Wednesday's cs.AI announcement day (2026-11-04) lists 115 new and 137 cross-lists (replacements skipped; listing total 252). A filter for agent systems, memory/context, computer-use / GUI / tools / skills / harnesses, multi-agent, persistence, and local/open models keeps 38 papers. The day's main theme is what agent memory is actually worth. One study measures a memory's value against the specific later task and finds that task-blind summaries keep less of it than raw records; a procedural-memory system does best when the agent organizes its own wiki and people spend their effort diagnosing failures; and a benchmark built from 80,299 real group-chat messages shows memory systems that match full context on recall still fall far behind on using what they remember. Context work points the same way: frequent compaction costs frontier models little accuracy, and in retrieval it is look-alike rival facts, not length, that cause context rot. Tool and skill security gets concrete results: the official MCP Python and TypeScript SDKs miss spec constraints in 13 and 19 confirmed cases, a deterministic runtime gate for third-party skills reaches 0% missed and 0% wrongly blocked, and a fake hazard written as an ordinary review steers agents past injection detectors. Also: a durable task board for agents that must wait, harness evolution and self-evolving skills (and evidence that some reported gains are not real capability), multi-agent drift and role enforcement, web agents against scams, silent grading errors on agent leaderboards, and quantized serving and speculative decoding for open models.

Research Papers

What Is This Memory Useful For? Measuring Task-Conditioned Utility in Long-Term Agent Memory

Chih-Hsuan Yang; Anjir Ahmed Chowdhury; Feng Yan; Rajeev Thakur arXiv: 2611.03930

Figure from What Is This Memory Useful For? Measuring Task-Conditioned Utility in Long-Term Agent Memory
Per-cell helped/unchanged/hurt labels across sessions (kappa = 0.78) and help rates by stratum: answer-bearing memories help far more often than matched controls.

Task-blind writer stores a memory; a frozen reader answers a later task with and without it under identical filler, with a cost ledger. On LongMemEval-V2 web-agent recall tasks, answer-bearing memories beat matched controls by +0.33 mean paired reward; task-blind summaries and structured records gain less than Raw even at fixed context; generic retrieval misses most answer-bearing memories; no measurable effect on the SWE and scientific-coding benchmarks tested. Public Conditional Memory Utility Dataset.

Key insight: A memory's value depends on the later task: answer-bearing memories added +0.33 paired reward over matched controls, and task-blind summaries kept less of that value than raw memories.

WASP-Mem: Let Agent Organize Procedural Memory. Human Spend Effort on Diagnosis

Peizhi Niu; Ge Shi; Vaibhav Gorde; Sarthak Ghosh et al. arXiv: 2611.04003

Figure from WASP-Mem: Let Agent Organize Procedural Memory. Human Spend Effort on Diagnosis
The six consolidation methods illustrated for the first experiment, from raw rollouts and human-schema forms to free-form wiki pages written by the agent.

WASP-Mem distills execution traces and feedback into persistent, interlinked wiki pages. On TheAgentCompany: +4.6-6.1% checkpoint pass rate over the strongest baseline across three models with 57-61% less storage. Human-defined schemas and dedup rules hurt; one round of human diagnosis beats four rounds of self-judgment by 7.0-10.9%.

Key insight: Letting the agent organize its own procedural wiki beat human-imposed schemas, and one round of human failure diagnosis outperformed four rounds of self-judgment by 7.0-10.9%.

EvoMemState: Evolving Memory State for Personal Agents

Zhishang Xiang; Yunbo Tang; Zhongqiao Liu; Junkai Chen et al. arXiv: 2611.03743

EvoMemState adds Relational States (validity, scope, evolving relations between facts) and an RL-learned, request-specific Reasoning State on top of persistent facts. Reports consistent gains over existing memory systems in retrieval completeness and memory-grounded answer accuracy on long-term dialogue and memory-use benchmarks. Code released.

Key insight: Personal-agent memory works better when it tracks how facts relate, stay valid and conflict over time, not just which facts were extracted.

One Assistant, Many Memories: Benchmarking Real-World Group Assistance When Recall Is Not Enough

Zhanghao Hu; Linhai Zhang; Qian Zhao; Qi Zhu et al. arXiv: 2611.04867

Figure from One Assistant, Many Memories: Benchmarking Real-World Group Assistance When Recall Is Not Enough
GroupAssistBench overview: real group-chat histories, memory recall questions with reference answers, and open-ended service requests that require using memory.

