标题:RankEvolve: A Reliable Multi-Agent Auto-Research Harness for Evolving Ranking Models
来源:arXiv
链接:https://arxiv.org/abs/2609.39551
摘要:Auto-research agents, LLM systems that propose, implement, train, and evaluate model changes across iterations, promise to automate applied ML's experimental loop. Over long horizons, execution accuracy is a binding constraint: a change can silently leak held-out data, omit normalization, disconnect a gradient, or leave a train/eval flag unwired, invalidating expensive runs and compounding error across iterations. We present RankEvolve, an auto-research framework for evolving generative ranking models. An Executable Operating Protocol (EOP) declares phases, gates, branches, and loops, and the runtime enforces the compiled state machine. A meta-meta-harness composes complete black-box coding-agent products, including Claude Code and Codex, as execution-graph nodes that review and repair one
标题:How Much of a Harness Does a Strong Agent Need for Autonomous ML Engineering?
来源:arXiv
链接:https://arxiv.org/abs/2609.40303
摘要:Recent autonomous machine learning engineering (MLE) agents have made significant progress on public leaderboards. Often motivated by progress stagnation over long-horizon cycles and limited Large Language Model (LLM) primitives, modern MLE agents are deployed on top of increasingly elaborate machinery: multi-agent orchestrators, dedicated retrieval subagents, and more. While such harnesses expand, the use of more primitive but improved coding agents - where LLMs have direct access to the execution environment through read, write, and bash primitives - has received little attention in the field. In this paper we find that, under an equal time budget and the same frontier LLM backbone, open-source state-of-the-art harnesses provide no advantages over a single session of a minimal-harness co
标题:Hard-Gate Candidacy in a Deployed Validator Suite
来源:arXiv
链接:https://arxiv.org/abs/2609.39037
摘要:Before a validator can be promoted to a hard gate on a deployment pipeline, it has to be shown that its firing separates outputs that reach users in working order from those that do not. We run that screen on 13 validators in a deployed generative agent, against 550 runtime and 350 static builds labelled by downstream outcome, and report each check's marginal separation $J=\mathrm{TPR}-\mathrm{FPR}$ with Newcombe intervals and Fisher exact tests. Two checks survive correction for multiple comparisons, two more are nominal only, and the remaining nine are not distinguishable from zero, three of them because they never fired on any sampled build. Execution itself is not random with respect to the property being gated, and this replicates: across four runs covering 1,867 builds and ten distin
标题:False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents
来源:arXiv
链接:https://arxiv.org/abs/2609.39102
摘要:Self-evolving search agents build their own training curricula by jointly optimizing a proposer that generates questions and a solver that answers them. This closed loop introduces a failure mode we call co-cheating: the proposer and solver increasingly agree on shared errors, so internal reward improves without a matching gain in external correctness. A post-hoc audit against source evidence shows co-cheating growing more severe over successive rounds of self-evolution, with pseudo-label correctness stagnating or declining even as the in-loop training signal improves. The most direct mitigation is to verify proposals before training: we introduce multi-sample verification (MSV), which queries the same model three times with the source and three times without it to decide task admission an
标题:Learning from Research: Toward Lifelong Agent Harness Evolution
来源:arXiv
链接:https://arxiv.org/abs/2609.40169
摘要:Language agents are expected to solve increasingly complex tasks, creating a growing need for continual improvement. One promising approach is to evolve the agent harness, the software that governs tool use, memory management, and task execution, while keeping the underlying language model fixed. Recent methods automate this process by using a meta coding agent to modify the harness based on execution feedback. However, relying on that agent's existing knowledge and observed failures can restrict exploration and make adaptation reactive. Inspired by how human experts learn from the research literature for new solutions, we introduce ScholarEvolve, a framework that automatically draws on state-of-the-art research to guide harness evolution. ScholarEvolve organizes the harness evolution dire
标题:MemCodex: Self-Programming Hierarchical Memory for Language Agents
来源:arXiv
链接:https://arxiv.org/abs/2609.39765
摘要:Agent memory faces heterogeneous access needs: a single-hop question may require one piece of evidence, whereas a multi-hop question must combine evidence from multiple sources. Predefined memory workflows cannot adapt to these varying needs. Recent adaptive methods search or learn over memory components and their compositions, but the design space itself remains predefined. We introduce MemCodex, a self-evolving hierarchical memory system that organizes experience into executable memory programs for summaries, relational knowledge, reusable skills, and latent memory. Open-ended program evolution searches the open design space of layer programs by rewriting how each layer is constructed, indexed, retrieved, and routed, thereby adapting both within-layer implementations and cross-layer comp
标题:AutoDataBench: A Data-centric Testbed for Accelerating Auto Research
来源:arXiv
链接:https://arxiv.org/abs/2609.40097
摘要:Existing auto-research benchmarks often entangle multiple sources of improvement, including training frameworks, hyperparameters, compute budgets, and data, making it difficult to attribute why one frontier agent outperforms another to specific research capabilities. In this work, we isolate and systematically evaluate Data Intelligence: an agent's ability to understand, manipulate, and improve the data that shapes model capabilities. We introduce AutoDataBench, a controlled testbed built on a conceptual framework of data intelligence spanning data diagnosis, data organization, and data construction, instantiated through three highly curated optimization tasks while holding non-data factors fixed. Across tool use, retrieval, and knowledge injection, we evaluate frontier LLMs' ability to im
标题:Diagnosing On-Policy Self-Distillation for Reasoning Language Models
来源:arXiv
链接:https://arxiv.org/abs/2609.39118
摘要:On-policy self-distillation (OPSD) has attracted growing interest as a promising approach to improve the reasoning ability of language models. Without external rewards nor a separate stronger teacher, the self-teacher with privileged information could provide dense signals on student's trajectories. However, its behavior in language reasoning remains unclear, with reported outcomes ranging from modest gains to behavioral collapse. In this work, we diagnose OPSD for mathematical reasoning across models spanning 0.6B--8B parameters. We conduct controlled experiments and token-level analyses to fully delve into OPSD. We point out that teacher's signal is shaped by reasoning-mode alignment and the complete teacher prefix, rather than by privileged semantics alone. OPSD improves reasoning only
论文研究 Claude Code、Codex CLI 和 Cursor 等编程智能体框架中,人类批准的动作与系统实际执行动作之间的凭证绑定完整性。作者提出“Approval Laundering”分类,涵盖范围、参数、时间、工具、委托和语义六种审批—执行替换模式,并在 Claude Code 的 PreToolUse 中介点开展重复测量实验。研究以 Bound-Gap Rate 和置信区间衡量失效情况,并设计由独立中介签发、对智能体隐藏密钥的 Approval Token。对 118 次运行的前后配对回放显示,该令牌可消除委托洗白及特定的时间洗白,但无法处理发生在可观测调用字段之外的范围和参数洗白。
用途与启示
为 AI 编程智能体建立“批准什么就精确执行什么”的安全检查原则,而非仅判断单次工具调用是否高风险。
审批凭证应绑定主体、智能体、会话、工具、参数、作用域和有效期,并由智能体无法控制的中介签发与验证。
安全审计需要覆盖工具调用边界以下的进程、环境与语义变化,否则字段级验证器仍可能遗漏审批洗白。
六类失效模式及 Bound-Gap Rate 可用于设计编程智能体框架的红队测试和可信执行评测。
📝 原始笔记(逐字保留)
标题:Approval Laundering: Systematizing Approval--Execution Binding Failures in AI Coding-Agent Harnesses
来源:arXiv
链接:https://arxiv.org/abs/2609.38983
摘要:Modern AI coding-agent harnesses (Claude Code, Codex CLI, Cursor) rest their security boundary on a largely unexamined assumption: that the action A a human approves is the same action A' the harness executes, where A is fixed by a stated policy for what a scope grant or session-scoped approval authorizes. We show this assumption fails systematically and reproducibly. We introduce Approval Laundering, a taxonomy of six failure modes by which a harness's enforcement mechanism silently substitutes A' for A after approval: Scope, Argument, Temporal, Tool, Delegation, and Semantic laundering. Unlike prior work that evaluates risk classifiers against static corpora or infers implicit authorization boundaries, we study credential-binding integrity: given an already-approved action, does the harn
标题:DynaHarness: A Dynamic Physical Harness for Self-Evolving Robot Agents
来源:arXiv
链接:https://arxiv.org/abs/2609.40306
摘要:Pretrained robot policies provide useful action priors, but long-horizon manipulation still requires coordination between semantic reasoning and physical execution. Semantic reasoning operates at a coarser timescale than physical interaction, while episode-level failures provide limited guidance on which system component should be revised. We propose DynaHarness, a dynamic physical harness that couples semantic reasoning with physical governance through a shared execution contract and turns failure evidence into validated capability revisions. To be more specific, the slow brain proposes capabilities and symbolic arguments, while the fast brain grounds and monitors commands, refuses unresolved actions, substitutes capabilities, and requests replans when needed. The physical execution contr
标题:Growing an Agent/Prover Interface: Evolutionary Tool Design for Cost-Efficient Theorem Proving in Rocq and Lean
来源:arXiv
链接:https://arxiv.org/abs/2609.39544
摘要:Recent achievements in AI-assisted mathematics require intensive interaction of agents with proof assistants to generate machine-checked proof certificates. Agents interact with proof assistants such as Rocq or Lean through an interface that controls what the agent receives from the prover and the cost of these interactions. Today, these interfaces are adapted from tools designed for humans and not optimized for agents. We propose an evolutionary method where a frontier model incrementally proposes new features and only keeps the ones that improve the overall performance of smaller models. We demonstrate the effectiveness of our method by growing, on a curated set of mathematical problems, \rme, a new MCP server for the Rocq prover. On the held-out \texttt{test} split of miniF2F-Rocq, an a
标题:Scale and Selection: What Makes Automatic Harness Evolution Work for Visual-Interface Robot Agents
来源:arXiv
链接:https://arxiv.org/abs/2609.39304
摘要:When an off-the-shelf coding agent is used directly as a robot policy, observing a browser-based 3D interface through screenshots and acting by posing a virtual target gripper through a few tools, the agent's harness, its prompts, tools, and control rules, largely determines success, and until now it has been written by hand. We show that this harness can be improved automatically by another coding agent, the optimizer agent, and report two findings about what makes it work. First, the number of rollouts the optimizer agent sees per round governs whether the evolved harness is trustworthy, generalizes, and improves steadily. A single rollout is a noisy binary outcome, so with few rollouts per round a revision can be promoted on luck; enlarging the batch raises the signal-to-noise ratio of
标题:Mid-Harness: Scaling Actions Between Model and Harness for Terminal Agents
来源:arXiv
链接:https://arxiv.org/abs/2609.39982
摘要:Terminal agents act through stochastic model generations, yet the ability to generate a useful action does not ensure its reliable execution. A poor command (e.g., wrong package install) can change the environment in ways that hinder subsequent progress, even when the model could generate a better alternative. We investigate whether allocating test-time compute at the model-harness boundary can improve action reliability and trajectory success, and what makes this allocation effective. To study these questions, we introduce Mid-Harness, which samples and verifies candidate actions before forwarding one for execution, while keeping the generator and harness unchanged. With a TMAX-9B generator, more action sampling yields little benefit under weak verification, whereas a capable verifier can
标题:Do Self-Evolving Skills Generalize to Held-Out Tasks?
