tackles the receiver-side context explosion and cross-agent representation redundancy in multi-agent latent collaboration by compressing multi-sender continuous thoughts into a fixed number of task-relevant slots via two-stage training, achieving a 2.46× inference speedup and 70.3% fewer output tokens while maintaining competitive task accuracy.

LatentEvolve: Self-Evolving Test-Time Scaling in Latent Space

Inspired by cognitive science's Complementary Learning Systems (CLS), LatentEvolve establishes an unsupervised self-evolving test-time scaling framework that alternates fast daytime episodic momentum transfer with slow nighttime parametric consolidation via a compact latent weaver, enabling frozen LLMs to learn how to scale computation across diverse tasks.

LatentMem: Customizing Latent Memory for Multi-Agent Systems

LatentMem addresses the fundamental bottlenecks of memory homogenization and long-context information overload in Multi-Agent Systems (MAS) by retrieving raw trajectories from a lightweight experience bank, composing role-aware continuous latent memories via a memory composer, and optimizing them end-to-end via Latent Memory Policy Optimization (LMPO).

LegalWorld: A Life-Cycle Interactive Environment for Legal Agents

Grounded in 75,309 paired first- and second-instance Chinese civil judgments, this paper introduces LEGALWORLD—the first five-stage life-cycle interactive environment spanning consultation to final appellate judgment—and LongJud-Bench, uncovering cross-stage error cascading and exposing dynamic courtroom advocacy as the primary frontier for legal agents.

LifeMem: Enabling Lifelong Experience Reuse for LLM Agents via Memory Distillation

LifeMem addresses action space mismatch and catastrophic forgetting in inter-environment lifelong learning by structuring heterogeneous interaction trajectories into workflow clusters constrained by directional statistics, distilling invariant high-level skills, and activating them alongside environment-aligned exemplars to achieve significant positive backward transfer across 10 benchmarks.

LLM Agents as Computational Typologists

AutoTypologist is the first LLM agent architecture designed for evidence-grounded computational linguistic typology, combining hierarchical grammar pre-retrieval, interlinear glossed text (IGT) quantitative analysis, dynamic evidence graph tracking, and post-hoc adversarial auditing to perform feature coding and cross-linguistic hypothesis testing across 25 diverse languages.

LLM Fine-tuning with an Unconstrained Fine-tuned Teacher

Proposes the Unconstrained Fine-tuned Teacher (UFT) framework, which reconceptualizes the teacher's role from passive knowledge transfer to active sample value discovery, balancing task adaptation and prior-capability retention via source-uncertainty-budgeted utility scoring and confidence-adaptive dual-teacher distillation.

LogicEnvGen: Task-Logic Driven Generation of Diverse Simulated Environments for Embodied AI

Addressing the limitation of existing embodied simulator generation methods that overemphasize visual realism while overlooking logical test diversity, LogicEnvGen introduces a top-down, LLM-driven environment generation framework that decomposes tasks into decision-tree behavior plans, extracts non-redundant logical trajectories via a heuristic minimal selection algorithm, and instantiates physically plausible 3D test cases using Z3 constraint solving.

LongDS-Bench: On the Failure of Long-Horizon Agentic Data Analysis

LongDS-Bench introduces the first long-horizon, multi-turn data analysis benchmark targeting evolving analytical states (updates, rollbacks, counterfactual perturbations, and multi-state compositions), revealing that top models achieve under 48.45% accuracy and over half of all failures stem from long-horizon state mismanagement and cascading error propagation.

MAS-on-the-Fly: In-Context Structural Adaptation of LLM-Based Multi-Agent Systems

MASFly introduces an in-context structural adaptation framework for LLM-based multi-agent systems that instantiates query-tailored agent topologies from an SOP repository and performs closed-loop execution-time supervision and dynamic reconfiguration with a global Watcher and personalized experience pool without updating model parameters.

MaS-VQA: A Mask-and-Select Framework for Knowledge-Grounded Visual Question Answering

Addressing the noise accumulation and cross-modal evidence decoupling in knowledge-grounded VQA, MaS-VQA introduces a unified Mask-and-Select framework that jointly prunes irrelevant visual regions and textual passages, using compact explicit evidence to constrain MLLM implicit knowledge generation without inference-time fine-tuning.