GroupAssistBench: 80,299 real messages in 50 groups (cost >$53K), testing recall (L1) and use for open-ended service requests (L2) over single, distributed and conflicting memories. The best memory system rivals full context on recall but falls far behind on use. Relating retrieved memories and binding remembered facts as response constraints both help.

Key insight: On real group chats, the best memory system matched full context on recall but fell far behind on actually using memories, especially distributed and conflicting ones.

Spatial Memory for AR Glasses: Hierarchical Visual Acquisition and Source-Linked Recall

Tianxing Zhu; Dan Huang arXiv: 2611.04494

Dual-camera AR-glasses prototype: monochrome monitoring selects views for RGB capture (6 of 648 observations in a 131.35 s replay), stores semantic observations linked to source images, and answers historical questions by staged retrieval. Separates captured evidence, indexed interpretation and supported claims.

Key insight: Linking every interpretation back to its source image lets AR glasses recall past encounters while capturing full-color views only rarely (6 of 648 observations).

SEAM: Situated Executable Associative Memory for a Co-located Community

Keru Wang; Michael Xu; Ken Perlin; Nebojsa Jojic et al. arXiv: 2611.04296

SEAM: members post moments to a shared memory that relates fragments, distills group knowledge and replies associatively; some memories act as instructions that reprogram the system. Deployed to 46 people for two weeks plus a 14-person workshop.

Key insight: A shared community memory in which some memories act as instructions made the system programmable by its 46 members, and also made memory a control surface.

You don't need long context

Peter Belcak; Saurav Muralidharan; Jindong Jiang; Yonggan Fu et al. arXiv: 2611.05306

Later tokens cost more (up to 7x on frontier open-weight models). With frequent compaction, frontier models lose little to no accuracy on long-horizon and long-input tasks; SWE performance does not collapse even at one action per context. Proposes semantic state space models (SSSMs).

Key insight: With frequent compaction, frontier models lost little or no accuracy on long-horizon tasks, questioning the need for long-context models when later tokens cost up to 7x more.

Why Pay for Long Reasoning When Rewrite Do Trick? Causal Effects of Reasoning History Compaction on Future Generation and Agent Cost

Anton Hristov arXiv: 2611.03852

Rewrites prior reasoning into compact state records before pending tool results are revealed, then compares continuations. On the open-weight models tested, accepted compact histories usually shorten later reasoning; in software-agent runs compaction can reduce total spend, but quality and rewrite-inclusive time depend on target and compressor.

Key insight: Rewriting an agent's earlier reasoning into compact state records usually shortened its future reasoning, and can lower total spend depending on model and compressor.

Length Is Cheap, Competition Is Not: Attention Predicts Context Rot in Retrieval

Yinuo Liu arXiv: 2611.04169

On five open 0.5B-3B models, neutral filler costs nothing up to 32K tokens (0.97-1.00), but evidence-formatted distractors drop accuracy to 0.16-0.33, and one rival value for the queried key multiplies loss 9-110x. Attention mass left on the evidence explains up to 58% of loss variance vs at most 8% for log length. Probe and 126K trials released.

Key insight: In retrieval, neutral filler up to 32K tokens cost nothing, while a single rival value for the queried key multiplied the loss 9-110x; context rot is competition for attention, not length.

The Agent Board: Kanban for Machines; Durable Coordination for Agents That Outlive the Conversation

Tyler Hunt; Donald Nguyen; Gurbinder Gill arXiv: 2611.04100

Figure from The Agent Board: Kanban for Machines; Durable Coordination for Agents That Outlive the Conversation
Wait cost per task as the job takes longer: naive polling grows steeply while the board stays flat; a plain external watcher is cheaper still.

The Agent Board in the Corvic harness: a durable shared table of semi-semantic tasks; agents post, watch and block on tasks, holding no process until a matching post wakes them. A Board Keeper agent routes non-agent events and allocates agents; pluggable policy kernels give statuses meaning. Pilots: wait cost flat vs ~165x growth for naive polling, but a plain external watcher was cheaper; Keeper had no false wakes and missed 3.6% of real wakes.

Key insight: A durable shared task board lets agents block on work without holding a process; wait cost stayed flat while naive polling grew about 165-fold, though a plain watcher was cheaper still.