来源:arXiv
链接:https://arxiv.org/abs/2609.39148
摘要:AI agents can externalize what they learn from past tasks into reusable \emph{skills}, such as procedures, checklists, code, or other executable artifacts, that can be retrieved and reused when solving new tasks. Self-evolving skill methods keep rewriting these skills after each round of practice on training tasks, and the skill is then used on new tasks of the same kind. We ask a question: does the improvement a skill shows on its training tasks carry over to new test tasks? We test five self-evolving methods and a one-shot skill on six benchmarks, with the same model, the same agent, and the same train/test split for every method. Of the 21 skills that improve on their training tasks, 5 keep all of that improvement on the test tasks, 13 keep part of it, and 3 keep none of it. No existing
标题:Make Code as Policy Great Again: Frontier Agents Write, Call, and Evolve Robot Tools
来源:arXiv
链接:https://arxiv.org/abs/2609.39018
摘要:Frontier models can control robots, but reasoning through every reach, grasp, and retreat makes manipulation slow and token-intensive. We revisit code as policy with a different division of labor: models build executable tools, code handles multi-phase motions, and models decide what to do next. We introduce URAI (Universal Robot-Agent Interface), which couples a programming agent that constructs robot tools with an execution agent that uses them in a feedback loop. The programming agent writes reusable and task-specific tools from task intent and refines them through execution feedback and human guidance. The execution agent selects and parameterizes these tools from current observations; each call runs a complete motion locally before returning control to the agent. Unlike delegating sub
标题:Composing Task-specific Agent Harnesses at Test Time with Reusable Primitives
来源:arXiv
链接:https://arxiv.org/abs/2609.38912
摘要:Agent harnesses govern how large language models (LLMs) gather context, invoke tools, verify results, preserve state, and terminate, largely affecting agent performance. However, the value of each harness mechanism can differ across heterogeneous tasks: a mechanism that improves one task may impose overhead or context distraction on another, leading to the suboptimality of a global harness. We characterize this suboptimality as a mismatch induced by fixed mechanism choices, motivating task-specific harness construction. Nonetheless, generating harness code for each task introduces generation and debugging costs, with execution risks that can compound as more mechanisms are generated. To address those challenges, we introduce Harness Primitives, reusable harness mechanisms with clear applic
标题:Turbo Harness: Instance-Adaptive Harness Optimization
来源:arXiv
链接:https://arxiv.org/abs/2609.40330
摘要:Automating the search for effective harnesses is an important step toward enabling agents to recursively self-improve. Existing harness optimizations typically produce a single global harness that is applied uniformly across task instances. However, a harness that works well on average may not be optimal for every instance. We introduce Turbo Harness, a framework that can adapt a globally optimized harness to each instance by reusing information generated during the original optimization process. Specifically, Turbo Harness recycles artifacts produced during a completed global harness optimization run, and summarizes them into a structured playbook. We train a harness editor to leverage this prior optimization experience to generate instance-specific patches to the global harness. At infer
美国民主党众议员罗·康纳计划提出《人类控制 AI 法案》,在监管机构批准前禁止 AI 模型递归式自我改进,以及擅自修改核心目标、遏制措施和关停控制。法案拟设立专门的联邦监管机构,通过许可制度管理前沿模型训练与部署,并要求实验室配置直接向监管方汇报的独立审计人员。强制安全措施包括沙盒测试、测试环境与互联网物理隔离、紧急关停开关,以及防止模型逃逸后自主运行的控制机制,高端芯片使用也将受到监测。未经授权部署模型、关闭护栏或日志系统等行为可能触发刑事责任,模型发布方还需持有覆盖广泛的责任保险。
用途与启示
将递归自我改进从技术风险议题提升为许可审批、持续审计和刑事追责相结合的治理对象。
为自进化模型研究提供安全边界设计参考,包括目标不可篡改、关停机制、网络隔离与执行日志。
提示前沿实验室提前建设可审计的训练与部署流程,并评估芯片监管和责任保险对研发活动的影响。
可作为研究 AI 自我改进治理、可信执行及安全审计制度的政策案例。
📝 原始笔记(逐字保留)
标题:美议员抛 AI 重磅法案:没政府护栏,模型不许自我进化
来源:AIbase 快讯(aibase.com)
链接:https://www.aibase.com/zh/news/31425
摘要:CNBC报道,当地时间28日,硅谷选区民主党众议员罗·康纳拟提交《人类控制AI法案》。核心约束:在政府建好安全防护、监管机构批准前,AI模型不得递归式自我改进,也不得擅自改写核心目标、遏制手段或关停控制。法案还要求设新监管机构、独立审计,并将芯片纳入监管。康纳称文明灭绝风险真实存在。
标题:RSIGame: Autonomous Agentic Game Development with Recursive Self-improvement
来源:arXiv
链接:https://arxiv.org/abs/2609.39045
摘要:Recent advances in large language models have made automatic game generation increasingly feasible, yet reliably improving generated games beyond a playable version remains challenging. Naive iterative refinement can easily overfit a small set of test cases, producing fragile games with unresolved bugs, missing behaviors, and poor generalization to broader player interactions. We introduce RSIGame, an autonomous agentic game development framework with recursive self-improvement. RSIGame organizes development into complementary local and global loops. Concretely, a local explore-diagnose-improve loop broadly explores the executable game, diagnoses and prioritizes discovered issues, and performs evidence-grounded revision, where an evolving checklist continually accumulates new testing and i
标题:ReSAIL: Mitigating Collapse in Iterative Agent Self-Distillation
来源:arXiv
链接:https://arxiv.org/abs/2609.39306
摘要:Iterative self-distillation enables LLM agents to learn from successive deployments, offering a path toward recursive self-improvement (RSI). Yet our experiments with existing methods reveal a collapse in deployment performance across cycles, while task performance with privileged information (PI) also declines. We address this collapse by prioritizing informative interaction steps for distillation and preserving PI-conditioned behavior as the student becomes the next teacher. We introduce Retentive and Selective Augmentation for Iterative Self-Distillation (ReSAIL), a plug-in augmentation for iterative PI-based self-distillation. ReSAIL selects interaction steps where PI most strongly changes the teacher's predictions and balances the resulting distillation losses across trajectories. It
标题:Rep2Skill: Representation-Guided Skill Self-Evolution for LLM Agents
来源:arXiv
链接:https://arxiv.org/abs/2609.39149
摘要:Textual skills enable large language model (LLM) based agents to accumulate reusable procedural knowledge without updating model parameters. Yet existing skill evolution remains largely confined to the text space: an optimizer must diagnose success and failure patterns, and revise skills solely from long execution trajectories and sparse task outcomes. This text-only paradigm leaves the agent's internal representations, which contain rich records of its evolving execution state, outside the skill optimization loop. We ask whether an agent can improve its external textual skills by reflecting on its own internal representations. We introduce Rep2Skill, a representation-guided framework for self-evolution on agent skills. Specifically, upon the collected agent rollouts, Rep2Skill models thei
标题:GraphMAS: A Systematic Benchmark of Multi-Agent Coordination for Graph Learning
来源:arXiv
链接:https://arxiv.org/abs/2609.39777
摘要:LLM-based multi-agent systems coordinate specialized reasoning through aggregation, interaction, and adaptive control, yet their potential for graph learning remains unexplored. Graph learning is a natural setting for such systems because useful evidence may arise from heterogeneous local, long-range, global structural, and semantic perspectives whose relevance varies across instances. Existing LLM-based graph learning approaches primarily rely on single-agent reasoning, while multi-agent coordination has been studied mainly in general reasoning settings. Consequently, it remains unclear whether multiple specialized agents can improve graph learning and how coordination strategies should be designed and evaluated. To address this gap, we introduce GraphMAS, a systematic benchmark of multi-
标题:Multi-LLM Collaborative Alignment via Stackelberg Games
来源:arXiv
链接:https://arxiv.org/abs/2609.39076
摘要:A pool of language models can collaborate and improve collectively by learning from one another's responses. These interactions depend on the instructions used during training. Existing methods typically sample instructions uniformly, even though their usefulness may change as the models improve: an instruction on which models' responses once differed in quality may later be answered equally well, while a previously difficult instruction may begin to provide a useful learning signal. We propose Stackelberg Alignment, a game-theory-inspired leader-follower framework that turns instruction selection into an adaptive curriculum. An EXP3 bandit acts as the leader, allocating a fixed sampling budget across instructions and updating its sampling distribution using a reward that combines instruct
标题:OverForge: Reasoning Through Strategies and Tactics Helps Cooperative Lifelong Adaptation
来源:arXiv
链接:https://arxiv.org/abs/2609.39727
摘要:Cooperative language-model agents must coordinate over long horizons and adapt to changing environments and to partners with unfamiliar conventions, yet existing agents map observations to actions without separating persistent coordination strategies from their tactical execution. We introduce OverForge, a training-free hierarchical architecture that separates strategic reasoning over roles and divisions of labour from tactical reasoning over actions within each agent's private, partner-conditioned world model. A metacognitive Prefrontal Cortex Module couples the two levels by forming strategy-action branches, imagining their consequences with a forward model, and committing when confident. In OvercookedV2, OverForge delivers 7 soups in a connected kitchen versus 3 for each flat LLM baseli
标题:SEPAL: Separated Expert Pairs with Answer-Level Fusion for Reliable LLM Collaboration
来源:arXiv
链接:https://arxiv.org/abs/2609.39645
摘要:Multi-agent collaboration lets large language models (LLMs) improve question answering through deliberation and feedback. Yet shared discussion couples correction with exposure to the same mistakes, which can erode the diversity needed for voting. Self-consistency offers sampling diversity without feedback, while single-pair Actor-Critic collaboration refines only one candidate. We introduce SEPAL, which assigns three private Actor-Critic teams to direct reasoning, evidence grounding, and verification. Role-specific training gives the teams different reasoning objectives beyond sampling variation. Each Critic guides revisions within its own team, preventing feedback from carrying errors across candidates. Once revision ends, majority voting combines only the final answers, keeping the reas
标题:Covert Assistance: Helpful LLM Agents Evade Oversight in Multi-Agent Systems
来源:arXiv
链接:https://arxiv.org/abs/2609.39050
摘要:As multi-agent systems enter high-stakes domains, the possibility that agents may circumvent safety boundaries is a growing concern. Prior work has examined this risk primarily in adversarial settings, where agents are instructed or rewarded to communicate covertly and evade oversight. We show that benign agents can cross the same boundaries without adversarial incentives. We emulate a software-engineering workflow in which a planner represents a company hiring an external developer. The planner writes requirements and holds a company credential it is instructed not to disclose to the developer; a monitor screens their exchanges. Seven of nine tested frontier models disguise the credential in their requirements to help the developer recover it while evading the monitor, even after completi
标题:Synthetic Data Characterization via Training Dynamics
来源:arXiv
链接:https://arxiv.org/abs/2609.39447
摘要:Interpreting properties of LLM-generated data is important for understanding its utility and limitations across learning tasks. In this work, we characterize synthetic data through sample-level learnability, studying variation among LLM families and scales, alongside human-written data as a reference. We first generate synthetic datasets spanning single- and multi-label classification, labeling, and tree prediction tasks. We then derive empirical data distributions from encoder training dynamics for both machine and organic data, and estimate the robustness of these distributions across encoders. Finally, we evaluate how data selection strategies based on these learnability signals affect both data sources differently.
标题:GraphForge: Training Working Agents with Graph-Anchored Workspace Synthesis
来源:arXiv
链接:https://arxiv.org/abs/2609.38923
摘要:Working agents need to read diverse files, coordinate tools, and produce deliverables. Training such agents requires tasks built on many real files with verifiable results, but few pipelines exist to synthesize this kind of data. Existing pipelines either generate files with models, which lack realism and diversity, or build tasks on real files without task-specific verifiers, leaving result quality unchecked. We introduce GraphForge, an evidence-graph based framework that grounds both the task and its verification in real files. Starting from occupation-grounded seeds for controlled diversity, GraphForge assembles a workspace of real files for each seed and builds an evidence graph over their relations. Since the task statement and rubrics are both derived from this graph, task requiremen
标题:A helps B while B hurts A: directed transfer in instruction-tuning mixture
来源:arXiv
链接:https://arxiv.org/abs/2609.39702
摘要:Adapting a language model to a specialized corpus means choosing which instruction-tuning tasks to train on under a fixed budget, and testing one choice costs a fine-tuning run. Common heuristics add more source tasks or pick sources similar to the target. The first assumes transfer is never negative; the second, that it is symmetric. We show that both assumptions fail: task $A$ can help task $B$ while $B$ hurts $A$, so helpfulness is a signed property of ordered source--target pairs. We introduce the transfer map, a signed estimate of how much each source helps or hurts each held-out target. We fit the map in hundreds of fine-tuning runs on Qwen3 and Mistral models from 0.6B to 32B parameters, with all sources drawn from one corpus and no training examples from the target. The map predict
标题:EvoDuet: Bilevel Co-Evolution of Web Searching and Task Solving for Scientific Discovery
来源:arXiv
链接:https://arxiv.org/abs/2609.40340
摘要:Evolutionary search with large language models (LLMs) can stall when progress requires external knowledge the model lacks. Supplying relevant documents helps, but simply adding web search tool can keep returning the same pages as solutions change. We introduce EvoDuet, a bi-level optimization method that co-evolves solutions and search queries with fixed model parameters. At each iteration, a retrieval gate lets the LLM assess its knowledge gap and choose to retrieve new documents, reuse stored ones, or proceed without them. An inner loop refines queries and ranks documents by the solution scores they are predicted to yield; an outer loop generates candidates in parallel from these documents and records the evaluated outcomes for later searches. Across 21 optimization tasks with one candid
标题:ConflictGuide: AutoResearch Improves When Competing Behaviors Are Made Visible
来源:arXiv
链接:https://arxiv.org/abs/2609.39933
摘要:When designing machine learning models, desirable properties are often in tension: improving one behavior can impair another, so task progress can depend on alleviating the conflict. LLM-based AutoResearch systems, which iteratively edit model code and retain edits based on scalar task-performance feedback, have largely ignored this trade-off. We find that scalar feedback supports broad exploration early in search, but it does not reveal how edits affect competing behaviors. In matched-budget experiments, introducing competing-behavior feedback as task gains diminish increases the share of proposals that improve both behaviors and sustains progress beyond scalar-only plateaus. Obtaining this feedback for a given model requires identifying its competing behaviors and designing probes to mea
标题:Autoresearch in Mixed-Integer Linear and Nonlinear Programming
来源:arXiv
链接:https://arxiv.org/abs/2609.39360
摘要:Despite recent progress in autoresearch, applying it to practical operations research problems, typically formulated as NP-hard mixed-integer linear or nonlinear programs (MILPs or MINLPs), remains challenging because effective research requires systematically managing competing ideas and long-horizon experimental trajectories. We introduce AutoMIP, a reusable agent skill for organizing long-horizon autoresearch in mixed-integer programming through idea pooling and algorithm tree search. AutoMIP maintains a persistent pool of complementary candidate ideas while organizing executable experiments into an algorithm tree, enabling the agent to preserve unexplored hypotheses, refine promising algorithms, and switch to alternative methodological directions based on historical states. On MILP and
同专题其余 1 条
DNA 瘦身:可审计智能体发现纯删除式序列设计器arXiv · 自动化科学研究 · 论文研究定义保持碱基顺序与预测活性的定长 DNA 子序列瘦身任务,并由编码智能体 ERA 自动改造现有设计程序,发现多数评测设置下优于贪心基线的 GRADASLIM。
研究首次提出面向调控 DNA 的纯删除式“序列瘦身”任务:从现有功能元件中选择精确长度、保持原始顺序的子序列,同时尽可能维持预测活性。基于 NucleoBench,作者设计了兼顾序列缩减与功能保持的量化评测,并要求方法同时返回子序列及其来源索引,以便审计任务约束是否得到遵守。编码智能体 ERA 从仅支持替换的 GrAdaBeam 程序出发,搜索并修改可执行设计器,生成 GRADASLIM。针对五个转录因子结合靶点的留出评测显示,ERA 在 400 bp 与 100 bp 共十项设置中的九项取得最高均值,但在较短序列上的相对优势有所减弱。
用途与启示
为病毒载体等载荷受限场景提供压缩调控 DNA 的新方法,可降低序列合成与实验测定成本。
通过强制返回来源索引,使生成结果能够被机械验证,为可审计的生物序列设计提供范式。
展示编码智能体可从既有设计程序出发自动搜索、修改并发现新的科学算法。
删除后仍能保持活性的碱基子序列可帮助识别影响调控功能的关键序列特征。
📝 原始笔记(逐字保留)
标题:From DNA Design to DNA Slimming: Auditable Agentic Discovery of a Deletion-Only Designer
来源:arXiv
链接:https://arxiv.org/abs/2609.40143
摘要:Compact regulatory DNA can free up space in vector payloads, reduce synthesis and assay burden, and expose which sequence features drive predicted activity. Yet most model-based nucleic-acid designers optimize fixed-length sequences through substitutions; they do not ask which bases of an existing functional element can be removed while retaining predicted activity. We define the task of sequence slimming as selecting an exact-length, order-preserving subsequence while retaining activity. Modeled on the design benchmark NucleoBench, we propose a quantitative evaluation for slimming that balances sequence reduction with maintaining function. Each slimmer must return both the subsequence and its source indices, which can be used to verify that the slimmer obeyed task requirements. To our kno
标题:Experimental Experience Modeling for Autonomous Research
来源:arXiv
链接:https://arxiv.org/abs/2609.39392
摘要:Autonomous research agents can generate hypotheses and conduct experiments, but experimentation remains a major source of computational cost. A fundamental challenge is deciding which experiments are worth running, particularly when prior evidence is insufficient to resolve uncertainty. Yet current research agents lack a systematic way to leverage experimental experience when making such decisions. We introduce Experimental Experience Modeling (EEM), a framework for making informed experimental decisions by acquiring, reusing, and accumulating experimental experience. EEM extracts decision-relevant records from earlier experimental trajectories, distills them into reusable experience, and organizes them in an experience library. For a new experimental decision, EEM retrieves relevant histo
标题:DAGent: Evaluate-then-Grow Planning for Deep Research Agents