MemDefrag: Latent Memory Defragmentation for Large Language Models

Addressing the long-standing challenges of accumulated positional encoding misalignment and the lack of retrieval tracing in long-term latent memory for LLM agents, this paper discovers an intrinsic tracing signal in middle Transformer layers and introduces MemDefrag—a training-free framework that performs inference-time defragmentation (density-based reordering and Top-\(K\) filtering) coupled with informativeness-guided proportional forgetting, boosting 50-step knowledge retention on NaturalQA from 17.4% to 43.0%.

MemGuard: Persisting Verifier Signals for LLM-Agent Memory Governance

MemGuard shifts trajectory verification from a one-shot admission filter to persistent lifecycle metadata, attaching reward, confidence, label, and uncertainty descriptors to govern candidate admission, conflict resolution, quality-aware retrieval, and archival across continuous agent streams.

MEMORA: A Memory-Enhanced Multimodal Committee Reranking Agent for Reasoning-Intensive Retrieval

Addressing the counterintuitive phenomenon where adding images degrades dense retrieval performance on complex reasoning tasks, this paper introduces MEMORA, a memory-enhanced multimodal committee reranking agent that combines episodic memory, dynamic intent factor decomposition, bipartite evidence graphs, and adaptive control to establish new state-of-the-art results on MM-BRIGHT.

MemoryForge: Synthesize Lifelong Memory for Human-Like LLM Agents

Departing from static textual persona injection, MemoryForge formalizes customized lifelong autobiographical memory synthesis from concise persona prompts via socio-historical context grounding, backward developmental planning, and multi-resolution simulation, empowering frozen LLMs to perform situation-relevant retrieval for remarkably authentic human-like behavior.

MENTOR: Reinforcement Learning via Flexible Teacher Guidance for LLM Reasoning

Addressing the dilemma between uninformative outcome signals and overly restrictive trajectory replication in tool-use distillation, MENTOR introduces an on-policy reinforcement learning framework with flexible, set-based teacher alignment and strict execution verification, significantly improving out-of-domain reasoning and tool generalization for small language models.

MetaBench: Benchmarking Two-Stage Reranking and Top-N Screening for Meta-Analyses

MetaBench standardizes early-stage meta-analysis study selection into a two-stage pipeline of high-recall candidate reranking and budgeted Top-N screening, integrating 7,801 PubMed meta-analyses, 4.2M candidate records, coverage-based tiering, and PICOS-targeted synthetic hard negatives to evaluate LLM screening agents under realistic human review workloads.

MobileGym: A Verifiable and Highly Parallel Simulation Environment for Mobile GUI Agents

MobileGym is a browser-hosted, lightweight Android-like simulation platform that models app ecosystems into fully controllable structured JSON states, enabling millisecond resets, zero-consequence operations, and single-server scaling to hundreds of parallel instances for online RL, where GRPO policies retain 95.1% of simulation gains on physical smartphones.

MoCA-Agent: A Market-of-Claims Code Agent for Financial and Numerical Reasoning

MoCA-Agent tackles silent miscomputations and opaque aggregation in financial and tabular question answering by decomposing inputs into typed atomic claims, clearing specialist trader orders into an audit trail of accept/reject decisions, and synthesizing executable Python code bounded by operation-aware verification and conflict arbitration.

RedNote-Vibe: A Lifestyle Social Media Dataset for Detecting AI-Generated Text and Analyzing Its Impact

The authors introduce RedNote-Vibe, the first longitudinal Chinese lifestyle social media dataset spanning 2020 to 2025 across 148k posts and 17 LLMs, accompanied by the PsychoLinguistic AIGT Detection Framework (PLAD) that extracts 31 theory-grounded cognitive features to achieve superior generalization to unseen models and uncover how strategic human-AI collaboration drives peak engagement.

MultiSessionCollab: Learning User Preferences with Memory to Improve Long-Term Collaboration

This paper introduces MultiSessionCollab, a multi-session collaborative problem-solving benchmark and a user-centric long-term memory architecture that leverages user preference enforcement signals to train agents via GRPO, significantly enhancing multi-session task success rates while reducing user interaction effort.

