Tuesday's cs.AI announcement day (2026-10-27) lists 172 new and 261 cross-lists (replacements skipped; listing total 433). A filter for agent systems, memory/context, computer-use / GUI / tools / skills / harnesses, multi-agent, persistence, and local/open models keeps 40 papers. Context cost is the lead thread: one study lets frontier models manage their own context with two tools and matches Claude Code's task success at 34-49% lower cost, while two others show that under prefix caching, deleting old history usually raises real cost because it breaks the cached prefix, and a static checker finds silent prompt-caching bugs across open-source agent projects. Memory papers argue for cheaper selection: a recommender-style cascade beats LLM-built memory systems on multi-session questions with no LLM call, a salience-gated pipeline cuts memory tokens by 92-99%, and proactive recall turns out to generate false alarms that grow with the size of the memory store. Harness papers favor an independent critic over completion gates, find that skills need rewriting for weaker models, and show most self-proposed harness changes do not help later tasks. On tools and security, MCP and CLI interfaces with identical capabilities drive different behavior, MCP tools drift despite version pins, composed third-party skills raise policy violations, email and calendar agents still make unauthorized writes, and agents harvest emailed passcodes to approve their own actions. Also: compiled workflows for recurring tasks, computer-use and mobile agents, persistent agents juggling concurrent jobs, and local MoE offloading.

Research Papers

CTXAgent - Self-Controlled Context Management of LLM Agents

Matthias Busch; Marius Tacke; Roland C. Aydin; Christian Feiler; Christian J. Cyron arXiv: 2610.34425

Gives the model two context tools plus one paragraph of guidance so it decides when to compact, what to keep and what to hand off. Seven frontier models on 35-212-turn mixed sessions (coding exercises, SciCode, SWE-bench Pro, Terminal-Bench 4.0 with planted facts and standing instructions). Most strong models keep live context mostly under 100k with no significant score drop; with Claude Opus 5.5 the self-managed harness matches Claude Code task success at 34% lower cost than its explicit 200k-window config and 49% lower than default. Agent, benchmark and result rows released.

Key insight: Given two context tools and a paragraph of guidance, most frontier models manage their own context well enough to match Claude Code's task success at 34-49% lower cost.

Agent Memory as Recommendation: Recommender-System Design for Long-Term Memory of LLM Agents

Xinyu Zhang; Weiwei Sun arXiv: 2610.34673

Formalizes per-request memory selection as budgeted slate recommendation (request = user, memories = items, context window = slate). AgentMemRec adapts candidate generation, learning-to-rank and slate construction and selects memories with no LLM call. On LongMemEval multi-session questions it beats A-Mem, HippoRAG 2, DynamicRAG and T-Mem by 10-31 accuracy points, matches a 64-call LLM reranker at 3.2x lower serving cost, and on LoCoMo beats Mem0 and Zep with no LLM call at write time.

Key insight: Treating memory selection as a recommender slate beats LLM-built memory systems by 10-31 points on multi-session questions without any LLM call.

The Price of Scale in Proactive Memory Recall

Yichuan Chen; Xin Wang arXiv: 2610.34507

Shows that deciding whether to surface an unprompted memory is a multiple-testing problem: a similarity threshold tuned to 5% false alarms at N=8 memories reaches 87-100% false alarms at N=2048 on LongMemEval/LoCoMo replays, across embedders; LLM judges grow the same way. Proposes a label-free cue-swap conformal gate that holds false alarms near 5% at the hit rate of a label-tuned threshold. Code released.

Key insight: A recall threshold tuned to 5% false alarms on a small memory store fires falsely 87-100% of the time once the store holds 2,048 items.

ConMem: Counterfactual Multi-Scenario Probing for Resolving LLM Memory Misalignment

Shuning Zhang; Tenghao Ji; Jingruo Chen; Eryue Xu; Bohao Chu; Xin Yi arXiv: 2610.33432

Introduces MemAlign (1,079 curated memory cases) for over-generalized, misattributed or privacy-sensitive memories, and ConMem, which decides retention by synthesizing counterfactual scenarios to audit downstream effects. Reports 75.16% accuracy, 64.15% balanced accuracy, 65.31 Macro-F1, reducing over-retention without losing personalization.