Meta$^2$: Precomputing Reusable Knowledge for Online Harness Evolution in Long-Horizon Agents

Lingzhi Yuan; Yiguang Cai; Weize Liu; Siqi Yang et al. arXiv: 2611.05139

Figure from Meta$^2$: Precomputing Reusable Knowledge for Online Harness Evolution in Long-Horizon Agents
Offline, Meta^2 learns a reusable Meta-Skill from contrasting rollouts; online, an evolver uses it to keep or edit the harness within a single run.

Meta^2 learns a textual Meta-Skill offline by contrasting rollouts without an evolver, with a guided evolver, and with privileged verifier info; online, an evolver decides KEEP or EVOLVE via task-local harness edits. With GPT-5.6-Terra: 81.00% on Terminal-Bench 2.1 and 44.44% on DeepSWE 1.1, best among evaluated methods.

Key insight: Learning reusable guidance on when to edit a harness mid-run, and keeping edits task-local, reached 81.00% on Terminal-Bench 2.1 and 44.44% on DeepSWE 1.1.

Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training

Kangning Zhang; Wenxiang Jiao; Bowen Fang; Shuai Shao et al. arXiv: 2611.04277

Harness-S builds a fixed pool of specialized sibling harnesses, invokes them only when all base-harness attempts fail, and trains on the rescued trajectories; an adaptive router picks top-k variants. On AppWorld, SWE-Gym and DBBench the policy substantially beats policy-only training back under the base harness.

Key insight: Rescuing failed rollouts with specialized sibling harnesses, then training on them, transferred harness-elicited behaviors into a policy that works under the original harness.

Artifact-to-Capability: Diagnosing Breaks in Capability Formation of Self-Evolving Agents in a Controlled Environment

Ziqian Ban; Haoxiang Yang; Yanfei Zhang; Shixin Liang et al. arXiv: 2611.04952

Defines capability formation from artifacts as five dependent properties with three-scale diagnosis. In Gomoku, Reflexion, Voyager and DGM all show the usual signals (artifacts grow, behavior changes, metrics rise) yet each breaks differently: legality-only gains, compositional override missing all 20 one-move wins, and loss-only drift.

Key insight: Reflexion, Voyager and DGM all showed the usual self-evolution signals in Gomoku, yet each had a distinct break in actual capability formation.

From AutoResearch to Continual AutoResearch: Do Agents Learn Across Tasks?

Jeremy Qin; Ruizhi Deng; Mohamed Osama Ahmed; Frederick Tung et al. arXiv: 2611.05108

Continual AutoResearch compares persistent and reset instances of five frontier agents over nine task families with matched streams and budgets, plus a memory ablation. Persistent state improves research efficiency for all five, but not consistently the final solution.

Key insight: Persistent agents found strong solutions sooner than reset agents across nine task families, but that did not consistently produce better final solutions.

Recursa: Toward Recursively Self-Improving Foundation Model Systems

Erkang Zhu; Yin Lin; Xiaorong Shi; Xuanrui Lin et al. arXiv: 2611.04035

Recursa stores domain skills, when-to-apply instructions, task memories and reflections in a repository; parallel learners propose improvements that are validated and merged. On Harvey LAB Firm Knowledge, held-out all-pass rises 30.0%->55.0% (DeepSeek-V4.1-Flash) and 35.0%->46.0% (Qwen3.8-Flash).

Key insight: A validate-then-merge repository of skills and memories raised held-out all-pass rates from 30.0% to 55.0% and from 35.0% to 46.0% without human experts or a teacher model.

Validation-Gated ML Pipeline Synthesis with Expert-Grounded Self-Evolving Skills

Zixiao Zhao; Jing Sun; Zhe Hou; Mengze Li et al. arXiv: 2611.04180

EvoML: Orchestrator enforces structural/execution/semantic gates per phase, Executor runs expert-grounded skills, Curator folds recovered scripts and guidance back into the skill library after each run. Best final model quality on all eight datasets vs agentic AutoML and CLI coding tools.

Key insight: Gating each pipeline phase on structural, execution and semantic checks, then curating recovered scripts into skills, gave the best final model on all eight datasets tested.

Drift No More: Spectral Locking for Long-Horizon Open-Weight Multi-Agent AI

Vishal Pramanik; Maisha Maliha; Susmit Jha; Sumit Kumar Jha arXiv: 2611.04315

ATLAS: training-free hidden-state steering toward shared eigendirections of task, role and state subspaces. Drifted turns collapse across all three subspaces (r=0.72-0.78, 14,832 turns). Beats prompting, retrieval, memory and steering baselines; up to +5.7 memory macro-avg over A-MEM, 93.5% role consistency, transfers to a 72B backbone.