Not All Evidence Is Equal: Source-Sensitive Calibration in Large Language Models

By holding propositional content constant while manipulating source credibility and grammatical/lexical evidential markers, this paper introduces a bilingual Turkish-English human-grounded benchmark for hearer-side uptake, demonstrating that humans exhibit robust source-sensitive calibration (+1.69 points) whereas LLMs show only weak, superficial sensitivity (+0.25 to +0.29 points) and lack deep epistemic grounding.

OffSeeker: Online Reinforcement Learning Is Not All You Need for Deep Research Agents

Addressing the excessive API costs and instability of online reinforcement learning (RL) in deep research agents, this paper introduces DeepForge—a lightweight end-to-end task synthesis pipeline—and OffSeeker-8B, an agent trained entirely offline using 33k SFT trajectories and 21k DPO pairs that matches 30B online RL systems while completely reducing training interaction costs to zero.

Omnilingual MT: Machine Translation for 1,600 Languages via Scalable Multilingual Pretraining

Omnilingual MT introduces a massively multilingual translation ecosystem covering over 1,600 languages by combining a 256K balanced tokenizer, backtranslation across 2,000+ languages, and the grammatically diverse seed dataset MeDLEy, demonstrating that specialized 1B–8B models (OMT-LLaMA and OMT-NLLB) strictly outperform 70B+ general-purpose LLMs and double the count of practically usable languages.

OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis

OpenResearcher decouples one-time online corpus bootstrapping from multi-turn trajectory synthesis and serves a controlled 15M-document offline search engine equipped with explicit search, open, and find primitives, generating 97K+ trajectories via GPT-OSS-120B that elevate a 30B student model to 54.8% on BrowseComp-Plus (a 34.0 percentage point gain over the base).

OpenSkill: Open-World Self-Evolution for LLM Agents

OpenSkill establishes the first target-task-supervision-free framework for open-world LLM agent self-evolution by decoupling external knowledge retrieval into implementation guidance and objective verification anchors, generating deterministic virtual unit test suites to iteratively diagnose and refine modular skills that achieve state-of-the-art automated pass rates and transfer seamlessly across distinct model families.

OptiMAS: Automatically Optimize Multi-Agent System

Addressing evolutionary instability and the isolation of insights across lineages in discrete search-based multi-agent evolution, OptiMAS introduces an end-to-end continuous optimization paradigm over a unified ReAct substrate, leveraging textual gradients, dual-track memory, and adaptive sampling to achieve steady, compounding performance gains across four challenging agentic benchmarks.

P-Search: A Parallel Plan–Act–Verify Framework with Decoupled GRPO for Deep Research

P-Search introduces a five-stage deep research framework—sequentially executing grounded background modeling, dependency-free evidence branching, parallel subagent search, and iterative verification closure—trained via behavior-level decoupled GRPO on star-shaped subgraphs to break single-path anchoring and cascade planning errors.

Peak-Then-Collapse and the Four Interface Channels of Knowledge Graph Tool Use

This paper systematically assesses the standard reinforcement learning with verifiable rewards (RLVR via GRPO) recipe on a stripped knowledge-graph tool interface, identifying a signature "peak-then-collapse" failure mode across reward designs, formalizing a four-channel signal-degradation diagnostic framework, and establishing a stable teacher-free self-distillation recipe achieving 40.0% Exact Match.

PlanBench-XL: Evaluating Long-Horizon Planning of LLM Tool-Use Agents in Large-Scale Tool Ecosystems

PlanBench-XL introduces an interactive benchmark comprising 1,665 tools and 327 multi-step retail tasks to rigorously evaluate LLM agents under partial tool visibility and retrieval-time explicit, implicit, and misleading failure blocks.

PredNav: A Predictive Cognitive Map for Zero-Shot Object Goal Navigation

Inspired by the predictive cognitive map of the hippocampal-entorhinal system, PredNav formulates zero-shot Object Goal Navigation frontier exploration as room-level structured prediction over an externalized topological map, integrating visual evidence, offline priors, and multi-view VLM verification to set new state-of-the-art results on MP3D and HM3D.

