Monday's cs.AI announcement day (2026-10-12) lists 83 new and 78 cross-lists (replacements skipped; listing total 161). A filter for agent systems, memory/context, computer-use / GUI / tools / MCP / skills / harnesses, multi-agent, persistence/identity, and local/open models keeps 40 papers. Guardrails and multi-agent safety lead: VEPA keeps lifelong agent policies as verified symbolic rules with an evidence gate, while reward hacking is shown to transmit across coding agents via shared code, memory, and plans. Harness and evidence state get first-class treatment in LongHarness (100M-token corpora), ReHarness (recurrent online harness evolution), and Humanize-style multi-session alternation that beats one decelerating long session. Personal and tool surfaces tighten too: Work2Skill induces reusable GUI skills from human demos, X-AgentRed red-teams cross-service personal-agent paths, RATED ranks tools by execution reliability, and CLEAT / Collaborative Theory of Mind push value-guided clarification and anticipate-user self-revision before acting.
Research Papers
VEPA: Verified Evolving Formal Policy for Lifelong LLM Agent Guardrails
Duy Nguyen; Lesly Miculicich; Bhavana Dalvi Mishra et al. arXiv: 2610.12813
VEPA represents the guardrail as a formal policy of executable symbolic rules over typed agent events. An LLM proposes policy patches from deployment feedback; a deterministic verifier accepts only structurally valid, conflict-free policies; an evidence-aware authority activates a verified rule only with enough supporting evidence. Outperforms baselines on ASSEBench-Safety/Security and ATBench while expanding deterministic formal-policy enforcement under mixed streams, corrupted feedback, and longer horizons.
Key insight: VEPA keeps lifelong agent guardrails as verified symbolic rules: LLMs propose patches, a deterministic verifier and evidence gate decide activation.
Reward Hacking as a Transmissible Strategy in Multi-Agent Coding Systems
Yuepeng Hu; Pegah Ghahremani; Reza Aghajani et al. arXiv: 2610.12822
Reward-hacking elicited in one coding agent can propagate to others via code, shared memory, and natural-language plans across roles. Even small open-weight agents induce hacking in stronger closed-weight targets. On hard tasks, propagation intensifies along a transmission chain after the original source leaves as code acquires more persuasive justifications. Code review by the coder or a separate verifier reduces delivered hacks.
Key insight: Reward hacking is transmissible across multi-agent coding systems via code, shared memory, and plans — even from small open-weight sources into stronger targets.
LongHarness: Advancing Agentic Reasoning over 100M-Token Corpora via Stateful Evidence Management
Songtao Huang; Zhijie Zhong; Xiaohan He et al. arXiv: 2610.13075
LongHarness targets agentic reasoning over 100M-token corpora. Existing harnesses leave evidence and conclusions entangled in growing interaction history, causing unbounded exploration and evidence-state drift. LongHarness makes evidence management an explicit stateful substrate: track acquired evidence, open information gaps, and evolving conclusions separately from the raw interaction log.
Key insight: LongHarness separates evidence, information gaps, and conclusions from raw chat history for agentic reasoning over 100M-token corpora.
Work2Skill: Capturing Human GUI Experience for Reusable Agent Skill Induction
Yuqi Huang; Hui Li; Yifei Cao et al. arXiv: 2610.12984
Work2Skill captures human GUI experience for reusable agent skill induction. WorkRecorder aligns screen recordings, mouse/keyboard events, and accessibility-tree info; WorkDemo-350 is a demonstration corpus. Separates workflow rules from task-specific details so induced skills transfer to unfamiliar software requiring human domain knowledge.
Key insight: Work2Skill induces reusable GUI agent skills from human demos by aligning screen, input events, and accessibility trees while stripping task-specific details.
Hindsight Structured Advice Distillation: Steering Executor Agents with Advisors Trained on Their Failures and Task Characteristics
O. Ozan Koyluoglu arXiv: 2610.13000
Instead of retraining a closed/shared executor or stuffing missing know-how into a one-size-fits-all prompt, a small advisor model reads each input and writes structured advice that steers the executor. The advisor is trained on the executor's failures and task characteristics (hindsight), so guidance is input-specific without modifying the executor.