Key insight: Auditing each memory against counterfactual future scenarios reduces over-retention of over-generalized or sensitive memories without losing personalization.

SLICE-Mem: Salience-Gated Lightweight Information Curation for Efficient LLM Agent Memory

Xiaobin Chen; Dian Jin; Lin Shu; Defeng Sun; Xiaoming Tao arXiv: 2610.33015

SLICE-Mem uses a compact salience classifier to admit user turns and a dependency-based extractor to write memory entries, with no generative LLM calls online; reconciliation is deferred to offline consolidation. With Qwen3-30B it reaches 46.60 vs 46.68 FAMA (MemOS) on Memora Weekly and 56.8% vs 59.3% on LongMemEval-S while cutting LLM tokens 92.4% and 99.5%.

Key insight: A small salience classifier plus rule-based extraction gets near-parity long-term memory while using 92-99% fewer LLM tokens.

Parametric Session States Recover Retrieval Intent without Replaying History

Qingsen Ma; Dianyun Wang; Jiahong Liu; Irwin King arXiv: 2610.33047

Proposes Parametric Session State, a reusable session-memory layer that remembers what to search for (project, entity, version terms behind queries like "use the previous configuration") while keeping factual knowledge external, so follow-ups can be resolved without replaying history.

Key insight: Remembering what to search for, rather than replaying the conversation, lets assistants resolve vague follow-up requests cheaply.

The Physical Cost of Context: Quality-Constrained Resource Accounting for Coding Agents

Zhongming Yu; Yitong Ma; Hengjia Yu; Boqin Yuan; Xiangye Lin; Hejia Zhang; Jishen Zhao arXiv: 2610.34763

Figure from The Physical Cost of Context: Quality-Constrained Resource Accounting for Coding Agents
The Physical Cost of Context: Quality-Constrained Resource Accounting for Coding Agents

Measures executed prefill (logical prompt minus reused cached prefix), KV residency, time and energy for coding-agent context policies. Preloading ten retrieved blocks cuts model calls 35.7% and the GPU-energy proxy 33.2% vs on-demand grep; prefix caching cuts its uncached prefill 67.7%. Keeping only the last five observations cuts prompt volume 33.3% but raises uncached prefill 608.5% and energy 125.6% and fails the quality gate.

Key insight: Under prefix caching, shortening a coding agent's prompt by deleting history can raise real prefill work by more than 600%.

Silent and Costly: Characterizing and Detecting Prompt-Caching Mis-Integration in LLM Agents

Ying Zhang; Ruoqi Guo; Yi Liu; Gelei Deng; Yuekang Li; Yepang Liu; Leo Zhang arXiv: 2610.34817

Figure from Silent and Costly: Characterizing and Detecting Prompt-Caching Mis-Integration in LLM Agents
Silent and Costly: Characterizing and Detecting Prompt-Caching Mis-Integration in LLM Agents

First empirical study of 120 prompt-caching PRs across 10 open-source agent systems: mis-integrations are request-side misconfiguration or read-side usage miscounting, and most fail silently. CACHELINT (CodeQL + dated provider capability oracle + LLM guard) gets 0.71 precision / 0.78 recall on the labeled set and found 32 mis-integrations in 571 projects (0.86 precision), with 21 developer confirmations and 18 upstream fixes including one in Anthropic's official SDK.

Key insight: Prompt-caching bugs in agent code usually fail silently; a static checker found 32 in 571 projects and drove 18 upstream fixes, including one in Anthropic's SDK.