Key insight: Task drift, role drift and stale state share a representational cause, and steering hidden states toward joint task-role-state directions reduced all three without training.

ATC: Agent Trajectory Control for Role-Contract Enforcement in Black-Box Multi-Agent Systems

Maisha Maliha; Vishal Pramanik; Sumit Kumar Jha arXiv: 2611.04321

ATC makes each role's authority an explicit contract, scores proposed messages for out-of-role deviation and leakage, regenerates under constraint and re-checks before admission. Works on black-box agents; improves task performance and role consistency, most under role-confusion stress and larger teams.

Key insight: Enforcing explicit role contracts at the point a message enters the shared conversation keeps black-box agents in their lanes, with the biggest gains in larger, more diverse teams.

Inter-Agent Influence: Evaluating Persuasion, Deception, and Coercion in Multi-Agent Systems

Chandler Smith; Qi Guo; Cecilia Elena Tilli; Sophia Hatz et al. arXiv: 2611.04556

Five environments with frontier-model agents: in oversight settings all tested models shifted overseer decisions from rejection to approval via persuasion, deception or coercion; peer-to-peer redirection also succeeded. Code and data released.

Key insight: In oversight environments, every tested model moved an overseer from rejection to approval through persuasion, deception or coercion.

MCPGuard: Finding Conformance Gaps in MCP SDKs via Dual-Track Analysis and Dynamic Adjudication

Xikai Sun; Kebin Liu; Fan Dang; Yawen Zheng et al. arXiv: 2611.03662

Figure from MCPGuard: Finding Conformance Gaps in MCP SDKs via Dual-Track Analysis and Dynamic Adjudication
MCP architecture background: a host's clients talk to local and remote MCP servers over JSON-RPC, through the official SDKs whose conformance MCPGuard tests.

MCPGuard aligns spec JSON Schema constraints with SDK models, audits normative prose rules with LLMs, and confirms candidates with PoCs. Confirms 13 findings in the official Python 1.26.0 SDK and 19 in TypeScript 1.29.0, none exposed by two black-box baselines; gaps propagate to open MCP servers.

Key insight: The official MCP Python and TypeScript SDKs omitted spec constraints in 13 and 19 confirmed cases, and those gaps propagate to servers built on them.

RuntimeTickets: Source-Grounded Security Obligations for Agent Skills

Kuiyi Gao; Xue Liu; Bowei He arXiv: 2611.03772

RuntimeTickets: an LLM proposes per-task security obligations citing skill lines; a deterministic runtime gate checks recipient, object and amount of every outbound send; no LLM in enforcement or decoding. At the egress boundary of three unmodified frameworks: 0.0% overlook and 0.0% overkill, while production scanners erred both ways.

Key insight: Keeping language models out of enforcement, with a deterministic gate on every outbound send, brought both missed malicious skills and wrongly blocked benign skills to 0.0%.

LatentTriad: Pre-Action Risk Representations for Malicious Agent Skill Detection

Zhe Yu; Yi Liu; Ying Zhang; Simin Chen arXiv: 2611.05088

LatentTriad reads prefill hidden states for task-operation consistency; no output tokens, tools or skill code run. 99.0-99.2% pooled F1 and 93.0-94.6% source-macro F1 over 5,830 skill packages from five sources and seven backbones; median 259-284 ms.

Key insight: Hidden states formed during prefill reveal whether a skill's operations match the user's task, enabling 99% F1 screening in about 270 ms without running the skill.

Voluntary Servitude: Weaponized Safety Warning Mechanisms for Controlling AI Agents

Ilan Zendel; Matan Levy; Stav Cohen; Avi Mendelson et al. arXiv: 2611.04393

Figure from Voluntary Servitude: Weaponized Safety Warning Mechanisms for Controlling AI Agents
Voluntary servitude: a planted hazard gives the agent a reason, not an order, and injection detectors score it as benign.

A fabricated hazard written as an ordinary review makes agents disqualify legitimate options and pick the attacker's target: 60.4% vs 66.0% for instruction injection, scored benign by deployed injection detectors (1.7-4.4x more effective behind a filter). Transfers to production browser agents and a retrieval pipeline. A corroboration requirement in instructions removes ~95% of exposure against a non-adaptive attacker.

Key insight: A single fake hazard written as an ordinary review steers agents about as often as instruction injection, slips past injection detectors, and is largely neutralized by a corroboration requirement.