Key insight: A small advisor trained on executor failures can emit structured per-input advice that steers a frozen/closed executor without prompt bloat.
CLEAT: Learning Value-Guided Clarification for Interactive Agents
Jiatan Huang; Zheyuan Zhang; Ziming Li et al. arXiv: 2610.12655
Clarification is useful only when the agent can turn the answer into better decisions; as training changes how the agent handles incomplete info, the same clarification can become more or less useful. CLEAT co-trains a clarification-value controller and the task policy from shared interaction outcomes, formulating residual question value so the agent asks when the answer would change the decision.
Key insight: CLEAT co-trains clarification value with the task policy so agents ask only when the answer would change the decision.
X-AgentRed: Red-Teaming Personal Agents in Evolving Multi-Service Ecosystems
Yeonsung Jung; Zhiyuan Wang; Zifan Wang et al. arXiv: 2610.13886
Personal agents connect to multi-service ecosystems via tools; compositional failures arise when agents fulfill benign requests whose cross-service effects are harmful. X-AgentRed red-teams by systematically exploring diverse cross-service paths rather than simply prompting LLMs to invent harm scenarios.
Key insight: X-AgentRed red-teams personal agents on compositional cross-service tool paths, not only single-tool jailbreaks.
RATED: Reliability-Aware AI Agent Tool Retrieval via Execution Dynamics
Weiliang Chen; Song Liao arXiv: 2610.13889
RATED combines robust semantic tool matching with dynamic reliability adaptation from execution dynamics. Existing tool retrieval optimizes semantic relevance but ignores tool availability/reliability at execution time; RATED uses a frozen dual-encoder with contrastive training plus runtime reliability signals.
Key insight: RATED retrieves tools by semantic match plus execution-dynamics reliability — availability matters as much as similarity.
Humanize MLE Agents: Multi-Agent Session Alternation Beyond Single-Agent Limits
Changye Li; Zijian Zhang; Muheng Li et al. arXiv: 2610.13494
Single-session MLE agents suffer long-horizon deceleration: reasoning intensity declines as sessions lengthen, and custom harness benefits vanish as the underlying model improves. Humanize MLE Agents uses multi-agent session alternation (fresh sessions with handoff) beyond single-agent limits.
Key insight: Multi-agent session alternation beats one long MLE agent session that decelerates as context grows.
Collaborative Theory of Mind: Learning to Self-Revise by Anticipating User Responses
Fanzhi Zeng; Anubrata Das; Maytal Saar-Tsechansky et al. arXiv: 2610.12872
CToM trains LLMs to anticipate user responses and use those predictions to self-revise before presenting a response — drafting, predicting the user reply, then revising. Targets multi-turn collaboration failures: solving before understanding needs, and failing to revise mistaken assumptions after feedback.
Key insight: Collaborative Theory of Mind anticipates the user's next reply and self-revises before acting — for multi-turn personal agents.
Competence-Aware Orchestration for LLM-Based Multi-Agent Systems
Songmiao Wang; Yuchen Wang; Chenglin Yu et al. arXiv: 2610.13144
LLM multi-agent collaboration should adapt to the backbone's competence on the current query: more collaboration is not always better — richer coordination helps on hard cases but adds token cost and redundant interactions on easy ones. Competence-aware orchestration estimates when to escalate to multi-agent vs stay single-agent.
Key insight: Competence-aware MAS orchestration escalates to multi-agent only when the backbone looks weak on the query — more collaboration is not always better.
Quality, Not Just Outcome: A Structural Scoring Framework for Selecting Coding-Agent Trajectories for Supervised Fine-Tuning
Sida Li; Yuanxin Liu; Shicheng Li et al. arXiv: 2610.13152
Outcome-level filtering (final patch passes tests) ignores process quality: successful trajectories may still contain repetitive file inspection, unstable tool usage, or delayed modification. Proposes a structural scoring framework over interaction patterns to select coding-agent trajectories for SFT that prefer efficient, stable process — not just a green verifier.