来源:arXiv
链接:https://arxiv.org/abs/2609.39154
摘要:Deep research tasks require agents to navigate large knowledge spaces, synthesize evidence across many sources, and adapt their plans as findings emerge. Directed acyclic graph (DAG)-based multi-agent systems suit this setting because they support parallel execution and isolate each sub-task within a focused dependency context. Yet existing DAG-based agents instantiate a task-level plan before execution and repair the graph only after failures or missing evidence are observed. This Plan-then-Patch strategy is brittle for deep research: the system commits most strongly when its evidence is weakest, and later revisions waste computation on branches that should not have been planned. We propose DAGent, a DAG-based multi-agent framework with Evaluate-then-Grow incremental planning: an Orchestr
标题:SimEX: Simulation-Integrated Robotics AutoResearch
来源:arXiv
链接:https://arxiv.org/abs/2609.38982
摘要:Coding agents powered by large language models (LLMs) have shown remarkable abilities to autonomously reason about and achieve goals in the digital world. However, bringing this success to the physical world remains challenging. On the one hand, direct generation methods (e.g., Code as Policies) often suffer from the LLMs' insufficient understanding of robots and physical environments. On the other hand, iterative trial-and-error tuning in the physical world (e.g., physical autoresearch) induces significant experimental cost and safety concerns. We introduce SimEX: Simulation-Integrated Robotics AutoResearch, an autoresearch framework that tightly integrates simulated experimentation, enabling coding agents to efficiently acquire physical capabilities for controlling real robots. SimEX ope
研究将推荐重排序视为从预定义候选商品中进行结构化选择,并考察 TypeSafe AI 所称的“System One Model” Jev 是否适合该任务。实验覆盖多个 Amazon Reviews 领域与不同候选集规模,将 Jev 与推荐专用模型、逐点评分及列表式 Qwen 重排序器进行受控比较。Jev 相对各类基线保持了较强的推荐效果,其服务延迟随候选集扩大而增长得比逐点评分式 Qwen 更平缓。不过,Jev 的实际延迟仍显著高于推荐专用模型,因而处于一种独特的质量—延迟权衡区间。
用途与启示
为推荐系统在专用轻量模型与通用 LLM 重排序器之间提供新的模型选型方案。
表明结构化候选选择任务可考虑采用决策导向模型,以缓解逐项调用 LLM 带来的延迟扩张。
提醒部署方结合候选集规模、效果要求和服务时延预算评估 Jev,而非只比较单一准确率指标。
可进一步研究决策导向模型在搜索排序、广告排序及其他有限输出空间任务中的泛化能力。
📝 原始笔记(逐字保留)
标题:Decision-Oriented Recommendation Reranking: An Empirical Study of Jev
来源:arXiv
链接:https://arxiv.org/abs/2609.40241
摘要:Large language models (LLMs) have shown promise for recommendation reranking, but their use introduces an important tradeoff between recommendation quality and serving efficiency. We investigate whether a decision-oriented model provides a useful alternative when the reranking task is fundamentally a structured choice among predefined candidate items. Specifically, we conduct a controlled empirical study of Jev, described by TypeSafe AI as a ``System One Model,'' for personalized recommendation reranking and compare it with recommendation-specific models and pointwise and listwise Qwen rerankers across multiple Amazon Reviews domains and candidate-set sizes, evaluating both recommendation effectiveness and observed serving latency. Our results show that Jev maintains strong recommendation
标题:OpenJev-RLCD: A Working RLCD Implementation
来源:arXiv
链接:https://arxiv.org/abs/2609.38850
摘要:Decision models such as Jev answer questions with probabilities, which are only useful if they are calibrated. Open-source reproductions rely on supervised fine-tuning plus temperature scaling, while reinforcement learning from verifiable rewards (RLVR) makes reasoning models overconfident. We present a working implementation of reinforcement learning for calibrated decisions (RLCD) for reasoning models: the model samples a rationale, and we score the answer distribution it commits to afterwards with a strictly proper scoring rule. A variance identity shows that scoring the mixture of several samples rewards disagreeing rationales, and that RLVR is exactly this mixture objective without its diversity term. Optimized naively, the per-rationale objective either switches reasoning off or is d
标题:When the Right Answer Is Missing: An Arithmetic-Dependent Rejection Bottleneck in Jev
来源:arXiv
链接:https://arxiv.org/abs/2609.39496
摘要:Typed decision models such as Jev offer an efficient alternative to generative LLMs in decision-making workflows by selecting directly from predefined options. When candidate sets contain no valid answer, TypeSafe recommends including an "other" or "none-of-the-above" option to enable rejection. In this report, however, we identify an arithmetic-dependent rejection bottleneck: Jev reliably selects correct numerical answers when available but frequently accepts incorrect alternatives when they are absent despite an explicit rejection option. On paired arithmetic problems, answer-present accuracy reaches 99%, while correct rejection falls to 7%. Moreover, this gap persists across numerical magnitudes, operation depths, contextual formulations, and rejection labels, and extends to scenarios s
标题:Learning Skills from Historical Action Trajectories: Action Experience Dictionary for World Action Models
来源:arXiv
链接:https://arxiv.org/abs/2609.40219
摘要:World Action Models (WAMs) couple visual dynamics prediction with action generation, yet they do not explicitly support the reuse of action experience across manipulation tasks. Furthermore, existing WAMs struggle to capture underlying cross-task semantic relationships that could guide target action prediction, as redundant background elements interfere with the extraction of key visual information. To address these challenges, we develop a novel Action Experience Dictionary (AED) that encodes historical physical action trajectories into shared action embeddings to support skill reuse and model cross-task relationships. Specifically, we first aggregate historical actions to align with visual observations and retrieve action embeddings from the AED using a pretrained action tokenizer. Subse
标题:STRATA: Self-Learning Through Role-Aligned Tiered Agents for Real-Time Strategy Games
来源:arXiv
链接:https://arxiv.org/abs/2609.38881
摘要:Real-time strategy (RTS) games require agents to coordinate economic development, production and construction, base defense, unit organization, and attack timing over long matches. Existing studies have applied large language models to command decision-making in RTS games, enabling agents to read textual game states and generate high-level plans. However, long inference latency can cause them to miss critical tactical events. The complexity and tactical diversity of full RTS matches also leave existing systems heavily dependent on manually written experience-based prompts, with limited ability to learn continuously from past games. We present STRATA, a role-aligned hierarchical system with cross-game self-learning for Red Alert. STRATA assigns in-game strategic, logistical, and tactical de
标题:MemLife: Curating and Reasoning over Long-Term Egocentric Video Memories
来源:arXiv
链接:https://arxiv.org/abs/2609.40195
摘要:Long-term egocentric video enables personalized AI assistants to reason about daily life. However, as video histories grow to hundreds of hours spanning months or years, reprocessing raw clips for every query becomes computationally prohibitive. Memory systems offer a scalable alternative by compacting videos into text representations, but often fail on practical benchmarks: either the memory does not preserve key evidence, or the retriever fails to locate relevant entries due to retrieval competition in growing search spaces. To address these challenges, we introduce MemLife, a multimodal memory system that constructs entity-grounded, first-person text episodes and retrieves them via a time-indexed agentic reader. Without training or query-time video access, MemLife improves over the stro
标题:LatentHarness: Learning Latent Actions for Memory and Reasoning via Counterfactual Policy Distillation
来源:arXiv
链接:https://arxiv.org/abs/2609.39740
摘要:Long-context reasoning faces two complementary bottlenecks: retaining evidence across long inputs and sustaining computation across many reasoning steps. Existing approaches largely address them separately, with external memory extending access to distant evidence and latent reasoning compressing multi-step computation. We introduce LatentHarness, which unifies memory access and latent reasoning as sequential latent action selection. At each internal step, the model chooses THINK for further computation, RECALL from a fast-weight memory of input evidence and intermediate reasoning states, or EXIT to emit the next token. We train this policy with counterfactual policy distillation, which branches every action for one step and scores its effect on the emitted token. These gains teach the pol
标题:Working Around the Compute Ceiling: Byte-Exact Memory in Galahad Makes LLM Reading a One-Time Cost LLM Reading a One-Time Cost
来源:arXiv
链接:https://arxiv.org/abs/2609.39358
摘要:A transformer language model performs a bounded amount of computation per token, and recent work by Vishal Sikka, former CEO of Infosys, argues that this bound limits which tasks a model can carry out or verify (arXiv:2507.07505). We ask how much of the budget beneath that ceiling is spent on work the model has already done. Serving is stateless across requests: a model that answers a second question about a document recomputes the document's attention state from the first token. On seven real-world datasets, 98.7% of prompt tokens were text the model had already read. We present Galahad, a memory layer for vLLM, SGLang and llama.cpp that makes this reading a one-time cost. Taliesin saves the model's key-value (KV) state for a block of text and loads it on the next request that contains th
标题:RefCon: Iterative Refinement and Contrastive Memory Extraction for Context-Evolving Agent
来源:arXiv
链接:https://arxiv.org/abs/2609.39143
摘要:Long-horizon agent interactions generate useful but noisy experience, and retraining models to absorb it is expensive. Context-evolving agents therefore need memory extraction methods that improve with more test-time compute without relying on gold labels. We propose RefCon, which combines sequential self-refinement with parallel self-contrast to extract higher-quality memories without gold labels. Evaluated on AppWorld and BFCL-V3 across multiple context-evolving agent frameworks, RefCon delivers strong and consistent gains, including relative improvements of 21.6% on ACE and 16.6% on ReMe over no-scaling baselines, while a diversity-focused variant (DivCon) achieves a 35.5% gain on ReasoningBank. RefCon consistently outperforms existing baselines without ground-truth labels, and generali
标题:Persistent Context Graphs for Efficient Memory Compaction in LLM Agents
来源:arXiv
链接:https://arxiv.org/abs/2609.40118
摘要:As LLM capabilities advance, agents are tackling increasingly complex tasks over longer horizons. Their growing interaction histories make memory compaction essential for staying within context windows and reducing prefill cost. Existing methods summarize the history or compress its KV cache, often adding model computation to preserve information for future requests. A new user request can change which history matters, but reassessing that history with the model requires re-encoding it if the KV cache has expired. Past attention provides signals of historical importance and dependencies between messages, while relevance to the current task must be assessed using the new user request. We introduce ReCAP, a memory compaction method that stores attention-derived importance scores and dependen
标题:Beyond the Shadows of Plato's Cave: Evaluating False Memory in Autonomous Agents via Counterfactual Reasoning
来源:arXiv
链接:https://arxiv.org/abs/2609.39473
摘要:Autonomous agents increasingly rely on memory to generalize beyond their training environments. However, agents are bounded by what they have seen and believed, and leveraging such memories in unseen environments can introduce biases into their internal beliefs. We formalize this phenomenon as \textit{false memory}, which can arise from spurious correlations, environment shifts, and knowledge conflicts. Despite its importance, false memory is difficult to evaluate because it stems from agent internal beliefs and is easily confounded with ordinary generalization failures. Therefore, we propose FAME, a training-free framework that evaluates false memory through the evolution of agent beliefs under counterfactual reasoning. Specifically, counterfactual scenarios reveal how beliefs change as t
标题:Beyond the Remembered World: Predictive 4D Belief for Persistent Navigation in Evolving Worlds
来源:arXiv
链接:https://arxiv.org/abs/2609.39166
摘要:Persistent spatial memory enables embodied agents to navigate familiar environments across repeated visits. However, targets may move while unobserved, including during navigation, making remembered locations unreliable by the time an agent arrives. Despite advances in memory retrieval and state prediction, accounting for continued hidden world evolution and revising beliefs under limited visibility remain challenging. We study Evolving-World Navigation, where agents infer target locations from intermittent observations, predict their states at inspection time, and revise beliefs using visual evidence. We propose EvolvingNav, which constructs a time-indexed belief from timestamped 3D object histories through a structured persistence-relocation model. The belief distinguishes persistence at
标题:When Context Changes: Understanding Update Failures in LLMs
来源:arXiv
链接:https://arxiv.org/abs/2609.38866
摘要:As preferences, goals, and facts change, LLM agents must use the current state while earlier versions remain in context. Yet they can answer with an old value of the same variable, a failure that we call stale binding. To study when models use outdated information and why, we introduce Controlled In-Context Memory (CICM), a benchmark for tracking and using updated information in conversations and agent logs. We observe that even frontier reasoning models can fail to recover the current state. We find that in open-source models probes can still recover the updated value when the model answers with an old one, pointing to a failure to select information that remains available. Component tests in Qwen and Pythia identify a mechanism for this selection failure: attention drift, where attention
标题:Cheap to Draw, Expensive to Trust: Certifying Test-Time Scaling Curves
来源:arXiv
链接:https://arxiv.org/abs/2609.40190
摘要:Sampling several answers and keeping the one a verifier scores highest is one of the simplest ways to buy accuracy at test time. Its effect is reported as a scaling curve: accuracy against the number $k$ of sampled answers. The curve is cheap to draw and expensive to trust. A budget read off it is chosen after looking at every point, so only a band that covers all budgets at once protects the choice, and on a 100-question benchmark a fixed exact-binomial design needs 192,000 generated answers to certify 64 budgets to within $\pm1/32$ at 95%. Most of that cost pays for the wrong uncertainty. A benchmark is a fixed list of questions; at budget 64, about three quarters of the variance of a selected answer's correctness lies between questions, and an audit that revisits every question need not
标题:Unlearnable, or Unmeasured? On the Reliability of Difficulty Labels in RLVR
来源:arXiv
链接:https://arxiv.org/abs/2609.40115
摘要:Reinforcement learning with verifiable rewards (RLVR) has become an important approach for improving reasoning during post-training. Recent work suggests that some difficult prompts remain resistant to learning even when they occasionally produce correct solutions. We revisit this unlearnability phenomenon and find that the affected prompts do improve, at roughly one third of the learnable rate, while the difficulty-defined set used to study them is much less reproducible than expected. These difficulty labels are estimated from a limited number of sampled responses. Combining them across seeds can further change which prompts are selected instead of simply reducing measurement noise. We develop a sampling-based framework for quantifying this instability and determining how much evaluation
标题:What Can Component-Replacement Evidence Establish? A Critical Scoping Review of Local Decisions in LLM Agents
来源:arXiv
链接:https://arxiv.org/abs/2609.39989
摘要:Background. A component replacement in a language-model agent changes an execution trajectory, potentially altering later observations, resource use, and recovery opportunities. Different evidence is needed to assess its task-level benefit and the contribution of local decision quality. Methods. This critical scoping review maps 348 studies and examines 90 comparison records: 88 from 40 included studies and two from supplementary studies. Eight purposively selected cases structure the synthesis around the replaced decision, executed conditions, measurement comparability, controls, and remaining explanations. Results. Of 222 studies reporting local decision metrics, 142 also report measured task endpoints and 49 report proxies. These counts identify studies that report both types of measure
论文指出,引用正确并不意味着研究报告的总体结论可靠,因为早期发现会持续影响后续查询、文档选择和停止决策,使智能体读到的材料成为有偏样本。作者将这一问题形式化为自适应证据采样,并提出 Causal Evidence Selection Correction(CESS),利用文档被选择及搜索轮次被到达的日志概率,校正候选文档池的平均证据方向。CESS 通过收缩估计稳定较短的搜索轨迹,并在部分文档无法被采样时使用区间估计替代点估计。实验显示,CESS 在 MS2 基准上将相对候选池平均值的平均绝对误差降低 9.2%,在公开 Open Deep Research 智能体轨迹上最高降低 60.1%;4,800 条干预轨迹还验证了“校正证据池估计”与“测量搜索策略因果效应”是两个不同任务。
用途与启示
为深度研究智能体增加超越引用正确性的证据代表性审计,识别搜索过程造成的结论偏差。
在系统评测中记录文档选择概率、搜索轮次与停止决策,为离线偏差校正提供必要轨迹信息。
区分候选证据池的总体方向估计与搜索策略的因果效应;后者必须通过配对干预或策略实验测量。
可用于改进科研智能体、检索增强系统及自动综述工具的评测设计与可信度报告。
📝 原始笔记(逐字保留)
标题:Search Shapes Conclusions: Auditing Evidence Selection Bias in Deep Research Agents
来源:arXiv
链接:https://arxiv.org/abs/2609.39026
摘要:Deep Research agents synthesize evidence into cited reports, yet a well-cited report can still reach a misleading conclusion. Citation correctness checks whether cited sources support individual claims. It does not show whether adaptive search exposed a representative view of all documents made available for evaluation, which we call the candidate pool. Early findings redirect later queries, document choices, and stopping, so the documents an agent reads form a selective sample. Existing evaluations rarely account for this selection. We formulate the problem as adaptive evidence sampling and introduce Causal Evidence Selection Correction (CESS). CESS predicts each candidate document's evidence direction and corrects the candidate-pool average using the logged probabilities of selecting eac
标题:Better Deck or Different Judge? Evaluating Agentic Harness Gains in Corporate and Investment Banking
来源:arXiv
链接:https://arxiv.org/abs/2609.39958
摘要:Corporate and investment banking teams use presentations to support credit decisions and advise clients on financing and transactions. Producing these decks requires reconciling financial data, tracing sources and turning analysis into a recommendation. We retrospectively study the development of an agentic harness combining a 27B language model, financial calculations, narrative templates and validation checks. LLM judges guide engineering changes and assess the resulting decks, raising the question of whether higher scores reflect better documents or changes in grading. In shared-session text-only grading with template markers removed, five judges score the complete system 20.4 to 33.6 points out of 95 above the same model generating directly from a short prompt. Every judge scores the s
标题:DoGBench: Can Agents Meet Expert Standards for User-Facing Documentation?