PRISM: Entropy-Guided Multi-View Exploration for Scientific Idea Generation

PRISM addresses the conceptual homogeneity of LLM-generated scientific ideas by estimating pairwise view orthogonality through directional hypothesis-transition Shannon entropy, selecting maximally orthogonal view subsets to condition backbone generators, and achieving superior novelty and impact across 500 scientific problems spanning 23 domains.

Privacy-Preserving Text Sanitization for Distributed Agents Collaboration via Disentangled Representations

To tackle both explicit identifier leakage and implicit stylometric signature exposure during cross-agent textual evidence exchange, DiSan introduces an orthogonal role-style disentangled encoder coupled with federated spherical prototype alignment, cutting answer-level PII exposure by 20× while maintaining 83.17% answer faithfulness in distributed multi-agent RAG.

PsyAgent: Constructing Human-like Agents Based on Psychological Modeling and Contextual Interaction

PsyAgent introduces a schema-first framework that couples a Big Five Individual Structure (IS) prior with Multi-Scenario Contexting (MSC) frames across eight arenas, utilizing parameter-efficient SFT and DPO to deliver stable trait fidelity and norm-aware contextual behavior in compact language models.

Q-CAT: A Non-Parametric Query Filtering System for Multi-Session Dialogue

Q-CAT addresses sensitive information leakage and time-varying corporate governance policies in enterprise LLM interactions by replacing retraining-heavy parametric classifiers with an embedding-based non-parametric retrieval system over a synthetic query pool, slashing inference latency to 36 ms while enabling zero-retraining policy adaptation.

Read the Paper, Write the Code: Agentic Reproduction of Social-Science Results

Develops an information-isolated agentic reproduction framework that extracts structured methods descriptions and blinded table templates from empirical papers, re-implements Python analysis code from scratch using only raw data without seeing original code, and achieves over 90% coefficient sign recovery on 48 human-verified social science papers while isolating paper underspecification as the leading barrier to automated reproducibility.

Reasoning and Tool-use Compete in Agentic RL:From Quantifying Interference to Disentangled Tuning

Addressing the fundamental bottleneck where reasoning and tool-use compete under shared-parameter agentic reinforcement learning (ARL), this paper introduces Capability Effect Attribution (CEA) to prove their near-orthogonal gradient conflict, and proposes DART—which freezes the pretrained backbone and deterministically hard-routes tokens to disjoint LoRA adapters—approaching the 2-Agent upper bound across 13 benchmarks.

Retrieval-Augmented LLM Agents: Learning to Learn from Experience

This paper introduces EXPRAG (an experience retrieval-augmented generation framework) and EXPRAG-LoRA (a retrieval-aware fine-tuning paradigm), showing that integrating episodic trajectory retrieval directly into LoRA fine-tuning substantially boosts generalization on unseen tasks and enables agents to learn from their own failures in-context without test-time parameter updates.

Risky Business: Measuring The Faithfulness-Safety Tension

The authors uncover a core alignment trade-off between chain-of-thought monitorability (faithfulness) and safety robustness, introducing the HazMart autonomous shopkeeper benchmark and Targeted Reasoning Replacement to show how models can be steered along asymmetric internal resistance and compliance pathways without degrading core capabilities.

RoboSPA: Can VLA Models Go Beyond Simple Scenes and Handle Spatial Reasoning Complexities?

Addressing the blind spots of existing VLA benchmarks confined to simple scenes and short-horizon instructions, RoboSPA establishes a 527K-trajectory benchmark across 5 difficulty levels, pairing Object-Normalized Target Accuracy (ONTA) with Progress Score (PS) to reveal dramatic performance collapses in fine-grained spatial reasoning and long-horizon execution.

SciLENS: RL-Driven Autonomous Agents for Scientific Localized Evidence Navigation and Synthesis

SciLENS introduces a fully localized autonomous research agent operating on a dual-tier 12-million-record scholarly store, pioneering the use of structured data-driven visualization as an actionable tool to overcome context bottlenecks and aligning the policy via citation subgraph consensus synthesis and reverse-decomposition rubric reinforcement learning.