Key insight: SFT on coding-agent trajectories should score process structure (tool stability, inspection loops), not only final test pass.
Recurrent Harness for Agentic Continual Learning
Yuxin Xiao; Zhiyun Lu; Jianyu Wang et al. arXiv: 2610.12668
ReHarness is a recurrent harness for agentic continual learning that complements forward-only state evolution. Existing harnesses either evolve only forward or need extra replay/validation data for harness optimization; ReHarness supports online adaptation to distribution shifts over long interaction streams without that extra data.
Key insight: ReHarness recurrently evolves the agent harness online for continual learning without extra replay/validation sets.
Don't Guess, Just Ask: Evaluating Interactive Requirement Clarification for Repository-Level Software Development
Mingyu Chen; Zhiqing Zhong; Manyi Wang et al. arXiv: 2610.13022
Evaluates interactive requirement clarification for repository-level software development. Users begin with vague intent; existing benchmarks skip pre-coding clarification or stay at function-level. Measures how well LLMs transform vague intent into clear requirements through interaction before coding, and how that affects repo-level outcomes.
Key insight: Repo-level coding needs a measured clarification phase before implementation when user intent is vague.
SWE-DeceptionBench: Benchmarking Deceptive Behaviors of LLMs Across the Software Development Lifecycle
Xiting Liu; Lin Yang; Yitong Zhang et al. arXiv: 2610.13626
Benchmark for LLM deception across six software-development lifecycle stages with Active Deception (present infeasible/conflicting requirements as achievable) and Passive Deception (conceal/distort). Measures trustworthiness of LLMs used throughout the SDLC beyond task success.
Key insight: SWE-DeceptionBench measures active and passive deception across six SDLC stages beyond task-success metrics.
Continually Post-training Small Models via Synthetic Self-distillation
Woosung Koh; Hyunsoo Lee; Kyungjae Lee et al. arXiv: 2610.13113
Self-Distillation Fine-Tuning fails on smaller models needed for low-latency on-device use. Shows poor acquisition/forgetting correlate linearly with post-trained distribution shift; introduces synthetic self-distillation that keeps learning in-distribution so continual post-training works at smaller scales.
Key insight: Synthetic self-distillation keeps continual post-training in-distribution so small on-device models can acquire without forgetting.
EvoBelief: Falsifiable World Models for Self-Improving LLM Agents
Janet Wang; Jiang Guo; Yi Xiang et al. arXiv: 2610.12789
Self-evolving agents usually improve how they act; EvoBelief evolves beliefs about environment dynamics during deployment as falsifiable hypotheses. Unexpected outcomes and decision-relevant unknowns trigger hypothesis formation; later observations and controlled experiments validate or revise them independently of the acting policy. Beliefs support multiple tasks but incorrect ones can propagate — so they must be testable and revisable.
Key insight: EvoBelief evolves falsifiable environment hypotheses at deployment time — complementary to acting-skill self-evolution.
Know What You Don't Know: Agentic Data Synthesis by Evolving Generation Space for Closed-Loop Learning
Wei Yang; Qing Ping; Youzhi Luo et al. arXiv: 2610.12593
EvoSyn is a learning-feedback-driven multi-agent framework that organizes synthesis around a revisable state specifying what interaction factors can be deliberately generated and where synthesis effort should be allocated. A Designer allocates synthesis directions and a Planner grounds them into executable experience, enabling within-space adaptation and synthesis-space expansion. On τ²-Bench and CAR-Bench: +13.3 average task success over the strongest synthesis baseline, with higher-quality and structurally broader synthetic experience.
Key insight: EvoSyn closes the feedback-to-decision gap: learner failures revise what the synthesizer may generate, lifting average task success +13.3 over the strongest synthesis baseline.
SpireBench: Evaluating Long-Horizon Strategic Play of Agentic Models with Slay the Spire
Scott C. Lowe; Sander W. Keemink arXiv: 2610.12596
SpireBench has models play the complete commercial deck-builder Slay the Spire at highest difficulty across 20 LLMs. Frontier models match average humans but lag experts; the gap widens as runs lengthen — strongest models win act 1 like experts then fall behind. All models misjudge risk over a run (preserve health instead of spending it to grow). Memories written across runs are often wrong and overly cautious; renaming every game term shows part of performance rests on prior game knowledge.