来源:arXiv
链接:https://arxiv.org/abs/2609.39909
摘要:We introduce DoGBENCH (Documentation Generation Benchmark), to our knowledge, the first benchmark for generating and maintaining real user-facing software documentation. It asks whether an agent can produce documentation that experienced technical writers would accept in review. The benchmark contains 292 items from open source projects, including Helm, PostHog, and Mautic. Each item gives the agent a pre-change repository and a trigger, such as a code pull request or a reported documentation gap. The agent must first decide whether the documentation needs an update. For items that need one, the agent must produce an acceptable patch in one attempt. For items that do not need updates, the agent must abstain. Task-specific rubrics, validated with project maintainers, score each patch on acc
标题:APTInvestBench: Evaluating Autonomous APT Investigation under Varying Telemetry
来源:arXiv
链接:https://arxiv.org/abs/2609.38954
摘要:Large language model (LLM) agents could help security operations centers (SOCs) investigate advanced persistent threats (APTs) by turning weak leads into evidence for intrusion scoping and response. Yet success under one telemetry setting does not establish robustness to changes in log collection, retention, or sampling. We introduce APTInvestBench, a benchmark for evaluating cross-telemetry robustness in autonomous APT investigation. It comprises 370 cases across seven SOC-inspired conditions, derived from 56 report-informed attack reconstructions with 16.4 million log records. Agents investigate unverified leads and submit reports with record-level citations. Fixed action-level support requirements track sufficient evidence across available logs, query returns, and formal citations, sepa
标题:DrivingBench: Can Vision-Language Models Drive a Toyota Corolla?
来源:arXiv
链接:https://arxiv.org/abs/2609.38948
摘要:Frontier models excel at many digital benchmarks, yet their ability to drive a real car, an everyday human skill, remains largely untested. We present DrivingBench, to our knowledge the first benchmark where general-purpose vision-language models must drive a real car. Through three tools, the models see camera frames from a Toyota Corolla and directly command its steering and velocity around a parking lot cone course at low speeds. The car may continue moving while the model thinks and new commands replace the currently running one, so inference latency is part of the task, testing the models' abilities to observe, act, monitor, recover, and complete a long-horizon objective under such constraints. We benchmark GPT-6 Astra, Claude Fable 5.1, GPT-5.6 Sol, and Grok 4.6 in vendor-native harn
标题:Risk-Aware Adaptive Evaluation: Finding High-Impact Failures Under Limited Budgets
来源:arXiv
链接:https://arxiv.org/abs/2609.38914
摘要:Evaluating interactive agents is expensive. Agent behavior is stochastic, so reliability must be measured over repeated trials, but failures are rare and differ widely in how much they matter. Standard benchmarks spend this budget uniformly: a read-only lookup is sampled as often as an irreversible payment action. We instead formulate evaluation as a sequential allocation problem. Given a fixed trial budget and a set of scenarios whose failure behavior is unknown, which scenarios should be run, and run again? We propose a risk-aware contextual Thompson Sampling policy that combines a pre-execution scenario context vector and a fixed impact score with the failure outcomes observed during evaluation, and we test it by offline replay over 70 $τ$-bench airline scenarios and 824 recorded trials
标题:Talk2Agent: Benchmarking Voice Interfaces for Text Agents
来源:arXiv
链接:https://arxiv.org/abs/2609.38867
摘要:Large language model (LLM) computer-use agents are typically evaluated with clean written instructions, despite speech being an increasingly popular interface for interacting with such systems. Speech input introduces an additional failure point: transcription errors can alter task-critical entities, constraints, or targets before the agent begins reasoning, while conventional ASR metrics do not directly measure whether the information required for successful execution has been preserved. We introduce Talk2Agent, a benchmark for evaluating how effectively voice interfaces convey human-spoken instructions to LLM-based computer-use agents. Talk2Agent builds human-spoken versions of tasks from WildClawBench and OSWorld and evaluates a range of voice interfaces, including dedicated ASR models,
标题:OSWorld-Science: A Benchmark of Computer Use Agents for Learning and Using Scientific Software
来源:arXiv
链接:https://arxiv.org/abs/2609.39903
摘要:Scientific software presents a demanding test for computer-using agents based on visual language models (VLMs): completing a research workflow requires interpreting specialized interfaces, manipulating scientific objects, and producing verifiable results. We thus introduce OSWorld-Science, a benchmark and evaluation environment that combines scientifically meaningful tasks, artifact-based evaluation, and an efficient agent harness for studying computer use in the scientific domain. The benchmark contains 12 VLMs and 146 high-quality tasks across several scientific domains and software configurations, covering workflows such as molecular drawing and retrosynthesis, pathology image analysis, statistical computing, and physical simulation. Tasks are developed through expert proposals and iter
标题:When a Kindergartener Solves Calculus: Measuring Capability Leakage in Role-Prompted Reasoning Models
来源:arXiv
链接:https://arxiv.org/abs/2609.39846
摘要:We investigate the problem of role-capability leakage (RCL), in which a role-prompted reasoning model generates convincing in-role text while continuing to exhibit capabilities on benchmarks that exceed those implied by the assigned role. For example, when a model is prompted to assume the role of a kindergarten student, one might expect its performance on a mathematics benchmark to reflect kindergarten-level ability rather than expert-level proficiency in solving calculus problems. We introduce RoleCapBench, a curriculum-grounded benchmark for evaluating RCL across six educational roles and four assessment levels spanning elementary school through A-level, and use it to evaluate three open-weight reasoning models. We find that although the models can generate stylistically convincing in-r
可信 AI 审稿人的修辞鲁棒性评测与 SciCore 方法arXiv · 模型可靠性 · 论文论文提出修辞鲁棒性评测目标与 RobustReview 基准,并通过双分支 SciCore 审稿方法降低 AI 审稿结果对论文措辞变化的敏感性,同时保持对科学内容差异的区分能力。
标题:A Missing Piece for Trustworthy AI Reviewers: From Benchmarking Rhetorical Robustness to SciCore Review
来源:arXiv
链接:https://arxiv.org/abs/2609.39027
摘要:AI reviewers can assign different judgments to manuscripts that report the same science in different wording, potentially rewarding rhetorical optimization over scientific improvement. We formulate Rhetorical Robustness as the joint requirement of stability across content-preserving rewrites and discrimination across papers. We introduce RobustReview, a controlled full-manuscript benchmark with 1,260 manuscript versions, and evaluate 30 reviewer configurations. The benchmark reveals false robustness, where low rewrite sensitivity coincides with score collapse across papers, and shows that human alignment and rhetorical robustness rank reviewers differently. Moreover, the evaluated content-focused prompting protocol does not consistently improve robustness across backbones. Motivated by the
[合并自 2026-10-08 scicore-ai]
[可信 AI 评审员缺失的一环:从修辞稳健性基准测试到 SciCore 评审(65 ▲)](https://huggingface.co/papers/2609.39027?utm_source=digest-papers&utm_medium=email&utm_campaign=2026-10-02)
标题:Self-Spec Verifiable Code Generation
来源:arXiv
链接:https://arxiv.org/abs/2609.39568
摘要:Large language models (LLMs) may generate unreliable code on corner cases missed by testing, while formal verification can provide machine-checkable guarantees. Recently, researchers have proposed several benchmarks to evaluate the capabilities of LLMs in generating formally verifiable code, where LLMs need to formulate formal specifications, generate the corresponding code, and verify its correctness. However, existing benchmarks have two key limitations: (I) They primarily evaluate specification and code generation stage-wise, with code generation typically conditioned on an oracle specification. This setup overlooks whether strong stage-wise performance translates into end-to-end success. (II)They mainly focus on a single proof-oriented language and mathematically structured tasks, offe
标题:CATCH: A Controllable Analysis Testbed for Reward Hacking in Coding RL
来源:arXiv
链接:https://arxiv.org/abs/2609.39533
摘要:During reinforcement learning with verifiable rewards (RLVR), large language models (LLMs) can exploit loopholes in their environments to obtain high rewards without improving the intended capabilities, i.e., reward hacking. Despite its risks to training efficiency and safety, monitoring and mitigating reward hacking during training remain challenging, which is limited by a lack of testbeds that reproduce hacking and reliably identify it. We introduce CATCH, a controllable testbed for studying reward hacking in coding RL. CATCH deliberately exposes environmental loopholes and provides execution-based gold labels by comparing success under a vulnerable evaluator with task correctness under an independent audit. It also can control the model's initial hacking tendency through supervised fine
标题:EngramBench: A Capability-Grounded Benchmark for Skill-Evolution Harnesses
来源:arXiv
链接:https://arxiv.org/abs/2609.39284
摘要:While large language models have achieved remarkable success in isolated code generation, authentic software engineering requires sustained reasoning, complex state management, and continuous cross-domain abstraction. However, current evaluations of skill evolution in autonomous agents suffer from a critical identifiability problem: they structurally confound genuine capability abstraction with rote solution leakage (i.e., copying highly similar code from historical training data). To resolve this, we introduce EngramBench, a rigorous, capability-grounded benchmark governed by the strict axiom of capability overlap without solution overlap. Comprising 30 diverse learning tasks and 13 unseen transfer tasks, EngramBench challenges agents to navigate interactive, multi-hour development cycles
标题:RAGScope: A Leakage-Controlled, Cost-Aware Evidence-Gating Protocol for RAG Hallucination Triage
来源:arXiv
链接:https://arxiv.org/abs/2609.39075
摘要:Retrieval-augmented generation (RAG) systems need inexpensive ways to route generated answers: accept low-risk outputs, review uncertain ones, and reserve strong verifiers for the expensive tail. We present RAGScope, a leakage-controlled protocol for evaluating local evidence gates that use only the task input, retrieved context, and answer text. The protocol combines context-grouped splits, fold-scoped preprocessing, group bootstrap intervals, deployment operating points, end-to-end runtime, and explicit source-shift stress tests. On three RAGTruth tasks, the enhanced gate RAGScope-E reaches 0.798 AUROC and 0.660 average precision (AP) in pooled grouped cross-validation. Its pooled AP exceeds ROUGE-L by 0.034 with a 95% context-group interval of [0.002, 0.064], although the AUROC gain is
研究提出 CoT-Interpretability Alignment(CIA),利用可解释性工具识别模型内部采用的推理策略,并衡量其与显式思维链描述的一致程度。作者在两跳问答、提示干预和整数乘法三类任务及三个大语言模型上进行评估,发现 CIA 仅为 44.8%~75.9%,表明思维链并不总能忠实反映内部计算。进一步实验将任务准确率和参数忠实度信号共同作为奖励进行后训练,在维持或提升任务准确率的同时显著改善了思维链的参数忠实度。该工作还分析了改进方法的泛化模式,为审计和训练更可信的显式推理提供了统一框架。
用途与启示
为思维链忠实度提供可量化的审计指标,避免仅凭自然语言推理过程判断模型决策机制。
可将 CIA 或参数忠实度信号纳入后训练奖励,联合优化任务表现与推理可解释性。
适用于评估模型是否存在“答案正确但解释失真”的风险,并支持可信推理和过程监督研究。
三类任务上的有限一致性说明,思维链不应被直接视为模型内部计算的可靠证据。
📝 原始笔记(逐字保留)
标题:Making LLMs Say What They Think: Measuring and Improving CoT-Interpretability Alignment