SecAgent: Efficient Mobile GUI Agent with Semantic Context

SecAgent tackles non-English data scarcity and visual history redundancy in mobile GUI automation by constructing CMGUI—a human-verified Chinese mobile dataset with 121k navigation steps and multi-choice action benchmark CMGUI-Bench—and proposing an autoregressive semantic context mechanism that distills multi-step visual trajectories into compact natural language summaries, enabling a 3B-scale model to outperform same-size baselines and match 7B–8B models.

Seeing is Believing? Evaluating Vision-Language Model Susceptibility in Agent-to-Agent Multimodal Persuasion

Addressing the security threat where autonomous vision-language agents are manipulated by adversarial peers, this paper introduces the MMPERSUADE benchmark and Persuasion Discounted Cumulative Gain (PDCG) metric to decouple semantic from visual influences via function-calling probabilities, discovering an "Alignment Paradox" where more capable models resist benign persuasion but become paradoxically more vulnerable under adversarial multimodal inputs.

Self-Corrected Speech Reasoning: Diagnosing and Improving Audio LLMs via Internal Verification

Addressing the "final-condition adoption failure" where speech LLMs transcribe or mention self-corrections yet fail to use them as active premises in reasoning, this paper establishes the 13,295-instance SCUR-Bench benchmark with a three-path diagnostic protocol, and proposes PACER—a GRPO framework combining structured condition rewards with speech-evidence contrast that reduces end-to-end and condition-related errors by 39.8% and 67.5%.

SHIFT: Self-reconstruction Harnesses Implicit Fine-grained Thinking for Retrieval

Addressing the dual representation and supervision mismatches between generative language modeling and dense retrieval, SHIFT generates continuous soft thinking tokens via vocabulary expectations, decouples contextual generation noise via residual projection and bidirectional attention pooling, and supervises latent reasoning through fine-grained next-token self-reconstruction, substantially outperforming prior rewrite-then-retrieve pipelines and implicit reasoning retrievers on reasoning-intensive benchmarks.

SIRI: Self-Internalizing Reinforcement Learning with Intrinsic Skills for LLM Agent Training

SIRI introduces a three-phase self-internalizing reinforcement learning curriculum that bootstraps policy interaction, mines compact skills from self-generated successful trajectories with online paired-rollout validation, and distills high-value skill-guided actions into policy parameters via hierarchical RL advantage weighting, achieving state-of-the-art agent performance with zero inference-time retrieval overhead.

Size Sound Symbolism Survives Genealogy Across 27 Language Families

By assembling a curated cross-linguistic dataset of 810 audio-validated size adjectives across 27 languages spanning 13 families, this work combines distance-stratified leave-one-language-out transfer with adversarial debiasing to demonstrate that phonetic size cues (/i/ for smallness and /o, a/ for largeness) robustly survive the suppression of genealogical and source-language shortcuts.

SkillBrew: Multi-Objective Curation of Skill Banks for LLM Agents

Formalizes skill bank curation for retrieval-augmented LLM agents as a constrained multi-objective Pareto optimization problem over utility, diversity, and coverage, and introduces a bi-level propose-then-verify framework driven by counterfactual leave-one-out replay to continuously evolve compact, high-performing skill repositories.

SkillMe: Evolving Agent Skills for Recurring Personal Workflows

Addressing the situated mismatch of generic skills under recurring personal workflows, SkillMe establishes a workflow-situated skill evolution framework where a high-capability Publisher induces task-scoped skills as temporary workspace files, heterogeneous Executors provide trace-grounded feedback, and evidence attribution drives closed-loop revision, yielding a +25.7 points mean gain and significantly lowered execution volatility.

SkillPyramid: A Hierarchical Skill Consolidation Framework for Self-Evolving Agents

SkillPyramid introduces a hierarchical skill consolidation and self-evolution framework that reorganizes flat skill libraries into a cohesive pyramidal topology featuring reusable low-level atomic operations and high-level abstract problem-solving patterns, enabling agents to compose, validate, and incrementally absorb new skills during execution to achieve a 38.0% average reward gain and a 27.7% step reduction across ALFWorld, WebShop, and ScienceWorld.

SkillsVote: Lifecycle Governance of Agent Skills from Collection, Recommendation to Evolution

SkillsVote establishes a full-lifecycle governance framework for agent skills across open-source profiling, pre-task agentic library search recommendation, post-task subtask attribution, and evidence-gated evolution, resolving context distraction and library pollution to achieve robust test-time online self-evolution and offline experience transfer on Terminal-Bench 2.0 and SWE-Bench Pro.