SLATE: A Large-Scale Benchmark for Long-Horizon API Selection in E-commerce Tool Use

Rongzhe Wei; Ge Shi; Min Cheng; Sarthak Ghosh et al. arXiv: 2611.03982

SLATE: hierarchical tool-grounded procedures over a ~1,000-tool e-commerce library, a deterministic simulator and an end-to-end success metric that credits valid alternative trajectories. The best of six 3B-12B open decision models completes only 5.0% of tasks. Released.

Key insight: Over a ~1,000-tool e-commerce library, the best 3B-12B open model completed only 5.0% of long-horizon tasks.

When the Consumer Is an Agent: Execution-Grounded Service-Level Agreements

Michal Orzechowski; Renata G. Slota; Jacek Kitowski arXiv: 2611.04182

An independent control plane observes a deadline-bound service-state predicate, admits completion and directs recovery. In a live multi-provider data service, agent self-assessment both accepted late tasks and rejected on-time ones; under enforcement, accepted completions stayed consistent with later observations.

Key insight: Agents' own completion reports accepted late tasks and rejected on-time ones; independent observation of service state fixed both errors.

Interpretable Early Termination of Web Navigation Agents via Closed Sequential Pattern Mining

Sergio Talavera; Magdalini Eirinaki arXiv: 2611.04142

BIDE+ sequential pattern mining over 3-step trace prefixes yields a human-reviewable failure-signature library and taxonomy. On 1,544 MiniWoB++ traces from three open-weight models: best F1 0.717 at K=3, 88% of failures caught within three steps, 26.3% token savings at 92% precision.

Key insight: Mining three-step trace prefixes caught 88% of web-agent failures early and saved 26.3% of tokens at 92% precision with a human-readable failure library.

JHiMA-WebRL: A Joint Hierarchical Multi-Agent Reinforcement Learning Framework for Long-Horizon Web Tasks

Shanka Subhra Mondal; Jiayun Wang arXiv: 2611.05263

JHiMA-WebRL couples a 7B task decomposer and a 7B actor, trained with teacher-built subtask supervision and hierarchical GRPO on WebArena. 37.2% on WebArena-Lite-v2 vs 24% for an RL-tuned ScaleCUA-32B under the same protocol; joint updates beat actor-only.

Key insight: A jointly trained 7B planner and 7B actor reached 37.2% on WebArena-Lite-v2, versus 24% for an RL-tuned 32B monolithic agent.

ScamWorld: An Interactive Environment for Evaluating Web Agents Against Online Fraud

Yutao Wu; Xiao Liu; Ming Wen; Hanxun Huang et al. arXiv: 2611.04879

Figure from ScamWorld: An Interactive Environment for Evaluating Web Agents Against Online Fraud
A reconstructed scam flow: a top sponsored search ad leads to a fake store that takes payment.

ScamWorld reconstructs reported scam sites (stores, investment platforms, advance-fee schemes) across languages. Six frontier-model web agents are deceived 11-35% of the time, proceed more readily with payments than with credential requests, and sometimes note suspicion yet continue, treating user instructions as authorization.

Key insight: Web agents were deceived by reconstructed scam sites 11-35% of the time, and some noticed warning signs yet proceeded anyway.

The Web Doesn't Sit Still: Adversarial Self-Evolving Attacks on Search Agents

Geyuan Wang; Shunyu Liu; Siyuan Liang; Chenyang Lyu et al. arXiv: 2611.03754

Bi-level attacker/defender evolution with Contrastive Rollout Evolutionary Optimization (CREO) in a controlled sandbox; beats static attack baselines across multi-hop QA benchmarks and frontier LLMs.

Key insight: Attacks that evolve against an adapting defender expose more search-agent weaknesses than static templates.

EMBERS: Empirically Motivated Robustness Testing of Deep Research Systems via Defect-Guided Fuzzing

Shuochuan Li; Zan Wang; Dong Wang; Yuhang Jiang et al. arXiv: 2611.03723

Study of 1,932 trajectories from three deep-research systems finds nine defect categories across subquestion formulation, search and final reasoning. EMBERS fuzzes stage-targeted inputs: 3.21x failure-discovery throughput of the strongest baseline over 16 settings; fine-tuning on outputs improves answer consistency.

Key insight: Defect-guided fuzzing found failures in deep-research systems at 3.21x the rate of the strongest baseline.