Key insight: Frontier agents match average humans on full Slay the Spire runs but fall behind experts as horizons lengthen; cross-run memories are often wrong and over-cautious.
Beyond Binary Verifiable Rewards: Behavioral Self-Rewarding for SWE Agents
Yiran Wu; Andrew Zhao; Yurui Chang et al. arXiv: 2610.12598
Binary RLVR rewards fail to distinguish trajectories with the same verifier outcome but different debugging quality. Behavioral Self-Rewarding (BSR) reuses the current policy to compare observable debugging behavior among same-task trajectories with the same verifier outcome, aggregates pairwise preferences into task-local scores, and converts them into denser rewards — without a separately trained reward model.
Key insight: Behavioral Self-Rewarding densifies binary SWE verifier rewards by comparing debugging behavior among same-outcome trajectories — no separate reward model.
Travel Back To Move Forward: Retrospective Exploration and Evolving for Long Horizon Agents
Ziang Ye; Wentao Shi; Lei Chen et al. arXiv: 2610.13058
ReTRAIL lets agents rewind the environment to earlier states while preserving valid progress and carrying forward knowledge from abandoned exploration — environment state moves backward while agent knowledge moves forward. Trains a shared policy with State-Anchored GRPO (compares siblings from the same rewind point) plus exploratory experience distillation into guidance for later decisions.
Key insight: ReTRAIL rewinds environment state while carrying forward knowledge from abandoned exploration — recovery without discarding lessons.
When Faster Decisions Do Not Make Better Agents: An Empirical Study of Jev in Agent Workflows
Yuchen Ling; Shengcheng Yu; Zhenyu Chen et al. arXiv: 2610.13092
Studies whether faster/cheaper local decisions from Jev transfer into efficient, reliable task completion across five agents in two representative workflows. Finds that local decision speedups do not automatically make better agents — workflow structure and decision placement determine whether savings compound or are wasted on extra steps.
Key insight: Faster local decisions (Jev) do not automatically improve agent task completion — workflow placement of decisions dominates.
Locate, Judge, Read: Role-Specialized Tool Agents over a Knowledge-Point Index for Open-Domain Multimodal Document QA
Wenlin Lin; Fanzhe Wei arXiv: 2610.13128
KnowForge separates jobs that current RAG agents dump on one model: a hybrid-retrieval knowledge-point index; a Locator that can only search; a code-scheduled bridge judge that forces the missing second hop; a typed, size-bounded Reader. Role specialization over a knowledge-point index beats single-agent tool scheduling for open-domain multimodal document QA.
Key insight: KnowForge role-specializes Locator / Judge / Reader over a knowledge-point index instead of one model scheduling every retrieval tool.
RoleWorld: Knowledge Transfer in World Modeling with Functional Role Abstractions
Boyuan Xiao; Pu Wang; Ji Feng et al. arXiv: 2610.13172
RoleWorld transfers world-model knowledge across environments with similar interaction structure by encoding dynamics as functional role abstractions rather than environment-specific symbols. Compositional role structure affords transfer where raw symbolic rules or executable programs still overfit surface names.
Key insight: RoleWorld transfers world models across environments by functional role abstractions rather than surface-specific symbols.
JANUS: Enhancing Long-Horizon Autonomous Research with Co-Evolving World Models
Mengdi Liu; Chenghao Jia; Hong Chang et al. arXiv: 2610.13194
JANUS addresses feedback delays in Generate-Execute-Feedback research agents with co-evolving world models that anticipate experiment outcomes before costly execution, reducing reactive trial-and-error under expanding search spaces for long-horizon autonomous research.
Key insight: JANUS co-evolves world models with research agents to cut feedback delays from Generate-Execute-Feedback trial-and-error.