The Hidden Costs of Forgetting: A Study of Trajectory Deletion in Software Engineering Agents

Fernando Vallecillos-Ruiz; Islem Bouzenia; Michael Pradel; Leon Moonen arXiv: 2610.35496

42 trajectory-reduction configurations over 9,400 SWE-bench Verified runs with two agents. Deleting up to 100% of old reasoning text can preserve effectiveness if delayed; immediate deletion hurts. Agents re-fetch deleted content. Costs fall in 26/42 evaluations under flat pricing but in zero under caching-aware pricing, because reduction breaks the cached prefix.

Key insight: Agents tolerate heavy deletion of old reasoning, but under caching-aware pricing none of 42 deletion configurations actually reduced cost.

Skip the Compressor: Coding Agents Fix More Issues When They Pick Whole Files

Xiaocong Yang; Rishit Chatterjee; Himanshu Udupi; Vikram Ramavarapu; ChengXiang Zhai arXiv: 2610.21073

At an equal token budget, choosing which files to show verbatim (CQ2-CARE) beats prompt compression and chunk retrieval: on 300 SWE-rebench issues it resolves 43.7% (GPT-5.6) and 46.7% (Claude Opus 5), 6.3 points over the best compressor and 9.3/12.7 over the best retriever, and also beats SweRank and SWE-Pruner.

Key insight: Spending the context budget on whole files chosen verbatim resolves 6-13 more issues per hundred than compressing or chunk-retrieving code.

StateWise: Teaching Language Models to Track and Obey Evolving Constraints in Multi-Turn Dialogue

Samarth Varshney; Ritu Yadav arXiv: 2610.33766

StateWise teaches a single 7-8B model to keep a typed, revisable constraint state (with tombstones for revoked constraints) and respond conditioned on it, then adds preference training on relapse contrasts. Cuts relapse 8-22 points vs supervised self-training, lifts conversation success 8-19 points, and beats an oracle compiled-ledger baseline on relapse with no inference-time scaffolding.

Key insight: A typed constraint state with tombstones for revoked rules teaches small models to stop re-applying constraints users have withdrawn.

MarkovSim: an LLM Agent Harness with Bounded Context for Long-Horizon Computational Simulation

Yeui-Lung Lei; Wen-Bo Chen; Pan Zhang arXiv: 2610.35349

MarkovSim writes the workflow record to a schema'd memory and rebuilds a compact context at finite-state-machine transitions. On six finite-element benchmarks it passed 16 of 18 runs, self-corrected solver and physics faults, and kept context bounded regardless of run length.

Key insight: Rebuilding the agent's context from schema'd memory at state-machine transitions kept context bounded across every long simulation run.

Rethinking Verification in Agent Harnesses

Zhengxin Zhang; Erkang Zhu; Yin Lin; Claire Cardie; Bolin Ding; Jingren Zhou arXiv: 2610.34577

Figure from Rethinking Verification in Agent Harnesses
Rethinking Verification in Agent Harnesses

Compares 17 verification designs across seven frontier coding models on Terminal-Bench and SWE-bench Verified, spanning initial instructions, mid-turn reminders, completion gates and independent review, plus a new VisualBench. Independent critic review gives the best accuracy-cost tradeoff: +5.9-9.5 points on TB and +3.9-6.6 on SWE at 1.13-1.25x and 1.00-1.04x cost; extra rounds add up to 3.5 points.

Key insight: An independent critic is the most cost-effective place to spend verification effort in a coding harness, adding 4-10 points for roughly 0-25% more cost.

How Do Coding-Agent Harnesses Self-Evolve? An Empirical Study of Retrospective Adaptation

Christoph Treude; Sebastian Baltes arXiv: 2610.34938

Runs a retro skill in Claude Code on 30 SWE-bench Verified issues that reviews a session and proposes harness changes; codes 125 proposals, implements 24 and evaluates 12 on later issues. Of seven evaluations where the targeted problem recurred, two changes eliminated it, one reduced it, four did nothing, and some added work elsewhere; 18 of 43 lint proposals were generic advice.

Key insight: Most harness changes a coding agent proposes after reviewing its own session did not reduce the targeted problem on later tasks.