SocialCritic: Grounding Multi-Agent Dialogue via Advantage-Based Endorsement Suppression

Addressing the operational risk where unsupported hallucinations in multi-agent dialogue cross control boundaries via cross-speaker endorsement, this paper presents SocialCritic—a learned receding-horizon controller that uses decoupled utility, norm, and epistemic sequence advantage critics with span-level targeted repair to suppress error propagation and boost actionable task accuracy.

SocialCritic: Grounding Multi-Agent Dialogue via Advantage-Based Endorsement Suppression

SocialCritic introduces a receding-horizon dialogue controller powered by decoupled utility, norm, and epistemic advantage critics with targeted span-level repair, suppressing cross-speaker endorsement of unsupported claims by 77.5% while lifting actionable accuracy.

ST-EVO: Towards Generative Spatio-Temporal Evolution of Multi-Agent Communication Topologies

Proposes ST-EVO, the first spatio-temporal evolving multi-agent system that utilizes a compact flow-matching generative scheduler to dynamically evolve spatial communication topologies across dialogue iterations, achieving an average accuracy improvement of 14.20% across nine benchmarks while drastically cutting token consumption.

STARK: Strategic Advantage Reinforcement with Confidence-Driven Exploration for Long-Horizon Agentic Learning

Addressing cognitive-mode collapse and single-axis credit assignment in agentic RL, STARK dynamically branches at low-confidence decision keypoints with injected reflection prompts and introduces a dual-track strategic and tactical advantage estimator, allowing a compact 1.5B model to outperform GPT-5 on out-of-domain benchmarks.

Stephanie2: Thinking, Waiting, and Making Decisions Like Humans in Step-by-Step AI Social Chat

Stephanie2 reformulates step-by-step interactive chatting into a step-wise decision-making dialogue agent featuring active waiting and adaptive pacing, explicitly choosing between sending a short message or waiting to listen, and modeling latency as thinking time plus typing time to achieve human-like conversational rhythm.

Streaming Communication in Multi-Agent Reasoning

STREAMMA replaces the prevailing "generate-then-transfer" multi-agent protocol with step-level streaming communication, exploiting the position-dependent reliability of reasoning chains to pipeline execution for up to 26.9× speedup while preventing error-prone late steps from misleading downstream agents, boosting reasoning accuracy by an average of 7.3 percentage points across diverse benchmarks.

SWE-MEM: Learning Adaptive Memory Management for Long-Horizon Coding Agents

SWE-MEM provides long-horizon coding agents with an autonomous, on-demand memory management tool, trained through curriculum fine-tuning and Memory-aware GRPO with trajectory checkpoint splitting and step-level credit assignment, achieving 43.4% and 60.2% resolve rates on SWE-Bench Verified under a compact 32K context limit while markedly lowering overall token consumption.

Syntactic Complexity Convergence in Dialogue: Analyzing Human-LLM Alignment Dynamics

By quantifying dependency-based syntactic complexity across four structural dimensions, this work reveals a fundamental dichotomy between within-metric convergence in human dyads and cross-metric compensatory alignment in human-LLM dyads (demonstrated via statistically significant negative Linear-CKA slopes), leveraging these structural dynamics to build an LLM-involved dialogue detector that attains 96.0% accuracy without language-model-based likelihood estimation.

T-Mem: Memory That Anticipates, Not Archives

T-Mem identifies that existing conversational memory systems fail under associative recall due to their reliance on surface-similarity matching; by establishing a 2x2 design space across granularity and orientation and precomputing write-time anticipatory triggers for all four quadrants, it collapses the cross-benchmark drop between factual and associative memory from 28–50 pp down to 5.45 pp.

TaxoAgent: Utility-Aligned Policy Optimization for Hierarchical Semantic Classification

TaxoAgent reformulates hierarchical semantic classification as an explicit coarse-to-fine decision trajectory, aligning an LLM policy with a composite non-decomposable utility spanning correctness, structural validity, prototype semantic proximity, and trajectory coherence via Group-Relative Policy Optimization (GRPO) to curb cross-branch errors and hierarchy violations.