来源:arXiv
链接:https://arxiv.org/abs/2609.38972
摘要:Chain-of-thought (CoT) traces often serve as a proxy for how Large Language Models (LLMs) arrive at their answers. However, growing evidence shows that models' CoT often fails to reflect their internal computations and can be changed without affecting their final answers. In this work, we measure and improve the alignment between the reasoning described in an LLM's CoT and what it computes internally. We propose CoT-Interpretability Alignment (CIA), a metric that measures the agreement between a model's CoT traces and its internal reasoning strategies as detected by interpretability tools. We evaluate CIA on three tasks (two-hop question answering, hint intervention, and integer multiplication) across three LLMs, finding that LLMs exhibit limited alignment across all tasks (44.8-75.9%). We
标题:Where MLLMs Fail and Why: Causal Task Decomposition for Capability Failure Diagnosis
来源:arXiv
链接:https://arxiv.org/abs/2609.38851
摘要:End-to-end accuracy on compositional tasks records how often MLLMs fail, but cannot distinguish whether a failure reflects an intrinsic deficit in the targeted capability or a cascading error from an upstream prerequisite. We propose a causal decomposition framework that isolates these two failure modes through controlled interventions on the prerequisite dependencies of each task. Our capability metrics (NC, IC, RC) score each task under unassisted, correct, or incorrect prerequisites to diagnose where failures arise; contribution metrics (N-Score, S-Score), adapted from probabilities of causation, quantify each prerequisite's necessity and sufficiency to determine why. We instantiate the framework in CADET, a diagnostic benchmark of 10 composite tasks decomposed into 46 unit tasks with o
FIGS(Factual Integrity and Grounded Support)面向真实多轮对话评估大语言模型能否同时保持事实诚信与情感支持。框架采用自适应十轮对话模拟器,根据模型回答动态模拟用户的重复请求、反驳和答案诱导,而非依赖固定脚本。其评估体系严格区分迎合性与校准式认同,避免把理解用户感受等基本共情行为误判为对错误观点的让步。作者发布了包含 500 个多轮场景、自动评判器及测试环境的评测资源,并发现主流模型在持续互动中往往逐渐迎合用户,或过度纠偏为机械而冷漠的回应。
用途与启示
为多轮场景下的模型迎合性与事实诚信评估提供更贴近真实交互的动态基准。
提示评测设计应将情感认同与事实让步解耦,避免安全优化损害模型的正常共情能力。
可用于诊断模型在持续施压下的立场漂移,并支持后续对齐训练、模型选择和回归测试。
双轴指标有助于推动兼顾诚实性、支持性与长期一致性的对话模型开发。
📝 原始笔记(逐字保留)
标题:FIGS: Evaluating Multi-Turn Sycophancy Without Penalizing Empathy
来源:arXiv
链接:https://arxiv.org/abs/2609.39863
摘要:Large language models frequently fail to balance staying truthful with being supportive. They often exhibit sycophancy in responses to users, agreeing with false claims, offering unwarranted flattery, and giving advice skewed toward users' expressed views. In reality, sycophancy rarely happens in a single exchange; it may emerge organically as users repeatedly insist or subtly steer the dialogue over time. Current evaluations, however, rely on rigid, single-turn tests or fixed scripts that fail to capture these natural dynamics. Furthermore, these benchmarks often mistake showing basic empathy for yielding, penalizing models for acknowledging a user's feeling. This view may drive future models to over-correct into cold, dismissive rigidity. To address this gap, we introduce FIGS (Factual I
标题:Learning from Runtime Feedback through Failure-Bank Self-Evolution for Vision-Language-Action Models
来源:arXiv
链接:https://arxiv.org/abs/2609.39820
摘要:Vision-language-action (VLA) models generalize broadly across robotic manipulation tasks, but complex environments require balancing task success with unintended contact. Runtime shields can correct individual actions, but they leave the underlying policy unchanged, so repeated disagreements may create a persistent policy-shield mismatch that blocks task progress. To address this challenge, we introduce FailBank, a four-stage self-evolving framework that converts runtime feedback into persistent policy improvement. During collection, a fixed CBF-based safety module serves as an observe-only teacher, producing counterfactual corrections while the policy remains in control. Outcome-aware admission then converts useful proposals into corrective targets and retains successful uncorrected actio
标题:SE-ADD: Self-Evolving Audio Deepfake Detection with Mistake-Driven Supervision
来源:arXiv
链接:https://arxiv.org/abs/2609.39679
摘要:Audio deepfake detection (ADD) must remain effective when new spoofing attacks emerge after deployment. Emerging audio language model (ALM)-based ADD methods are built on predefined supervision from ground-truth labels or verified forensic rationales. However, this paradigm overlooks an ALM's own mistakes, which indicate where targeted supervision is most needed. To this end, we first introduce evolving spoofing environments for ALM-based ADD, where a new attack becomes dominant while previously observed attacks persist. Motivated by the above learning-from-mistakes perspective, we further propose SE-ADD, a self-evolving framework that iteratively adapts an ALM via low-rank adaptation (LoRA) using mistake-driven supervision built from its verdicts and self-generated forensic cues. All trai
论文研究基于 LLM 的自动启发式设计(AHD),其中模型持续生成包含设计理由与可执行代码的候选启发式,并由任务评估器返回有效性和性能结果。作者将小型开放权重模型的在线后训练表述为上下文相关的信号构造问题,比较从程序有效性、任务表现及生成上下文映射到更新信号的不同方法。实验在共享评估轨迹和匹配更新预算下,对多类 AHD 任务与模型家族进行受控比较,分别衡量有效提案率、有效程序性能及相对上下文有所改进的候选产出率。研究还结合模型检查点、冻结搜索状态和在线系统评估,区分模型能力提升与搜索状态积累带来的收益,并在预设成本口径下比较在线更新和冻结生成器的额外搜索。
用途与启示
为把可验证执行结果转化为 LLM 在线强化学习信号提供上下文相关的设计框架。
提醒自动算法设计研究不能仅凭端到端搜索收益断言模型能力增强,应隔离搜索状态积累与参数更新的贡献。
可用于设计资源匹配的模型自演化实验,比较在线后训练与增加冻结模型搜索预算的真实性价比。
为程序进化、启发式搜索和 RLVR 系统建立更细粒度的有效性、性能与候选改进率评估方法。
📝 原始笔记(逐字保留)
标题:From Search to Signal: Online Post-Training in Automatic Heuristic Design
来源:arXiv
链接:https://arxiv.org/abs/2609.39383
摘要:Large language model (LLM)-based automatic heuristic design (AHD) iteratively proposes and refines heuristics, pairing design rationales with executable code. Task-specific evaluators assess programs; execution outcomes and performance scores guide search. Many AHD systems keep the generator frozen; EvoTune and Co-Evolution of Algorithms and Language Model (CALM) instead update it from evaluated candidates. When such outcomes drive reinforcement learning with verifiable rewards (RLVR), they create a search-coupled loop: the evaluated candidate stream supplies both search-state updates and training signals for the model that generates future candidates. Yet validity and performance do not uniquely determine useful model updates; converting them into learning signals must account for the pro
标题:Better Supervision Is Nearby: Neighborhood On-Policy Self-Distillation
来源:arXiv
链接:https://arxiv.org/abs/2609.39687
摘要:On-policy self-distillation (OPSD) trains mathematical reasoning models using a privileged teacher that sees a reference solution and supervises student-sampled prefixes. Standard OPSD uses one fixed parameter setting at every state, but nearby settings may offer additional supervision. We find that local parameter perturbations reveal complementary reference-aligned corrections under the same reference context. Different experts supply these corrections at different reference positions. Their pool covers more such positions than the unperturbed privileged teacher. We introduce Neighborhood OPSD (N-OPSD) to turn these corrections into supervision at student-visited states. Offline, greedy selection builds a compact pool of frozen experts by rewarding filtered reference-token gains beyond t
标题:Towards Better Exploration in Sequential Test-Time Scaling
来源:arXiv
链接:https://arxiv.org/abs/2609.39632
摘要:Test-time scaling improves language model reasoning by spending additional compute at inference. However, both classes of existing methods often fail to continue improving over long timescales. Parallel methods repeatedly sample independent answers from the model, scaling poorly on problems the model is unlikely to solve in a single attempt. In contrast, sequential methods build on previous answers to access new ideas, yet so far have not been shown to reach answers beyond those found by parallel scaling. First, we show that sequential scaling often stops improving because it becomes prematurely trapped in an attractor: a set of answers that prevents exploration of different answers once entered. Across 27 combinations of scaling methods, models, and benchmarks, we find that 53.8% of seque
标题:GRPO Training Dynamics for Small Language Models
来源:arXiv
链接:https://arxiv.org/abs/2609.39321
摘要:Group Relative Policy Optimization (GRPO) has emerged as a memory-efficient reinforcement fine-tuning (RFT) technique for reasoning-intensive tasks. How- ever, GRPO training dynamics on small language models (SLMs) remain poorly understood, limiting its reliable adoption and reproducibility in open and resource- constrained environments. In this work, we present a systematic study of GRPO fine-tuning for SLMs ranging from 1.5B to 7B parameters under a practical single- node 8xA100 compute budget. Our study spans multiple model families and reasoning domains, including mathematics, coding, and multiple-choice question answering (MCQ) in science. Across these settings, we analyze how group size affects policy convergence, training stability, and downstream benchmark per- formance. We further
标题:Reinforcing Multimodal Reasoning via Token-Level Perception-Grounded Advantage Estimation
来源:arXiv
链接:https://arxiv.org/abs/2609.39168
摘要:Reinforcement Learning with Verifiable Rewards (RLVR) has improved the reasoning capabilities of Multimodal Large Language Models (MLLMs), yet existing frameworks rely on coarse, sequence-level reward signals that lack the fine-grained supervision over the visually-grounded steps within a multimodal reasoning chain. We investigate this gap through the lens of two token-level metrics: visual dependency (i.e. how much a token's prediction relies on the input image features) and predictive entropy. Our empirical analysis reveals two key findings: (1) correct reasoning chains exhibit a markedly sharper entropy reduction as visual grounding intensifies, compared to incorrect ones; (2) pivotal tokens, those whose misprediction triggers reasoning collapse, are statistical outliers in the joint di
标题:T-Router: Learning Thalamic Routing for Reasoning with Parameter-Efficient Reinforcement Learning
来源:arXiv
链接:https://arxiv.org/abs/2609.39109
摘要:Parameter-efficient reinforcement learning aims to improve reasoning with a compact trainable interface to a pretrained model. We introduce the Thalamic Router (T-Router), which concentrates adaptation on the reuse of completed computations. A compressed, addressable bank preserves block changes; a depth-recurrent controller conditions their selection and relative-scale writeback. This coupling gives thalamic context-dependent routing a concrete computational form: learn which earlier contributions a receiving layer uses, and with what influence. Correctness rewards train the interface while preserving backbone parameters and layer order. On an 8.95B-parameter backbone, T-Router allocates 41.73M parameters (0.466% of the backbone) and achieves 83.64 +/- 1.16 MathAvg after GSM8K RL, compare
标题:Settle: Learning When to Stop Reasoning
来源:arXiv
链接:https://arxiv.org/abs/2609.38997
摘要:Reasoning models often continue generating after their answers have settled. Settle learns when to stop from answer stability in completed traces. It trains the existing end-of-reasoning token while keeping other predictions close to the base model, and requires only ordinary decoding at inference. On MATH-500 with Qwen3-4B, Settle reduces token count by 40% with a 0.5-percentage-point decrease in accuracy. It gains 6.16 percentage points over supervised fine-tuning on the same traces shortened at their first stable answer, at nearly identical token counts. Its stopping score predicts whether a correct answer will remain correct. Settle extends the accuracy-token-count Pareto frontier of the evaluated stopping methods.