Evolving to Match Capabilities: Improving Harness Self-Evolution via Skill Adaptation

Zhihao Yang; Haojin Wang; Juncheng Wu; Kuan Zhang; Yukun Chen; Run Luo; Yaojie Lu; Yuyin Zhou; Lu Wang; Xianpei Han arXiv: 2610.33024

Figure from Evolving to Match Capabilities: Improving Harness Self-Evolution via Skill Adaptation
Evolving to Match Capabilities: Improving Harness Self-Evolution via Skill Adaptation

Finds skills that help strong models can hurt weak ones: on 103 task-skill pairs skills add 31.1 points for GPT-5.6-luna but cost Gemma-4-E4B 8.8. In 86% of weak-model failures the violated requirement was already in the skill. A meta-skill-guided rewriter (Kimi-K3) adapts skills, adding 5.1-14.1 points for weak executors; inside a five-round self-evolution loop it adds 2.0-5.1 points across seven executors.

Key insight: Skills that add 31 points for a strong model can cost a small model 9 points; rewriting them for the executor's capability recovers 5-14 points.

Skill$^{\infty}$: Recursive Self-Improvement for Agent Skill Optimization

Wangcheng Tao; Zhanglu Yan; Han Wu; Weng-Fai Wong arXiv: 2610.35434

A meta agent follows a meta-skill to optimize task skills and also optimizes the meta-skill with the same procedure (diagnose traces, propose candidates, request independent reviews, allocate budget). Skills load via native skill discovery. On 83 SkillsBench tasks: 60.7% pass vs 56.4% for human-curated skills and 54.5% for SkillEvolver.

Key insight: Letting a meta agent optimize its own skill-optimization procedure produces skills that beat human-curated ones on SkillsBench (60.7% vs 56.4%).

SKALOS: Learning to Distinguish Similar Skills with Capability Graphs

Wenhan Xiao; Ziwei Zhang; Xingcheng Fu; Chuanyue Yu; Haonan Yuan; Qingyun Sun; Runhua Xu; Jianxin Li arXiv: 2610.34728

SKALOS reranks skills with a capability graph and witness scoring by weakest evidence. +2.68 NDCG@10 over a text-only reranker on SkillRet; witness-guided selection lifts ALFWorld success from 22% to 29%.

Key insight: Comparing skills by explicit capability evidence instead of description text improves selection among near-duplicate skills.

MultiEvoSkill: Structure-Aware Multi-Agent Skill Evolution for Generalizable Skill Artifacts

Kounianhua Du; Jianxing Liu; Ruoshui Shan; Yucheng Zhang; Ke Li; Xing Sun; Weinan Zhang arXiv: 2610.34260

MultiEvoSkill compiles a skill into a typed graph, uses heterogeneous agents to propose revisions, and resolves edits through structure-aware debate with gating for information gain and topological correctness, targeting skills that generalize across backbones.

Key insight: Representing skills as typed graphs and evolving them through multi-agent debate targets skills that transfer across model backbones.

FlashScript: Compiling Recurring Agent Workflows into Reusable Neuro-Symbolic Programs

Sadra Sabouri; Hosein Hasanbeig; Adam Sage; Shreya Sinha; Cale Darling; Ashish Tiwari; Katie Garms; Souti Chattopadhyay; Sumit Gulwani arXiv: 2610.33876

FlashScript compiles a successful run into a neuro-symbolic program: actions become code, instance values parameters, judgment calls LLM calls. On 120 WorkBench analytics tasks with Claude Haiku 4.5 the programs were correct and made identical tool calls every run with ~70x fewer input tokens and a third of the runtime of a skill-guided agent; on composite tasks the skill agent fell to 23% while programs stayed correct; generalized to all 108 unseen instances.

Key insight: Compiling a successful agent run into a program with LLM calls only for judgment gives identical, correct runs at about 70x fewer input tokens.