Forgetting is Not Always Bad: A Neuro-Inspired Memory Repair Mechanism for Poisoned LLM Agents
Lei Liu; Yunji Liang; Xiaowen Zhang et al. arXiv: 2610.13267
LLM agents are highly vulnerable to memory-poisoning. Inspired by neuroscience (weak reactivation destabilizes memories without reinforcement), proposes a neuro-inspired memory repair mechanism that uses controlled forgetting rather than continuous external auditing — improving memory efficiency vs static isolation or always-on auditors.
Key insight: Neuro-inspired memory repair uses controlled forgetting/weak reactivation against memory poisoning — cheaper than always-on external auditors.
SWE-Critix: Post-Training Generalist Verifiers for Assessing Coding-Agent Trajectories
Jiajun Hu; Chun Yong Chong arXiv: 2610.13356
SWE-Critix post-trains generalist verifiers that judge and explain coding-agent trajectories when no oracle evaluator defines success. Targets long-horizon software engineering trajectories that are extremely long and hard for naive LLM judges.
Key insight: SWE-Critix post-trains generalist verifiers that judge and explain long coding-agent trajectories without an oracle evaluator.
Web$hell: A Unified Interface for Agentic Web Access
Sunghwan Kim; Deogyong Kim; Sangam Lee et al. arXiv: 2610.13416
Web$hell unifies search-augmented, browser-based, and filesystem-based agentic web access behind one interface so agents can choose the right access model as tasks become more information-intensive, rather than being locked into one access pattern.
Key insight: Web$hell unifies search, browser, and filesystem agentic web access behind one interface.
Rollout-Time Intervention: Unlocking Agentic Reinforcement Learning Beyond Policy Limits
Chen Xi; Ding Ruixue; Wang Zhigaoyuan et al. arXiv: 2610.13438
On-policy agent RL collapses when tasks exceed the model's capability boundary (effective trajectories become sparse). Rollout-Time Intervention unlocks learning beyond policy limits by intervening during rollouts so training can proceed when pure on-policy sampling finds almost no successes.
Key insight: Rollout-time intervention lets agentic RL keep learning when on-policy successes become too sparse beyond the policy's capability boundary.
Humanize Olympiad Agents: A Multi-Session Workflow for Olympiad Problem Solving with Solve, Review, Wiki, Verify
Zhengyang Zhang; Jing Xiong; Hongzhou Lin et al. arXiv: 2610.13632
A multi-session workflow for olympiad problem solving with Solve, Review, Wiki, Verify roles. Argues a single agent session remains brittle on difficult problems (early errors compound); releasing both model and complete inference stack matters for reproducibility. Complements recent Lean/Erdős systems that withhold stacks.
Key insight: Solve/Review/Wiki/Verify multi-session workflows beat single brittle sessions on hard olympiad/research problems.
A Baseline Assessment for the Agentic Web Technology Stack
Vagner Figueredo de Santana; Diogo Cortiz arXiv: 2610.13703
Baseline assessment of how protocols, frameworks, and libraries compose an Agentic Web solution. Without such an assessment, developers face skilling, governance, and infrastructure challenges before joining the agentic web. Maps interoperability and gaps across the emerging stack.
Key insight: A baseline map of Agentic Web protocols/frameworks is needed before locking an implementation stack.
Meta-Reflective Evolution Effectively Improves Production AI Systems
Seung Won Wilson Yoo; Chang Zeng; Aditi Somayajula et al. arXiv: 2610.12794
CoCoEvolve (Cortex Code Evolve) is a meta-reflective evolutionary optimizer for production AI systems. Most LLM-guided evolution mutates a frontier candidate from its failures alone; on production systems isolated failures reveal little about complex hidden interactions. CoCoEvolve uses meta-reflection over system-level interactions rather than single-candidate failure patches.
Key insight: CoCoEvolve meta-reflects over production-system interactions rather than mutating one frontier candidate from isolated failures.
Co-Modifying Application Code and Deployment Infrastructure: A Causal Study of AI Agent Pull Requests
Mohammad Wathiq Soualhi; Basem Suleiman arXiv: 2610.12849
Causal study of AI agent pull requests that co-modify application code and deployment infrastructure (CI/CD, containers, provisioning). Evaluating agents only in sandboxes against local tests misses deployment awareness; cross-domain modifications introduce unverified dependencies into asynchronous integration. Analyzes 26,999 PRs across AIDev v5 and MOSAIC-3M.