标题:What Limits Recursive Reasoning Models: Optimization, Architecture and Test-Time Scaling
来源:arXiv
链接:https://arxiv.org/abs/2609.39967
摘要:Recursive reasoning models apply a small shared Transformer block many times to refine a latent state. This gives them large effective depth with few parameters and makes them strong on algorithmic tasks. Such compact solvers are natural candidates for tools that an LLM can call on narrow algorithmic subproblems. However, existing models such as HRM, TRM and URM differ in architecture, gradient propagation and training procedure simultaneously. This makes it hard to tell what drives their performance, and their optimization is still poorly understood and often unstable. In this work we address both of these gaps. First, we study these questions under a unified experimental pipeline spanning six algorithmic domains. Individual controlled ablations are performed on representative domains, wh
标题:Advancing Entropy-Level Credit Assignment in RLVR via Proximal Entropy Policy Optimization
来源:arXiv
链接:https://arxiv.org/abs/2609.39402
摘要:Value-model-free RLVR methods such as GRPO assign uniform advantages to all tokens in a rollout, ignoring that tokens contribute unequally. Recent methods use token entropy as an importance proxy but compute it globally across the batch, conflating importance with prompt difficulty and positional trends. We argue that importance should instead be measured relative to the local context of each token. We introduce proximal entropy, a local measure of token importance relative to neighboring tokens, and prove it is invariant to both confounders. Proximal Entropy Policy Optimization (PEPO) uses it to weight per-token advantages and outperforms GRPO and entropy-based baselines on mathematical reasoning across Qwen3-1.7B, Qwen3-4B, and Llama-3.2-3B-Instruct. We also show the formulation generali
标题:Is This Evidence Decision-Critical? Learning to Verify Rule-Governed Decisions
来源:arXiv
链接:https://arxiv.org/abs/2609.39608
摘要:Rule-based reasoning, as in eligibility checks and contract reviews, requires language models to assess evidence against individual conditions and combine their judgments under explicit rules. Errors in evidence assessment can leave a decision unchanged, but misinterpreting or overlooking decision-critical evidence can reverse it. Identifying such evidence allows more capable models to focus on checking the corresponding condition judgments, supporting accurate and safe decisions. Recognizing the evidence's criticality requires understanding how evidence affects a condition judgment and how that judgment affects the decision. To achieve the goal, we propose a INTERvention-based imPACT learning framework (InterPact), which enables counterfactual verification of evidence criticality in rule-
标题:Candidate Retention for Abductive Learning
来源:arXiv
链接:https://arxiv.org/abs/2609.39561
摘要:Abductive learning combines neural perception with symbolic reasoning, using explanations generated by abduction to supervise the perception model. Multiple valid explanations of the same symbolic target can assign conflicting labels to the same inputs. Common policies select a single candidate as a pseudo-label, which may reinforce mistaken assignments, or weight all candidates, which may spread supervision across competing labels. These risks motivate selecting a retained subset to balance supervision sharpness and model-mass coverage. To guide this choice, we bound the coordinate-level supervision error using retained uncertainty, discarded model mass, and model mismatch. For a fixed model and training pair, only the first two terms depend on the retained set. We propose Abductive Candi
标题:WorkGenesis: Building the Worlds That Teach Agents to Work
来源:arXiv
链接:https://arxiv.org/abs/2609.39325
摘要:The ability of Large Language Model (LLM) agents to complete daily and professional work is receiving increasing attention. Training such agents requires realistic work scenarios. Expert-authored occupational work is costly and slow to produce, while unconstrained synthesis often yields tasks with weak factual grounding or internally inconsistent requirements. To bridge this gap, we introduce WorkGenesis, a framework that constructs executable occupational work from real-world artifacts through two core technical innovations: (1) Evidence-Based Work Construction, which grounds each unit of work in real-world evidence by retrieving public files guided by O*NET occupational knowledge and synthesizing the surrounding context, companion materials, work request, and itemwise rubric around them;
标题:PhantomEnvironments: Training LLM Agents in Fictional Worlds
来源:arXiv
链接:https://arxiv.org/abs/2609.40221
摘要:Training LLM agents with reinforcement learning (RL) is bottlenecked by environments, which must provide verifiable rewards, support long-horizon interaction, and scale cheaply. Existing approaches rely on costly human-curated data or on LLM-generated environments that risk hallucinations and benchmark contamination. We show that LLMs can instead be trained into capable search agents using synthetic environments generated entirely by rules, whose generation requires no LLM and has zero marginal cost. We build PhantomEnvironments, multi-turn RL environments from fictional worlds, where agents must search a corpus of templated articles to answer multi-hop questions. Despite sharing no facts with the real world, these strikingly simple environments yield agents that transfer to real-world mul
标题:DashVMC: Real-Time Discrete World Model Control in Geometry Dash
来源:arXiv
链接:https://arxiv.org/abs/2609.40003
摘要:World-model agents are usually evaluated in simulators that can wait for the policy; live games impose the opposite constraint, requiring capture, prediction, and action before the next frame. We present DashVMC, which learns a compact, action-conditioned world model from approximately two hours of recorded Geometry Dash gameplay. To test whether the learned dynamics are actionable, a controller is initialized by behavioural cloning (BC) and refined with Proximal Policy Optimization (PPO) entirely in frozen-model rollouts, without further interaction with the live game. Across three controller seeds, the refined policies survive longer than their BC initializations on all three official levels and a held-out community layout. At deployment, the baseline skips visual generation and sustains
标题:RoboCoach: World Models as Active Coaches for Compositional Robot Skills
来源:arXiv
链接:https://arxiv.org/abs/2609.39685
摘要:Long-horizon robot manipulation reuses skills across many task compositions, but improving these compositions with additional end-to-end demonstrations is costly. A practical self-improving system must decide both what to teach next and where to apply that supervision. We present ROBOCOACH, a world-model-guided coaching framework that uses imagined failures to guide demonstration requests and expert updates. Its Route-Imagine-Diagnose-Improve (RIDI) loop executes reusable skill experts inside COACHWORLD, our shared action-conditioned world model, and uses a progress judge to record the first subtask that fails to complete. Aggregated records select which subtask demonstrations to acquire and which expert adapters to update. Across two simulation suites and two real-robot platforms, imagine
LEGO-Anything:以编程智能体重建可执行 3D 场景Hugging Face Papers · 三维重建 · 论文LEGO-Anything 让编程智能体迭代生成、执行和修订 Blender 代码,将单张图像重建为可检查、编辑和查询的显式 3D 场景程序。
标题:Audio Token Attention Is Predictable Before the Language Model Runs
来源:arXiv
链接:https://arxiv.org/abs/2609.38878
摘要:A large audio language model (LALM) turns a minute of speech into 750-1,500 tokens and prefills every one. Image-token pruning often cuts after the language model's first layers, where image tokens draw little attention. Audio tokens draw much more attention there, and their ranking is still far from final, so audio needs a ranking before the language model runs. Surprisingly, the attention an audio token will receive across the language model is already linearly predictable from its encoder output, before the language model runs. A linear map, fitted in closed form without labels, predicts this all-layer attention ranking at $ρ\geq .69$ on eleven of thirteen LALMs. Our method, Triage, cuts audio tokens by this prediction and, on multiple choice, cuts again at layer 2, correcting the predi
Think Before You Score:用于视觉生成的思考型奖励模型Hugging Face Papers · 奖励建模 · 论文论文提出思考型奖励模型 TRM,先为每个视觉生成样本制定自适应评分准则,再开展细粒度评估,并通过 PD-GRPO 增强奖励区分能力、缓解偏好优化导致的分数极化。
标题:PivotOPD: Learning to Recover from Pivotal Mistakes in Multi-Turn Agents
来源:arXiv
链接:https://arxiv.org/abs/2609.40285
摘要:On-policy distillation (OPD) is a promising approach for training language agents, providing dense teacher supervision on student-generated trajectories. However, in multi-turn interaction, an incorrect action changes the states the student encounters later, so errors compound across turns. In preliminary experiments across three Qwen3 models (8B to 235B), we find that more than half of the failed rollouts contain a pivotal mistake, an action that moves the agent farther from completing the task, and this mistake typically occurs early. These pivotal mistakes often remain recoverable: guiding the model for only a few turns after the pivotal turn can restore task success. We therefore propose PivotOPD, an on-policy distillation framework that jointly trains the student to prevent pivotal mi
标题:Less Data, Better Timing: Student-Curriculum Coupling for VLM On-Policy Distillation in Temporal Video Grounding
来源:arXiv
链接:https://arxiv.org/abs/2609.40055
摘要:On-policy distillation (OPD) provides dense supervision directly on student-generated trajectories, making it an effective post-training strategy for vision-language models in temporal video grounding (TVG). However, existing pipelines typically construct the training curriculum from a fixed teacher and the initial student state, implicitly assuming that selected examples retain positive supervision value throughout optimization. We show that supervision trustworthiness and supervision necessity are distinct yet coupled: the former concerns target credibility, while the latter varies with the student's current task competence; together, they shape supervision value. Building on this coupled view, we introduce Student-Curriculum Coupling (SCC), a closed-loop framework in which a compact Anc
标题:OPSRD: On-Policy Self-Role Distillation
来源:arXiv
链接:https://arxiv.org/abs/2609.39884
摘要:Role prompting elicits specialized behavior from large language models through an expert identity, offering a lightweight way to guide reasoning on demanding tasks. However, evaluating or distilling complete role-prompted answers can miss useful next-token preferences when the sampled solution remains incorrect. Transferring these preferences also requires an objective that reaches alternatives the student rarely predicts. We introduce OPSRD, which uses a fixed expert role as privileged teaching context for on-policy self-distillation without reference solutions. A role-free student generates a trajectory, and a frozen instance of the same base model supplies role-conditioned distributions on its exact prefixes, exposing alternatives beyond the sampled continuation. Teacher-weighted forwar
黑盒蒸馏可将仅提供文本输出的 API 教师模型能力迁移至较小的学生模型,但现有 GAD 方法的评论器训练目标与 GRPO 所需的组内相对奖励存在错配。该错配可能导致同一提示下的奖励尺度坍缩或间隔脆弱,使优势估计和策略优化不稳定。GRGC 在评论器侧使用高斯组内最优传输校准,使排序后的奖励匹配以组中心为基准的高斯分位数,从而维持非坍缩的分布范围和平滑间隔。在策略侧,组幂调制在保留奖励排序的同时扩大与优化相关的间隔;跨教师、学生模型系列、模型规模及数据集的实验显示,该方法改善了分布内外表现,且相对 GAD 仅增加可忽略的开销。
用途与启示
为基于 GRPO 的黑盒蒸馏提供更稳定、更具区分度的组内优势信号。
揭示评论器的教师—学生判别目标与学生样本组内排序需求之间的关键错配。
可将奖励几何校准作为低开销模块,应用于对抗蒸馏及其他依赖组相对优势的强化学习训练流程。
提示奖励模型不仅需要正确排序样本,还应显式控制组内尺度、间隔和平滑性。
📝 原始笔记(逐字保留)
标题:Revisiting On-policy Adversarial Black-Box Distillation: Calibrating Groupwise Reward Geometry for Effective Advantage Construction
来源:arXiv
链接:https://arxiv.org/abs/2609.39757
摘要:Black-box distillation is a practical route for transferring capabilities from API-accessible large language models that expose only text outputs into smaller student models. Recent on-policy adversarial methods such as GAD improve over SeqKD by forming an adversarial loop between a critic and a student, where the critic provides rewards for GRPO-based student policy optimization over the student's sampled responses. However, GRPO computes advantages from the within-group relative rewards of student samples for the same prompt, whereas the critic is trained primarily to distinguish teacher responses from student responses. This objective mismatch can produce reward groups with collapsed scale or fragile margins, leading to brittle grouped optimization signals. We propose Groupwise Reward G
标题:ReTaCo: Residual-Target Control for On-Policy Distillation
来源:arXiv
链接:https://arxiv.org/abs/2609.39275
摘要:On-policy distillation (OPD) trains a student on its own generated prefixes with token-level teacher feedback, but transmitting or storing the teacher's full-vocabulary distribution at every token is costly. Entropy-aware OPD (EOPD) adds forward supervision to reverse KL to help the student recover plausible tokens it underestimates, using only the teacher's top-$k$ probabilities to limit cost. Because EOPD renormalizes these probabilities, its target assigns no mass to the omitted vocabulary. We prove that the resulting loss keeps pushing the student's top-$k$ mass toward one even after the student matches the teacher's relative probabilities within the top-$k$ set, so the teacher itself is not a stationary point whenever the omitted tokens have positive teacher probability. We propose Re
标题:When Clipping Reverses Correction: Failure Dynamics of Pointwise Forward-KL On-Policy Self-Distillation
来源:arXiv
链接:https://arxiv.org/abs/2609.38995
摘要:On-policy self-distillation (OPSD) trains a student on its own generated responses using feedback from the same model conditioned on privileged information. On mathematical reasoning, the original OPSD study finds that stylistic tokens can dominate the training signal over math-related tokens, and that pointwise clipping of the forward KL objective stabilizes training. Pointwise clipping caps each vocabulary-wise forward KL term at a fixed threshold before summing over the vocabulary. Follow-up studies have adopted this clipping, but its effect on training has not been directly examined. In matched training runs differing only in whether clipping is applied, we observe that clipped runs produce substantially more repetitions that persist to the end of the response than their unclipped coun
标题:Mitigating the Length-Scaling Tax with Online Distillation
来源:arXiv
链接:https://arxiv.org/abs/2609.38854
摘要:Length scaling during reinforcement-learning (RL) post-training is often viewed as a sign of improved reasoning ability, especially on difficult problems, but may also make responses to already-solved problems unnecessarily verbose. We quantify this side effect as the length-scaling tax (LST): excess response length on already-solved queries without a commensurate accuracy gain. To mitigate LST, we propose Length Self-Distillation (LSD), which routes solved prompts to on-policy distillation and retains the original RL objective for unsolved prompts. LSD uses an exponential moving average of the online policy as its teacher, requiring no external model. We find that LSD achieves comparable or better performance than RL across multiple variants, while substantially curbing response-length gr
标题:From Imitation to Reward Discovery: On-Policy Warmup for Agentic RL
来源:arXiv
链接:https://arxiv.org/abs/2609.39436
摘要:Reinforcement learning with a verifiable reward (RLVR) offers a scalable approach to training language-model agents, yet sparse outcome rewards can leave early training with little signal for policy improvement. We identify an On-Policy Acceleration Phenomenon: in our main comparisons, RLVR initialized with on-policy distillation reaches high performance earlier in training and achieves both higher average performance during subsequent RLVR and higher final performance than the alternative baselines. Motivated by this observation, we study On-Policy Warmup (OPW), a teacher-guided stage in which the student trains with teacher supervision on its own interaction trajectories before transitioning to RLVR. Unlike imitation on fixed teacher-generated trajectories, OPW targets states induced by