Teaching by Learning: Harnessing LLMs Collaboration for Adaptive Data Synthesis via Cognitive Load Feedback

Addressing the issues of coarse outcome feedback and sequence-level averaging dilution in multi-model collaborative data synthesis, this paper proposes CoFAG, a framework that quantifies internal cognitive load via a dilution-resistant Inference Stress Index (ISI) based on token bifurcation entropy and iteratively optimizes instruction synthesis through a three-agent closed loop with Direct Preference Optimization (DPO).

The Autonomy Tax: Defense Training Breaks LLM Agents

This paper exposes a fundamental capability-alignment paradox termed the "autonomy tax," where defense fine-tuning against prompt injection overfits surface shortcuts, destroying multi-step agent execution through immediate Step-1 tool incompetence, unrecoverable retry cascade timeouts (up to 99%), and high bypass rates under sophisticated attacks.

The Selection Effect: Why Compact LLM Verifiers for Emotion Classification Fail Behind Encoder Confidence Gates

By formalizing inference-time verification as cost-aware selective routing across closed-label emotion benchmarks, this paper discovers a fundamental selection effect: confidence gates forward an ambiguous and novel slice where compact LLM verifiers fail to beat the supervised encoder itself, rendering pure encoder abstention and split-conformal gating strictly superior on the Pareto frontier.

TIDE: Trajectory-based Diagnostic Evaluation of Test-Time Improvement in LLM Agents

Addressing the failure of static outcome metrics to capture multi-turn interactive dynamics, this paper introduces TIDE, a lightweight and agent-agnostic diagnostic framework that decomposes Test-Time Improvement (TTI) into optimization efficiency (AUV), behavior adaptation vs. loop stagnation (LR), and working memory utility (MI).

Time is Not a Label: Continuous Phase Rotation for Temporal Knowledge Graphs and Agentic Memory

Addressing the static-dynamic dilemma where discrete timestamp metadata either buries permanent facts under recency sorting or requires expensive destructive updates, ROMEM introduces a drop-in temporal reasoning module that models time as continuous phase rotation in complex vector space modulated by a semantic speed gate, achieving non-destructive geometric conflict resolution in append-only agentic memory.

TokenCliff: Benchmarking Output Token Budget Sensitivity in Multi-Turn LLM Agents

TokenCliff introduces a controlled benchmark of 660 deterministic tasks across four core capability domains and a multi-step split, sweeping episode-level output token budgets from 200 to 8,192 tokens across 10 LLMs (33,000 episodes) to reveal performance cliffs of up to 32 percentage points and demonstrate that peak task performance and model scale do not guarantee budget elasticity.

ToolForge: A Data Synthesis Pipeline for Multi-Hop Tool Interaction in LLM Agents

ToolForge introduces an efficient data synthesis pipeline that converts QA seed data into verified multi-hop, multi-turn, and self-reflective tool-use trajectories without executing large-scale real-world tool APIs. By combining 19 base virtual tools with dual-gated variants, four interaction paradigms, paired golden and noisy retrieval simulations, and a two-layer validation module (MLV), fine-tuning Qwen3-8B on just 4,250 synthetic samples achieves a 21.6% relative gain over the Tool-Integrated Reasoning (TIR) baseline across five out-of-domain QA benchmarks and outperforms models trained on 25k samples on BFCL.

ToolWeave: Structured Synthesis of Complex Multi-Turn Tool-Calling Dialogues

ToolWeave establishes a modular, verifiable synthetic data generation framework that constructs curriculum-driven tool graphs with built-in dataflows, extracts motif-aligned workflows, and tracks parameter provenance across fine-grained plans, driving the share of true multi-step tool turns from under 7% to over 40% and cutting parameter hallucinations from over 50% to under 20%.

TraceProbe: An Execution-Based Benchmark for Diagnosing Safety Failures in LLM Agents

TraceProbe establishes an execution-based diagnostic benchmark of 155 executable sandbox scenarios categorizing risk sources into non-adversarial boundary overreach, internal authority compromise, and external attacks, leveraging normalized traces, structured evidence packets, and model-free recorded action replay to achieve step-level localization of latent trajectory violations and actionable guard placement.