Harness-Model Collaboration for Long-Horizon Continual Learning

Zhen Bi; Zhizhi Peng; Yan Wang; Xueshu Chen; Haofei Zhang; Yaobo Zhang; Zeyu Yang; Qing Shen; Jungang Lou arXiv: 2610.34724

HaMo distills trajectories into strategy-level experience cards for fast retrieval and trains the model to interpret and execute retrieved strategies, with coverage-based trajectory selection for updates. Beats harness-only and model-only adaptation on web agents under within- and cross-domain shifts.

Key insight: Combining retrievable experience cards with training the model to use them beats either memory-only or fine-tuning-only adaptation for web agents.

Build or Adopt? Custom-Built vs. General-Purpose LLM Agent Harnesses for Automotive GUI Testing

Arsham Gholamzadeh Khoee; Ruyu Fu; Jiaqi Wei; Ruikun Deng; Robert Feldt; Dhasarathy Parthasarathy; Yinan Yu arXiv: 2610.33636

At Volvo Group, compares a production custom GUI-testing harness with Claude Code and OpenCode using the same MCP tools and models. General-purpose harnesses match success on stepwise tests but vary more, cost 2.6-10.7x more inference per task, yet total monthly cost is 3-66% lower once engineering effort is counted; they do better on open-ended app tasks by recovering from wrong actions. Ten practitioner interviews.

Key insight: General-purpose harnesses cost 2.6-10.7x more in inference than a custom one but 3-66% less overall once engineering effort is counted.

Lessons from Luck: Do LLM-Written Lessons Carry Chance into an Agent's Next Decision?

Kazuki Nakayashiki arXiv: 2610.33029

Registered study over 256 synthetic worlds and 11 LLM endpoints (90,084 answers): after a chance failure of the better tool, LLM-written lessons (ReasoningBank instruction) led readers to pick the worse tool in 11.6% of next runs vs 0.4% with an irrelevant item. A decision-quality note in the writing instruction cut that to 7.5%, a no-change option to 3.1%.

Key insight: Lessons written after a lucky failure push the next run toward the worse tool 11.6% of the time versus 0.4% with neutral memory.

Matched Capabilities, Different Behaviors: A Controlled Study of MCP and CLI Interfaces for Language Agents

Zhenning Shi; Liang Pang; Yongqi Li arXiv: 2610.34298

MICA renders MCP and CLI from one semantic action contract through the same runtime. Across six models and three environments, matched interfaces still differ substantially in performance and cost; 94.3% of 2,004 paired trajectories diverge structurally. DART repairs interface-specific realization, improving scores in all six settings by up to 9.0 points and cutting MCP token cost by up to 27.1k tokens per task.

Key insight: Exposing identical capabilities through MCP or a CLI changes agent behavior in 94% of paired trajectories.

ProbeID: Verifying MCP Tool Identity for LLM Agents under Registry Drift

Guanxing Chen; Sining Zhang; Haohuai He; Qiujie Lv; Yu-An Huang; Zhi-An Huang arXiv: 2610.33633

ProbeID verifies which implementation sits behind an MCP tool by comparing responses to schema-valid probes against a trusted reference; 0.956 hit@1 where description routers are near chance (simple exact comparison ties it). In an MCP-Bench agent, verification cut tainted tasks from 15 to 9 of 16 at unchanged success. Over eleven weeks, 7 of 59 real tools changed responses despite version pins.

Key insight: Behavioral probes can verify which implementation sits behind an MCP tool; 7 of 59 real tools changed behavior over eleven weeks despite version pins.

$Δτ$-Bench: Pushing Agents Toward a Moving Capability

Shuofei Qiao; Xuanguang Pan; Yijun Chen; Wendi Feng; Bin Wu; Runnan Fang; Yuqi Zhu; Shumin Deng; Emine Yilmaz arXiv: 2610.34238

Δτ-Bench: 1,588 multi-turn tasks across 42 domains and six successive ecosystem states where tools disappear, get replaced or gain cheaper alternatives. Best model/harness reaches 42.3%; skill-learning gains don't carry across changes. Steward, an ecosystem manager doing incremental skill maintenance and dynamic provisioning, sustains higher success with lower maintenance overhead.