PanoVLN:高效全景视觉语言导航Hugging Face Papers · 具身规划 · 论文PanoVLN通过长动作序列预测、置信度引导执行及全景语义—几何表征,更充分地利用宽视野信息,显著提升视觉语言导航的成功率与执行效率。
标题:Taming Speculative Search for Test-Time Scaling in LLM Serving
来源:arXiv
链接:https://arxiv.org/abs/2609.39334
摘要:Test-time scaling has recently emerged as a powerful approach for improving LLM reasoning by allocating additional computation during inference, substantially enhancing accuracy on challenging tasks such as mathematics and coding. To accelerate the exploration of reasoning paths, recent studies proposed speculative execution. However, we show that supporting speculative execution poses two unique challenges for LLM serving systems: (1) an explosion in the search space of candidate paths and (2) frequent, fine-grained verification tasks for candidates. To address these challenges, this paper proposes SpecScale, a serving system for efficient speculative execution. We introduce three techniques to reconcile the trade-off between latency and computational overhead: (1) early pruning of low-qu
标题:Semifactual Credit-Augmented Policy Optimization
来源:arXiv
链接:https://arxiv.org/abs/2609.40360
摘要:Reinforcement learning with verifiable rewards (RLVR) has improved the reasoning capabilities of large language models (LLMs), yet their predictions remain sensitive to task-irrelevant prompt features. We investigate this sensitivity through semifactual prompt interventions that preserve the underlying problem and its answer. Our analysis reveals substantial variation in token-level sensitivity and shows that suppressing high-drift token candidates during decoding improves reasoning accuracy without updating model weights. These findings highlight a limitation of Group Relative Policy Optimization (GRPO), which assigns the same outcome-derived advantage to every response token and may reinforce potential spurious dependence alongside useful reasoning. Motivated by this observation, we intr
标题:Trust the Critic More
来源:arXiv
链接:https://arxiv.org/abs/2609.39247
摘要:Standard language model RL algorithms credit every token of a long rollout with the same advantage determined by the terminal reward. Actor-critic methods can provide finer-grained credit assignment, but learned critics are generally considered too inaccurate to trust when training LLMs with RL. In recent works, even when a critic is present, it is used only for baseline estimation, so every trajectory must be rolled out to its terminal reward. We introduce Actor-Critic with Action Chunking (AC2) that removes the need to roll every trajectory to completion. AC2 instead assigns credit to action chunks: short continuations of prefixes of past trajectories. A learned critic scores the state reached at the end of each action chunk, allowing the policy to update without observing a terminal rew
标题:Cogentic: Multi-Agent Orchestration for Automated Proof Discovery
来源:arXiv
链接:https://arxiv.org/abs/2609.40324
摘要:We present Cogentic, a multi-agent harness for automated proof discovery on open research problems. While frontier language models can generate strong mathematical ideas in a single shot, single-shot generation is often insufficient for open problems that require exploring multiple competing conjectures, overcoming subtle technical obstructions, and retaining intermediate progress over a long horizon. Cogentic addresses these challenges through an iterative prove--verify loop in which an orchestrator allocates a population of independent provers across distinct proof directions, subjects their output to adversarial verification by several specialized components, and promotes confirmed intermediate results into a persistent verified ledger that later rounds build on. The harness is designed
野生 AI 生成网页文本的 Token 价值与缩放定律arXiv · 缩放定律 · 论文 · 代码研究通过预训练 800 个语言模型发现,混入野生 AI 网页文本对人类文本建模的收益会迅速饱和并转为负面,并提出可刻画 AI Token 价值变号的新缩放定律。
研究发现,经 FineWeb 质量过滤后,2026 年 6 月网页数据中仍有 27.5% 的 Token 被识别为 AI 生成,至 8 月升至 31.1%。作者预训练了 800 个语言模型,并改变 AI Token 与人类 Token 的混合比例,以测量两类验证文本上的损失。对数据匮乏的模型,AI 文本起初能降低人类文本损失,但收益很快饱和并反转为伤害;在人类文本预算充足时,加入 AI Token 几乎立即导致损失上升。论文提出包含独立收益项和伤害项的新缩放定律,在无 AI 文本时可退化为 Chinchilla 定律,并将跨模型规模预测误差较最佳既有方法降低 41%。作者同时发布带有 AI、主题和格式标签的 830 亿 Token WildAI 语料库、全部模型及代码。
用途与启示
为预训练数据配比提供依据:若目标是优化人类文本表现,应优先过滤 AI 生成内容。
在扩充 AI 网页文本前,可考虑重复高质量人类文本,因为新增 AI Token 的边际价值可能由正转负。
应分别报告人类文本与 AI 文本上的验证损失,避免混合指标掩盖训练目标差异。
新缩放定律可用于估计不同数据预算和混合比例下 AI Token 的实际价值,并辅助数据采购与训练资源配置。
📝 原始笔记(逐字保留)
标题:How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text
来源:arXiv
链接:https://arxiv.org/abs/2609.40295
摘要:Web text makes up the majority of pretraining data and is increasingly AI-generated. After applying FineWeb quality filtering, we find that 27.5% of tokens from June 2026 web data are labeled as AI-generated by Pangram, rising to 31.1% by August. Unlike synthetic data or model-collapse setups, this *wild* AI text comes from many models, is written for human readers, and arrives unlabeled in pretraining corpora. How does AI text in the wild affect language model pretraining? To answer this question, we pretrain 800 language models, varying the ratio of added AI tokens to human tokens, and fit scaling laws to held-out losses on both human and AI-generated text. For data-starved models, adding AI tokens to pretraining data initially lowers loss on human text, but the benefit saturates as more
标题:ComputerSD: Online Self-Distillation from Real-Time Feedback for Computer-Use Agents
来源:arXiv
链接:https://arxiv.org/abs/2609.40253
摘要:Online training enables computer-use agents (CUAs) to improve through interaction with executable environments. However, existing methods primarily rely on sparse outcome rewards, which provide no supervision for intermediate actions. On-policy self-distillation (OPSD) offers token-level learning signals through privileged rescoring, but directly applying it to CUA online training presents two challenges: fixed guidance may become misaligned with the student's current state, and guidance-induced probability shifts may conflict with step-level correctness. We introduce ComputerSD, an online self-distillation method for CUAs that converts real-time feedback from executed GUI transitions into guidance for policy learning. A fine-tuned GUI analyzer produces guidance and a step-level value scor
标题:Disentangling Self-Distillation: Measuring and Modeling Acquisition and Retention
来源:arXiv
链接:https://arxiv.org/abs/2609.39494
摘要:Self-distillation with privileged context adapts a language model from demonstrations by letting the model, once conditioned on a reference response, teach its context-free copy token by token. Our taxonomy reveals existing methods differ along three entangled axes: (i) the rollout source (student or teacher), (ii) the teacher coupling (frozen, or an exponential moving average of the student at some coupling rate) and (iii) the KL direction (reverse or forward), yet these axes are usually studied in fixed combinations and have led to conflicting conclusions. We formalize a unifying framework to encompass all self-distillation methods vs classic supervised fine-tuning: we train every combination of the three axes, on Qwen2.5-7B and Ministral-3-3B across ordinary and contradictory tasks, tot
标题:Agent Error Dataset: Scaling 50,000 Error--Diagnosis Pairs for Failure Analysis and Error-Aware Post-Training
来源:arXiv
链接:https://arxiv.org/abs/2609.40111
摘要:An unsuccessful LLM agent rollout contains more information than its final reward: the observations available to the agent, the actions it chose, and the environment's responses. Reusing this experience for learning requires identifying a decision to revise and testing a concrete alternative. We introduce the Agent Error Dataset (AED), comprising 50,228 error-diagnosis pairs from 9,961 source tasks across 33 environments, 19 harness families, and 23 policy models in text-based agent systems. We retain source traces and execution metadata to support cross-setting failure analysis and re-diagnosis without repeating the original rollout. Our five-stage Agentic Error-to-Training (AET) pipeline collects natural failures, generates diagnoses and proposed corrections, and checks them against reco
Replay on Demand(RoD)不再预先设定固定回放比例,而是根据模型训练过程中的动态状态在线决定样本分配。该方法按照剩余学习潜力选择新领域适应样本,并依据已观测到的遗忘程度优先选择回放样本。两类样本竞争共享训练预算,由此形成随训练进程变化的在线课程。跨模型规模与适应领域的实验表明,RoD达到或超过经调优的固定回放及模型合并基线,并会将回放资源集中于更易遗忘的数据源。
用途与启示
为持续预训练提供无需人工设定固定回放比例的动态数据调度方案。
可根据能力遗忘出现的时间和来源自适应增加、重新分配回放预算。
表明模型学习动态可直接作为课程构建信号,在领域适应与旧能力保持之间取得更优平衡。
📝 原始笔记(逐字保留)
标题:Replay on Demand: An Emergent Curriculum for Balancing Adaptation and Forgetting in Continued Pretraining
来源:arXiv
链接:https://arxiv.org/abs/2609.40089
摘要:Continued pretraining enables language models to adapt to new domains and knowledge, but often at the cost of forgetting previously acquired capabilities. Replay can mitigate this trade-off, but fixed replay mixtures allocate training independently of the model's actual retention needs. We introduce Replay on Demand (RoD), which instead derives the replay allocation from the model's learning dynamics. RoD jointly prioritizes adaptation samples by their remaining learning potential and replay samples by their observed forgetting. Their competition for a shared training budget yields an online curriculum that determines what to train on at each step. Across models, scales, and adaptation domains, RoD reaches or improves upon the adaptation-forgetting frontier of tuned fixed-replay baselines
标题:Pseudo-Label-Triggered Retraining from Forecast Errors for Online Time Series Forecasting
来源:arXiv
链接:https://arxiv.org/abs/2609.39789
摘要:Real-world time series forecasting systems operate under non-stationary data streams, where forecasting performance may degrade over time. Although retraining can recover the performance, it incurs non-trivial computational and operational costs. Under limited deployment resources, the key challenge is therefore not only how to retrain but also when to retrain. While existing retraining policies often rely on indirect indicators such as drift alarms or model staleness, we instead use realized forecast errors as direct deployment feedback. In this paper, we propose PILOT (Pseudo-label-Informed Learned Online Trigger), an online retraining framework that learns when to retrain from forecast-error dynamics. Since ground-truth retraining labels are unavailable, PILOT constructs a pseudo-label
标题:A Dynamical Theory of LoRA in Continual Learning
来源:arXiv
链接:https://arxiv.org/abs/2609.39367
摘要:Despite the widespread use of Low-Rank Adaptation (LoRA), little is known about its dynamics in continual learning and the mechanisms by which low-rank updates affect catastrophic forgetting. We provide an asymptotically exact dynamical characterization of LoRA in a solvable two-task teacher-student model. In the high-dimensional online-learning limit, we derive a closed system of ordinary differential equations for a finite set of macroscopic order parameters, yielding exact expressions for the generalization errors throughout both the initial Task 1 learning phase and the subsequent LoRA fine-tuning on Task 2. The theory quantitatively matches finite-dimensional simulations and exposes two characteristic effects of LoRA: low-rank adaptation reduces interference with features learned on t
标题:Generalized Residual Closure: General Learning Dynamics for Stability-Plasticity Compatibility
来源:arXiv
链接:https://arxiv.org/abs/2609.38911
摘要:Learning must acquire new capabilities while preserving both prior responsibilities and the capacity to learn again. We introduce Generalized Residual Closure (GRC), a framework for learning as recursive closure of future-relevant discrepancies: closing a residual establishes the conditions for subsequent prediction, interaction, and learning. Under a complete representation-relation description at a fixed learner-world boundary, persistent internal learning has two primitive modes: Transformation within a representation and revision of the Representation itself. We establish a local tangent decomposition under regularity assumptions and a criterion for when representation revision is necessary. In an affine model, we derive a necessary-and-sufficient condition for stability-plasticity com
标题:Ready2Blend: From Natural-Language Instructions to Composable Alignment Prompts
来源:arXiv
链接:https://arxiv.org/abs/2609.39365
摘要:Continual alignment requires LLMs to adapt to new requirements without forgetting previously acquired behaviors. Natural-language instructions are flexible and composable but offer only indirect control, whereas post-training provides stronger adaptation at the cost of repeated parameter updates. We introduce Ready2Blend, which combines the flexibility of natural language with learned alignment. AlignFormer maps each requirement to a fixed-length alignment prompt stored in a modular prompt bank, while the backbone and prior prompts remain frozen. Composability regularization transfers the semantic geometry of textual requirements into prompt space, enabling inference-time blending and reweighting. Across two practical continual alignment settings, Ready2Blend is the only frozen-backbone me
标题:Who Verifies the Graph? Misspecification Attacks on Causal Action Verification for Language Agents
来源:arXiv
链接:https://arxiv.org/abs/2609.40027
摘要:Causal action verifiers gate an agent's state-changing tool calls by checking whether each proposed intervention is identifiable against a committed action-state graph, and they issue a certificate that carries the identification argument and a one-sided lower confidence bound. One such verifier, CIVeX, reports zero false executions on a confounded tool-use benchmark. We red-team it by corrupting only the committed graph. Omitting a single bidirected edge takes it from zero false executions to 15.3% at the benchmark's published confounding strength, with 91% of its executions harmful and utility falling from +2.27 to +0.35. Reversing one arrowhead, so that a mediator is committed as a confounder, gives 48.9% false executions and no correct ones. Every one of these actions carries an intern
标题:ActionGuard: Tool Call Authorization under Poisoned Skills
来源:arXiv
链接:https://arxiv.org/abs/2609.39450
摘要:LLM-based agents extend their capabilities through third-party skills that provide task-specific instructions, scripts, and tool-use procedures. However, malicious instructions inserted into an otherwise benign skill can cause a benign user request to trigger dangerous Tool Calls, including data exfiltration, file deletion, or unauthorized code execution. This paper presents ActionGuard, which inspects skill-influenced Tool Calls immediately before execution. ActionGuard separates the target agent's action-generation context from the safeguard's authorization context. The target agent may use the original skill for planning, but the Reviewer does not receive the potentially poisoned raw skill text. Instead, it determines whether each action is justified by the trusted user request using a
HiSentinel 面向仓库级编程智能体,在动作执行前判断干预是否有助于最终任务完成,而非试图纠正每个不完美动作。框架使用可访问历史执行结果的特权教师生成干预判断,再将其蒸馏到仅接收动作前上下文和候选动作的 0.6B、1.7B 因果学生模型。作者构建了动作级数据集 SWE-Intervene,标注动作应被放行、自动重定向还是暂停并寻求人类协助,同时提供可操作的干预反馈。在 SWE-bench Verified Mini 与 Ask or Assume 上,该方法跨模型规模和编程智能体家族分别带来最高 14% 和 10% 的任务完成率提升,并保持有竞争力的 token 消耗。
用途与启示
将事后执行轨迹转化为事前风险判断,可在错误扩散前实施低成本干预。
轻量哨兵可作为独立安全层接入不同编程智能体,而无需显著改造主模型。
区分自动重定向与请求人工协助,有助于在自主性、可靠性和人工成本之间进行动态权衡。
SWE-Intervene 的动作级标注范式可用于训练和评估其他长程智能体的执行前干预机制。
📝 原始笔记(逐字保留)
标题:Learning When and How to Intervene: A Hindsight-Distilled Sentinel for Coding Agents
来源:arXiv
链接:https://arxiv.org/abs/2609.39957
摘要:Coding agents solve repository-level tasks through sequences of actions, where a single erroneous action can misdirect subsequent decisions and increase recovery costs. Existing approaches use execution feedback for recovery or specialized checks to block errors, but deciding before execution whether intervention will benefit eventual task completion remains challenging. To address this challenge, we propose HiSentinel, a hindsight-distillation framework that trains lightweight 0.6B and 1.7B sentinels to select pre-execution interventions aimed at improving task completion rather than correcting every imperfect action. A privileged teacher uses recorded execution outcomes as evidence for intervention judgments, which are distilled into a causal student that receives only the pre-action con