Truth or Tribe: How In-group Favoritism Prioritizes Facts in Persona Agents

The paper introduces the "Truth or Tribe" triadic simulation framework adapted from social psychology's minimal group paradigm, demonstrating that persona LLM agents exhibit pervasive in-group favoritism by adopting incorrect claims from identity-similar peers while discounting ground-truth facts from dissimilar peers, and proposes three effective prompt-based mitigation strategies.

TSHMem: A Temporal-Semantic Hierarchical Memory Graph for Long-Term Conversational Agents

TSHMem organizes long-term dialogue turns, self-contained statements, hierarchical summaries, and entities into a heterogeneous memory graph, integrating query-conditioned temporal constraints directly into graph transition probabilities and bidirectional hierarchical ranking to achieve state-of-the-art long-term memory retrieval using fewer evidence tokens.

Validity Is Not Visibility: A Lifecycle FSM for Agent Memory

Addressing the structural conflict between current-intent and historical-intent queries in agent long-term memory, this paper proposes MEMFSM, a training-free retrieval-side framework that decouples world-state validity from query visibility via a four-state finite-state machine (FSM) and an intent-aware router, achieving state-of-the-art results across heterogeneous benchmarks including the newly curated VALIDMEM.

WeaveBench: A Long-Horizon, Real-World Benchmark for Computer-Use Agents with Hybrid Interfaces

Addressing the limitation of existing computer-use agent benchmarks that evaluate visual GUI or command-line channels in isolation, WeaveBench establishes 114 real-world, long-horizon tasks across 8 domains requiring non-substitutable cross-interface orchestration on deployed agent runtimes, accompanied by a trajectory-aware judge that detects shortcut cheating to reveal that frontier backbones attain at most 41.2% PassRate.

When Denser Credit Is Not Enough: Evidence-Calibrated Policy Optimization for Long-Horizon LLM Agent Training

To eliminate the divergent anchor bias in long-horizon LLM agent RL where rare lucky actions receive inflated advantages under small-sample rollouts, this paper proposes Evidence-Calibrated Policy Optimization (ECPO), combining action-level shrinkage estimation with variance-gated credit weighting to achieve robust policy improvements with negligible computation overhead.

Why Act Without Asking? Evaluating Mobile GUI Agent Clarification Strategies

Addressing the pervasive "Act-without-Asking" pathology where mobile GUI agents almost never query users under ambiguous goals (clarifying in only 0.1% of cases), this paper introduces the step-level MID-Bench benchmark with a strict user simulator and shows that train-free logit cascades boost ambiguous-task success from 0% to 40%–56% without degrading clear-instruction accuracy.

World Models for Policy Refinement in StarCraft II

StarWM introduces the first player-view, action-conditioned textual world model for StarCraft II and integrates it into a lightweight Generate–Simulate–Refine decision loop (StarWM-Agent), achieving consistent win-rate gains (+30%, +15%, and +30% against Hard, Harder, and VeryHard built-in AI) and dramatically reducing supply block rates through foresight-driven macro refinement.

WorldLines: Benchmarking and Modeling Long-Horizon Stateful Embodied Agents

Introduces WorldLines, the first benchmark evaluating long-horizon stateful embodied household assistance under partial observability, and proposes ObsMem, an observer-grounded memory framework that separates event evidence, structured state trails, epistemic beliefs, and commitments to prevent state-overwriting hallucinations and enable executable embodied planning.

WRIT: Write-Read Intensive Trajectory Synthesis for Multi-Turn User-Facing Agents

WRIT factorizes the complexity of multi-turn user-facing agent trajectories into two orthogonal axes—the number of write decisions and the read evidence burden per decision—and demonstrates that fine-tuning a compact 4B model on just 2K synthesized trajectories outperforms GPT-5.1 no-think on \(\tau^2\)-bench while dramatically slashing inference token usage.

You Only Align Once: Propagating Cooperative Behaviors in Multi-Agent Systems through Seed Agents

This paper introduces Alignment Propagation, showing that distilling principled cooperative reasoning into a single 14B open-weight seed agent via SFT more than doubles group cooperation in unmodified collectives purely through natural-language dialogue, while generalizing zero-shot across environments and model families.