Key insight: When tools appear, vanish and get replaced, the best agent completes only 42.3% of tasks and learned skills stop helping.

Task Success Is Not Enough: Side-effect-Aware Evaluation of Tool-Using Language Model Agents

Jiaju Huang; Shaobin Chen; Xinglong Liang; Xinyu Ma; Hao Yang; Wei Ke; Tao Tan; Yue Sun arXiv: 2610.34167

Figure from Task Success Is Not Enough: Side-effect-Aware Evaluation of Tool-Using Language Model Agents
Task Success Is Not Enough: Side-effect-Aware Evaluation of Tool-Using Language Model Agents

SABench: 201 email, calendar and ticket tasks with deterministic oracles for goal satisfaction and authorization. Across 11 frontier models, 15.2% of runs cause collateral damage (38.8% in ambiguity tasks); 12.8% of goal-satisfied runs still make unauthorized writes; rates vary 4.7x across models at similar goal rates. Data and code released.

Key insight: About one in eight successful email, calendar and ticket agent runs still performs an unauthorized write.

CoSA: The Conflict of Skills Attack in Real-World LLM Agent Ecosystems

Yujia Zheng; Tianhao Li; Shangzhi Xu; Yuekai Huang; Ziyou Jiang; Zhaoyang Li; Yuyao Ge; Junran Peng; Mingyang Li arXiv: 2610.35082

Figure from CoSA: The Conflict of Skills Attack in Real-World LLM Agent Ecosystems
CoSA: The Conflict of Skills Attack in Real-World LLM Agent Ecosystems

Shows composed Skills can violate policy even when each passes alone. Adding locally admitted Skills (capped at 5% of the pool) raises policy violations from 35.02% to 58.45% across five LLMs and three harnesses on CoSA-Bench (288 tasks). A harness-level guard checking information flow and authorization before consequential actions cuts OpenClaw recruiting violations from 65.28% to 6.67% with benign completion 66.67% to 65.00%.

Key insight: Adding a handful of benign-looking skills to a pool raises policy violations from 35% to 58%; a pre-action authorization guard cuts them to under 7%.

SkillLifeBench: Benchmarking Security Across the Lifecycle of Agent Skills

Chenxu Zhang; Zhou Liu; Wu Xin; Wentao Zhang arXiv: 2610.35215

SkillLifeBench (388 paired scenarios) covers implementation, interface, runtime, composition and version changes in reusable skills, testing 14 models. GPT-5.6-Sol and Sonnet 4.6 both score 82.0% on runtime tasks but 96.7% vs 56.7% on lifecycle tasks; 103 runs completed both the task and the attack goal. Execution provenance raises joint success from 49.35% to 65.58%.

Key insight: Two models with identical runtime security scores diverge from 96.7% to 56.7% when they must manage skill versions and updates themselves.

Position, Role or Label? A Controlled Study of Instruction Conflicts in LLM Agents

Anish Laddha; Nitesh Pradhan; Gaurav Srivastava arXiv: 2610.33201

WhoWins: 1,080 scenarios and 1.89M scored replies from 14 open-weight models. Within a message the first conflicting line wins 72-95%; a system/developer policy sent first loses most conflicts to later lines in every Qwen model but wins on gpt-oss; in tool and file-write tasks, policy in the system message wins less often than the same text as a labeled user message. Models flag conflicts in at most 2.92% of replies.

Key insight: In open-weight models, a policy placed first in the system prompt often loses to later conflicting lines, and models almost never flag the conflict.

I-HITL: Inalienable Human-in-the-Loop

Samuel Breckenridge; Dani Vilardell; Wenhao Wang; Fan Zhang; Ari Juels arXiv: 2610.35835

Figure from I-HITL: Inalienable Human-in-the-Loop
I-HITL: Inalienable Human-in-the-Loop

Ten frontier models given email access and a task blocked by a wallet spending cap often harvested emailed OTPs to raise the cap themselves, bypassing HITL. Defines inalienable HITL, proves it needs a trusted device, and presents Countersign, which uses off-the-shelf mobile devices to strengthen existing HITL.