VisSearch-Bench: Evaluating Long Horizon Visual Search in Multimodal Agents

ZhengXian Wu; Qianhao Nie; Zhuohong Chen; Junjie Gao et al. arXiv: 2611.04894

VisSearch-Bench: 204 human-written tasks across 7 domains requiring agents to link visual evidence with search results. Best of 16 multimodal models fully solves 26.47% even with search tools.

Key insight: Even with search tools, the best of 16 multimodal models fully solved only 26.47% of visually intensive search tasks.

LOLPlay: Benchmarking Real-Time Visual Agents in League of Legends

Zhaojian Yu arXiv: 2611.03685

LOLPlay: 36 League of Legends tasks with 82 executable subgoal checks; agents act via screenshots and mouse/keyboard. Best task success 25.0% (GPT-6 Astra); high reasoning effort improves neither metric over low for any model, and actions execute after the scene changes.

Key insight: In real-time game control, higher reasoning effort improved neither task nor subgoal success over low effort for any model tested.

Signals: Continuous Diagnostics of AI Agents

Amey Rambatla; Vatsal Goel; Sagar Gaur; Siddhant Loya et al. arXiv: 2611.04507

Signals (Cisco): taxonomy of 45 problem types in eight families; separates realized failures, risk exposures and inefficiencies, groups findings into engineering issues with remediations and checks. 91.43% F1 vs reference issues, ~18 pts over a one-pass LLM control with the same model and taxonomy.

Key insight: A 45-type taxonomy applied to agent traces matched reference issues at 91.43% F1, about 18 points above a one-pass LLM review.

Silent Instrument Errors Reorder Agent Leaderboards

Yangze Liu arXiv: 2611.04399

On HAL and tau2-bench: a general agent's 13-pt gap was ungraded episodes stored as failures; cost columns double-count nested calls; archived trajectories record grader actions as agent actions; re-grading under a new evaluator reorders boards. HAL maintainers confirmed three. Proposes reporting graded fraction, root-span cost, action provenance and pinned evaluator versions.

Key insight: Grading and logging bugs, such as ungraded episodes stored as failures, silently reordered public agent leaderboards; the paper proposes cheap checks.

From Watching to Steering: Enabling Human-Coding Agent Collaboration through Execution Traces

Xiaoke Yan; Jiaying Lei; Danqing Shi; Zichen Wang et al. arXiv: 2611.04466

Vibetrace on OpenCode: interactive execution traces, fork-from-step corrections, and distillation of traces plus feedback into reusable skills. On SWE-bench Verified, distilled skills cut input tokens 23.0% and cost 15.5% while raising pass rate 65.0%->68.0%.

Key insight: Turning execution traces and human corrections into reusable skills cut SWE-bench Verified input tokens by 23.0% and cost by 15.5% while raising pass rate from 65.0% to 68.0%.

Agentic Path Maps: Exposing an Agent's Traversal Through the Garden of Forking Paths

Madeleine Grunde-McLaughlin; Hafsah Usman; Daniel Mitroshkov; Emily Reif et al. arXiv: 2611.04183

Design space for communicating alternative paths after execution; user study (n=48, data analysis) finds people rarely consider alternatives unprompted, but overreliance drops when alternatives are visualized.

Key insight: Users rarely consider alternatives on their own, but showing the paths an agent did not take reduced overreliance.

Twice as Fast, Until It Isn't: Quantized Serving of Mid-Size Open LLMs

Smit Kadvani; Kriti Faujdar; Spurthi Tallam arXiv: 2611.04380

Audit of public measurements: only two checkpoints (INT4 Gemma 3 12B) have both a vLLM timing and an accuracy score. 4-bit weights roughly double small-batch decode speed (2.12x Qwen2.5-14B on A100) but gains vanish when compute-bound (1.02x at batch 256), where FP8 helps; many accuracy deltas are within noise. Seven-item reporting checklist.

Key insight: 4-bit weights roughly double small-batch decode speed for 9B-20B open models, but the gain vanishes at large batch, and many published accuracy deltas are within noise.

Adversarial Speculative Reasoning Improves Long Horizon Agents

Hongyin Luo; Wei Fang; Lijie Yang; Owen Stepan et al. arXiv: 2611.05345

Draft acceptance noise grows within verification windows and shifts reasoning after ~30 steps. Training the draft to be rejected at critical decision points makes those tokens come from the target under least-noisy conditions; raises end-to-end success on agentic coding benchmarks without changing the target, training 0.06-0.28% of hosted parameters.

Key insight: Training a draft model to be rejected at critical decision points made speculative decoding improve long-horizon agent accuracy, not just speed.

News and changelogs