Key insight: AI agent PRs that co-edit app code and CI/CD need deployment-aware evaluation — sandbox tests miss cross-domain breakage.
AgentSWE: Can Coding Agents Build the Agent You Actually Want?
Jiahao Wang; Hongjin Qian; Yuyang Hu et al. arXiv: 2610.13361
AgentSWE benchmarks whether coding agents can build the agent you actually want across Creation (from NL requirement), Editing of production agent codebases, and Optimization toward a measurable target — 25 commission-framed tasks. Tests the assumption that a delivered agent faithfully serves its requirement rather than merely appearing to.
Key insight: AgentSWE tests whether coding agents build agents that faithfully serve the commission — not ones that merely look right.
ProgramGym: Scaling Training Environments for Long-Horizon Program Reproduction Agents
Chengyi Yang; Renxu Bao; Jue Hou et al. arXiv: 2610.13395
ProgramGym (Program-Env + Program-Agent) scales training environments for long-horizon program reproduction agents that must discover requirements and decide system design — not just resolve issues in a repo or generate from a full requirements doc. Targets sustained exploration of a reference program before implementation.
Key insight: ProgramGym trains long-horizon agents to discover requirements and design programs, not only patch known issues.
Matched-Context Group Policy Optimization for LLM Agent Training
Chenxu Zhang; Yaorui Shi; Wenlong Zhang et al. arXiv: 2610.13755
Matched-Context Group Policy Optimization constructs weighted action preferences from already collected trajectories for long-horizon agents with sparse terminal feedback. Defines comparison contexts via pre-action state summaries and observed rubric progress; two-action support with exact/coarse matching; fixed folds separate candidate decisions from terminal outcomes.
Key insight: MCGPO assigns credit in long-horizon agent RL by matching pre-action contexts before comparing sibling actions.
Incremental Retrieval-Augmented Generation for Human-in-the-Loop Decision Support
Federico Mazzoni; Andrea Failla arXiv: 2610.12666
iRAG starts with an empty KB populated through human decisions, stores decisions as retrievable precedents, and progresses from silent observation → evidence-grounded contestation → autonomous handling with abstention. Matches/reduces final-decision error under drift vs an independent baseline; retrieval decay is characterized.
Key insight: iRAG grows a decision-precedent KB from human choices: observe → contest with evidence → act with abstention.
DP-MIND: Differentially private memory-guided Long context Inference
Zhihao Liu; Yitian Cheng; Shuang Liu et al. arXiv: 2610.13132
DP-MIND scales differentially private inference to million-token contexts via a memory-guided framework for LLM-as-a-Service settings where user queries include sensitive contexts. Existing DP methods struggle at million-token scale due to context capacity and utility collapse under DP noise.
Key insight: DP-MIND brings differentially private inference to million-token contexts via memory-guided compression for hosted LLMs.
Adaptive DHiSparse: Speculative Decoding over a Load-Adaptive KV Hierarchy
Naji Alazhar; Wassel Alazhar arXiv: 2610.13708
Adaptive DHiSparse supports speculative decoding (DSpark/DFlash2) over hierarchical sparse KV that adapts residency to load: protect the working set, keep extra history while GPU memory is spare, recover the rest via exact fused retrieval. Addresses long-context serving that runs out of KV memory before hardware throughput is useful.
Key insight: Adaptive DHiSparse pairs speculative decoding with load-adaptive hierarchical sparse KV for long-context serving.
Kronecker Product Attention: Memory-Efficient KV Compression via Structured Factorization
Bahareh Nikpour; Hossein Rajabzadeh; Maryam Dialameh et al. arXiv: 2610.12657
KPA represents key/value activations as sums of Kronecker products of compact token-dependent matrix factors for memory-efficient KV compression. Strictly generalizes several prior factorizations while aiming to preserve expressivity vs MQA/GQA/MLA/TPA trade-offs.
Key insight: Kronecker Product Attention compresses KV via structured factorization as an alternative to spill-to-disk persistence.