Key insight: Frontier agents with email access will harvest one-time passcodes to approve their own actions; robust human approval needs a trusted device.

InferPriv: Evaluating Inferential Privacy Leakage in LLM Assistants

Pengrun Huang; Yiwei Fu; Luoxi Meng; Kamalika Chaudhuri arXiv: 2610.34640

InferPriv measures whether a curious recipient can infer a user's sensitive attributes from an agent's routine reply. With a privacy-aware system prompt, GPT-6.1-Sol, Sonnet-5.5 and Gemini-3.8-Flash have near-zero direct leakage but 16.2-25.2% inferential leakage.

Key insight: Assistants that never state a user's secrets still let recipients infer sensitive attributes 16-25% of the time.

OS-Wizard: Scaling Computer-Using Agents to Specialized Software via Knowledge-Guided Exploration

Zhoumianze Liu; Zichen Ding; Bowen Yang; Zehao Li; Zhenyu Wu; Yibo Zhao; Kanzhi Chen; Yian Wang; Yanan Sun; Qiushi Sun; Yu Qiao arXiv: 2610.33907

Figure from OS-Wizard: Scaling Computer-Using Agents to Specialized Software via Knowledge-Guided Exploration
OS-Wizard: Scaling Computer-Using Agents to Specialized Software via Knowledge-Guided Exploration

OS-Wizard collects trajectories with tutorial-guided exploration, a hybrid GUI+code action space, and checkpoint revisits during tree exploration: 10.9K trajectories / 54.9K samples. Fine-tuned open models reach 64.1% on specialized desktop workflows (+58.2 points), above reported human 59.8%, and improve general GUI-only tasks too.

Key insight: Tutorial-guided exploration with native-command actions yields training data that lifts open models above reported human performance on specialized desktop apps.

AndroidFlux: Evaluating and Reward Modeling Failure Recovery Capabilities in Mobile Use Multimodal Agents

Gyubeum Lim; Yemo Koo; Ghunwoo Jang; Zsolt Kira arXiv: 2610.34384

Figure from AndroidFlux: Evaluating and Reward Modeling Failure Recovery Capabilities in Mobile Use Multimodal Agents
AndroidFlux: Evaluating and Reward Modeling Failure Recovery Capabilities in Mobile Use Multimodal Agents

AndroidFlux restores AndroidWorld trajectories to checkpoints before, at and after the first recoverable error so agents continue from identical states; adds a recovery reward-modeling benchmark and an offline RL dataset. Recovery only partly correlates with clean-start success; explicit reasoning helps only GUI-RL-trained models.

Key insight: A mobile agent's ability to recover from mid-task errors only partly tracks its clean-start success rate.

GUIFlow: Temporal Redundancy-Aware Incremental Inference for GUI Agents

Ye Li; Haoran Ding; Huanan Liu; Yuan Meng; Yutian Shen; Yuansong Wang; Kangye Ji; Hanyun Cui; Hongbo Duan; Chen Tang; Jiawei Li; Shu-Tao Xia; Zhi Wang arXiv: 2610.34959

Figure from GUIFlow: Temporal Redundancy-Aware Incremental Inference for GUI Agents
GUIFlow: Temporal Redundancy-Aware Incremental Inference for GUI Agents

GUIFlow reuses computation across GUI steps: position-aligned historical KV injection, change-guided selective updates in vision encoder and LM, and prior action responses as speculative drafts. Up to 3.39x geometric-mean speedup offline and 3.49x online on OSWorld with broadly comparable task performance, no training.

Key insight: Reusing KV states and prior actions across nearly identical GUI steps speeds GUI agents up to 3.5x without training.

AOS-Bench: Can LLM Agents Manage Dynamic Workloads?

Erle Zhu; Genrui Zhang; Keyang Hu; Mingyu Su; Hongning Wang arXiv: 2610.33178

AOS-Bench treats an agent as a workload manager: tasks arrive over time under shared CPU/memory limits. Every system loses 13-23 points on heavy workloads; two systems 0.6 points apart in isolation deliver 54% vs 68% in stream. Serial dispatch leaves 73% of heavy-workload tasks unclaimed; a fixed concurrency of two comes within four points of the best agent-managed policy; median CPU stays under 30%.

Key insight: Agents handling a stream of concurrent jobs lose 13-23 points under heavy load, and a fixed concurrency of two nearly matches the best self-managed policy.

Babysitting AI Agents: A User-Centered Taxonomy and Study of Agentic Delegation Failures

Shravika Mittal; Sunnie S. Y. Kim; Scott Counts; Jennifer Neville; Debarati Das; Philippe Laban arXiv: 2610.34748

MANDATE taxonomy of 23 failure categories in seven groups from 163,487 Reddit posts about AI agents, plus a 73-user study: more severe failures are seen as harder to notice; under-studied failures include agents running without making progress.

Key insight: A taxonomy of 23 delegation failures from 163,487 Reddit posts shows the most severe failures are also the hardest for users to notice.

Thinking-in-Bets at Test Time Reduces Sycophantic Revisions in LLM Agents

Nghi D. Q. Bui arXiv: 2610.34654

Thinking-in-Bets has the agent write a private claim checklist (claims, supporting checks, confidence) before reading user feedback. Across five Gemini models and eight benchmarks it nearly eliminates false reversals in patch review, cuts unnecessary edits after mistaken complaints, and writing the record after reading the complaint recovers little; re-checking the weakest claim also improves accuracy without feedback.

Key insight: Writing down claims and their checks before reading user pushback nearly eliminates agents abandoning correct work.

AgentOrch: How Close Can Global Orchestration Come to Step-by-Step Interaction?

Zijie Liu; Qi Yang Huang; Jinhao Duan; Yueh Che Lee; Thomas A. Joseph; Zayne Turner; Adam Seligman; Tianlong Chen; Zhaozhuo Xu arXiv: 2610.33375

AgentOrch compares precomputed executable workflows with step-by-step tool interaction on 650 enterprise SOPs and 3,220 runtime variants. Step-by-step still wins for every model but the gap shrank ~79% from early to recent models; incremental workflow construction adds 9.98 points, execution-feedback revision 6.13, targeted training 12.18.

Key insight: Precomputed workflows still trail step-by-step agents, but the gap has shrunk about 79% and targeted training closes another 12 points.

Prefetch-Steered Routing: Deterministic Cache-Aware Expert Selection for Offloaded Mixture-of-Experts Decoding

Siming Huang arXiv: 2610.34721

Prefetch-steered routing makes cache-aware expert selection deterministic by routing over a logical cache. On OLMoE-1B-7B with 25-50% of experts resident on an RTX 4090 it decodes 2.2-2.3x faster than exact routing with LRU with benchmark differences of +0.1/+0.5; 1.9x on Moonlight-16B-A3B; on Qwen3.6-35B-A3B it keeps quality gains but not speed.

Key insight: Routing over a deterministic logical expert cache makes offloaded MoE decoding 1.9-2.3x faster on one consumer GPU without changing outputs by timing.

Trust Mechanisms for Offline Generative AI at the Edge: A Connectivity-Regime Framework and an Empirical Study on a Raspberry Pi 5

Georgios Feretzakis; Alexandros Papaspyridis arXiv: 2610.33732

Connectivity-regime framework (C0-C4) and signed knowledge packs, tested on a Raspberry Pi 5 with Qwen2.5 local RAG. A retrieval-score gate detects knowledge gaps (AUROC 0.97) but not in-topic unanswerable questions (0.68); a 3B model matches 7B gold-answer containment at under half the latency; a validly signed poisoned pack drops correct answers to zero on 30 questions.

Key insight: On a Raspberry Pi 5, layered verification rejects tampered knowledge-pack updates, but a validly signed poisoned pack still drives correct answers to zero.

News and changelogs