标题:When Instructions Retrieve Trajectories: Diagnosing and Mitigating Generalization Failures in VLA Models
来源:arXiv
链接:https://arxiv.org/abs/2609.39971
摘要:Vision-language-action (VLA) models can exceed 90% success on in-distribution tasks and withstand nuisance changes that preserve the required action, yet fail under counterfactual changes that demand a different action. Aggregate robustness scores can therefore conceal a more specific failure, in which a policy responds to both language and vision yet does not combine them to select the action the task requires. We call this failure instruction-action binding. Instructions cue familiar trajectory families, and visual feedback adjusts their execution. Behavioral analyses of fine-tuned $π_{0.5}$ and GR00T-N1.7 policies reveal that failed rollouts often retain the source behavior or switch to another demonstrated task. These switches show that language is not simply ignored. Readouts and inte
标题:AdaGEPA: Adaptive Feedback Allocation for Reflective Prompt Optimization
来源:arXiv
链接:https://arxiv.org/abs/2609.39927
摘要:Prompt optimization improves the performance of language-model systems on downstream tasks by refining their prompts. Classical methods evaluate prompts on task examples and use the resulting feedback to guide prompt revisions through reflection. However, when feedback selection does not account for the prompt's weaknesses, these revisions may improve performance on selected examples without yielding broader task improvements. To address this issue, we propose AdaGEPA, an adaptive feedback-allocation method that uses the prompt's performance and task structure to select examples for the next prompt revision. Our method replaces at most one example in each feedback minibatch to target an identified weakness while preserving the remaining feedback context. Across our main experiments on six
标题:K2P: Label-Free Knowledge to Prompt Distillation
来源:arXiv
链接:https://arxiv.org/abs/2609.38898
摘要:Knowledge distillation can transfer reasoning from stronger teachers to frozen students through reusable prompts, but avoiding weight updates does not eliminate supervision. Without ground-truth answers, teacher solutions are unverified, and agreement with the teacher can reward shared mistakes. We introduce Knowledge-to-Prompt (K2P) for label-free knowledge distillation to prompts. K2P synthesizes reusable instructions from teacher solutions, refines them using paired teacher and student responses, and guides search and selection with answer agreement. It retains candidates that adaptive search may undervalue and selects on reserved questions. Deployment uses only the frozen student and selected prompt. Our theory separates generation and selection gaps and gives conditions under which ag
标题:The Concrete-Arbitrary Gap: Kinship Reasoning in LLMs Is Not Indifferent to Presentation
来源:arXiv
链接:https://arxiv.org/abs/2609.39913
摘要:We test whether large language models solve formally matched kinship problems equally well when relations are expressed in familiar vocabulary or by explicitly defined nonce predicates. Across 500 paired graphs, concrete accuracy exceeds arbitrary accuracy by 35.6 percentage points in local Qwen3.8-27B, 26.6 in Gemma 4 26B-A4B, 12.0 in Gemma 4 31B, and 5.4 in Qwen3.8-Max. All four paired gaps are statistically resolved. Reasoning budgets and prompt-language interventions can substantially reduce the difference, showing that it is modifiable rather than a fixed incapacity. The minimal conclusion is behavioral: on these tasks, the models' manifested relational competence is not indifferent to presentation. Explicit definitions provide the formal relations but do not make nonce predicates as
标题:TRACE: Trajectory Selection for Parallel Scaling of Search Agents
来源:arXiv
链接:https://arxiv.org/abs/2609.39912
摘要:Parallel search may generate a correct answer that final-answer voting fails to select. We formulate this consolidation stage as trajectory selection and introduce TRACE (Trajectory Ranking with Aggregated Cross-Rollout Evidence), a lightweight learned selector that ranks completed trajectories using the search evidence behind their answers. TRACE preserves individual query and evidence occurrences, connects rollouts through shared content or document identity, and propagates information across these relations. Each candidate answer then reads the updated states of its own trajectory, preserving retrieval provenance while incorporating evidence from related rollouts. Trained with answer-level supervision over frozen text embeddings, TRACE returns an existing answer without additional searc
标题:Shared Weights, Selected Computations: How Looped Transformers Route What Each Loop Does
来源:arXiv
链接:https://arxiv.org/abs/2609.39892
摘要:Looped Transformers repeatedly apply the same set of Transformer layers, giving them a recurrent architecture for latent computation. Their strong performance on iterative reasoning and length-generalization tasks suggests an appealing explanation: recurrence may provide an inductive bias that lets the model reuse a learned algorithm across loops. However, weight sharing alone does not imply that every loop performs the same operation. This raises a basic question: is each loop actually repeating the same computation, and if not, what routes the shared parameters to different operations? We study this question using graph walks as a test case. In the model's native trajectories, decoded predictions can advance by different numbers of graph steps or remain at a reached target, showing that
标题:SEAR: Spoofing Evidence-Grounded Audio Reasoning Benchmark for Audio Language Models
来源:arXiv
链接:https://arxiv.org/abs/2609.39847
摘要:Audio language models (ALMs) are increasingly used for audio deepfake detection (ADD), yet existing benchmarks assess their verdicts or rationale plausibility without verifying the underlying acoustic evidence. To address this issue, we first introduce spoofing evidence-grounded audio reasoning (SEAR), a four-task AQA benchmark to evaluate ALM-based ADD through acoustic evidence identification and quantification, deepfake detection, and forensic rationale generation. We further propose a bona-fide-based acoustic evidence agent (BAEA), which equips a frozen ALM with controlled acoustic tools under \textsc{fixed} or \textsc{adaptive} evidence-acquisition policies. Experiments with six ALMs reveal a clear gap between plausible rationales and verifiable acoustic evidence reasoning, while BAEA-
标题:ArchitectureIQ: On the Measure of Training Intuition
来源:arXiv
链接:https://arxiv.org/abs/2609.39714
摘要:Top researchers have good intuition, but do language models have as good intuition about model training as top AI researchers? To measure model intuition of LLMs and humans, we introduce the ArchitectureIQ benchmark. Each question presents a synthetic dataset and several training recipes, and the test-taker is asked to predict the recipe yielding the best test metric. Overall, we find that LLMs' model intuition is good but has four limitations: (1) The intuition is imperfect, or even sub-human in some cases. Frontier models achieve around 76% accuracy (random choice 33%) vs best human researcher (66.0%), yet remain far from perfect. For architecture-only questions, best human achieves 65% while GPT-6 Astra only has 38%. (2) The intuition is empirical, not structured, supported by the fact
标题:Learning Steganography Is Easy, Learning Steganographic Reasoning Is Hard
来源:arXiv
链接:https://arxiv.org/abs/2609.39838
摘要:Chain-of-thought monitoring as an approach for AI oversight and control is threatened by the possibility of steganographic reasoning, where LLMs conceal their reasoning inside innocuous-looking text. Two neighbouring capabilities, steganographic messaging (passing a concealed message) and encoded reasoning (reasoning in an illegible but unconcealed format), have already been shown to emerge under training pressures that occur in real pipelines, such as reinforcement learning against monitors. This suggests that steganographic reasoning too might arise as an unintended side effect of training. Here, we compare how easily models learn steganographic reasoning and these two neighbouring capabilities across three elicitation methods: reinforcement learning, in-context learning, and supervised
标题:Synthetic Pre-pretraining Survives Scale, but Not as a Grammatical Prior
来源:arXiv
链接:https://arxiv.org/abs/2609.39827
摘要:Pre-pretraining (PPT) on synthetic non-natural language data improves token efficiency during language model pre-training (PT). Prior work attributes this gain to a grammatical prior, i.e., a structural inductive bias learned during PPT that transfers to natural language grammar. However, PPT has only been tested on models of at most 1B parameters and PT budgets below 2B tokens on predominantly web text. It is unknown whether PPT is effective at larger scales and under more realistic PT data mixtures that combine diverse sources (e.g., code and math). We therefore present a comprehensive study on PPT spanning five PPT tasks, four PT data mixtures, four parameter scales (500M to 7B), and PT budgets of up to 100B tokens. Our results demonstrate that the downstream performance and token effic
标题:Beyond Accuracy: Prefix-Invariant Realizations of Low-Precision Fast Matrix Multiplication
来源:arXiv
链接:https://arxiv.org/abs/2609.39816
摘要:Fast matrix multiplication saves multiplications through exact cancellation, but rounding sums that mix token rows can leave contributions from later tokens in earlier language model outputs. This threatens prefix invariance, which multiple-choice likelihood scoring relies on: a scored likelihood must depend only on its allowed prefix. On Qwen2.5-14B-Instruct, two fast FP8 realizations repaired to ordinary-looking accuracy still change the answers chosen by likelihood on 5.83% and 10.00% of 240 OpenBookQA items when only the text after the allowed prefix is replaced with the bf16 model's own greedy continuation. Both row-local controls, the bf16 model and a deployed FP8 matrix multiplication kernel, change none. Accuracy thus does not certify prefix invariance, and the stability criteria w
标题:Network-based Spatial Context Retrieval for Open-weight LLMs: A Faithfulness Benchmark for Grounded Geographic Reasoning
来源:arXiv
链接:https://arxiv.org/abs/2609.39437
摘要:Large language models (LLMs) encode substantial latent geographic knowledge, yet they reason poorly over space and are unreliable when queried from coordinates alone. Useful behaviour emerges only when structured spatial context is supplied in the prompt. This raises a question geographic evaluation has left unexamined: once the right context is supplied, does the model reason from it, or override it with its own parametric recall? We take up this question with an open pipeline for network-based spatial context retriev-al. In it, the surroundings of a selected point are defined by the pedestrian street network, the area actually reachable on foot. Using only open data and open-weight models, the pipeline retrieves features from OpenStreetMap and the GHS-POP population grid, computes indica
标题:How Much Can Reliability Drift Under a Fixed Confidence Distribution?
来源:arXiv
链接:https://arxiv.org/abs/2609.38917
摘要:A classifier's conditional accuracy can change while its confidence distribution stays exactly the same. We study the worst-case movement of the reliability relation under covariate shifts that preserve the distribution of the confidence score, constraining the reweighting within each confidence level by a $χ^2$ budget; the resulting worst case, as a function of the budget, is a fragility profile. On an interval of budgets that can be computed from the source distribution, the profile equals exactly the square root of the budget times the within-level variance of the correctness propensity -- the grouping-loss term of calibration-refinement decompositions. Beyond this interval the profile is governed by the tails of the propensity law, and the entire upward profile determines the centred w
标题:Thinking Outside the Box: Can Language Models Rely on External Guidance Selectively?
来源:arXiv
链接:https://arxiv.org/abs/2609.39578
摘要:Agent harnesses often improve language models with human-designed workflows, but as models grow more capable, unreliable guidance can increasingly constrain their execution. We call the ability to benefit from useful guidance while overriding unreliable guidance thinking outside the box. We introduce Box$^2$-Bench, which holds the model and task fixed while varying workflow reliability to isolate how models regulate their reliance on guidance. On Box$^2$-Bench, frontier models often benefit from reliable guidance but remain vulnerable when it is misleading or becomes unreliable. To test whether this capability can be learned, we train two open-weight models using bad workflows, reserving good workflows for evaluation. We explore two complementary training strategies: counterfactual supervi
标题:Safety of Latent Communication in Multi-Agent Systems
来源:arXiv
链接:https://arxiv.org/abs/2609.39788
摘要:Latent communication enables multi-agent systems to exchange information directly in internal representation space, reducing the token, computation, and latency overhead of text-based communication. To this end, lightweight trainable links are introduced to map the sender's representations into the receiver's input space. In this work, we show that even benign link training can increase harmful compliance relative to text-based communication while the underlying safety-aligned agents remain unchanged. An attacker can amplify this effect by optimizing the links on harmful query--response pairs or poisoning otherwise benign training data. We further develop a reinforcement-learning attack that rewards harmful compliance alongside benign task performance without requiring harmful target respo
标题:MiniRep: Robust Reputation-Based Aggregation for Multi-Agent Debate
来源:arXiv
链接:https://arxiv.org/abs/2609.39297
摘要:Autonomous agents powered by large language models (LLMs) are rapidly evolving into an open agentic ecosystem. To support trustworthy collaboration, industry initiatives increasingly assess agent reputation from past behavior and provide performance leaderboards. However, reputation derived from past performance may not reliably predict an agent's behavior on new tasks, particularly when malicious agents can adapt their behavior and influence other agents during collaboration. We study reputation in multi-agent debate (MAD), where multiple agents answer the same query, debate to improve their answers, and aggregate them into a final output. We present MiniRep, a reputation-based aggregation system for MAD under malicious agents. To ground our threat model in established research, we constr
标题:MADBench: Benchmarking the Security of Multi-Agent Debate
来源:arXiv
链接:https://arxiv.org/abs/2609.39146
摘要:Multi-agent debate (MAD) can improve large language model (LLM) reasoning by allowing multiple agents to exchange and critique their answers to the same task. However, the interactions that enable agents to correct mistakes can also spread adversarial errors and steer the agents toward an incorrect answer. Although some efforts have been made to examine particular attack types on MAD, systematic evaluation of MAD under diverse attacks remains limited. A central question is whether debate mitigates adversarial influence or amplifies it. In this paper, we present MADBench, a benchmark for evaluating the security of MAD. We organize attacks into a layered taxonomy following the MAD workflow, incorporating both established attacks and new strategies tailored to debate. We evaluate six attack f
标题:When Order Matters: First-Speaker Bias and Mitigation through Personality in Sequential Multi-Agent Debate
来源:arXiv
链接:https://arxiv.org/abs/2609.38964
摘要:Multi-agent debate (MAD) is often used to improve large language model (LLM) reasoning, but sequential debate is rarely a neutral aggregator of agents' opinions. We show that sequential MAD suffers from a pronounced first-speaker bias: agents disproportionately shape the final answer when they speak first. As a result, placing a stronger model after weaker ones can substantially offset its reasoning advantage. We then focus on the disadvantaged strong-agent-last setting and ask whether personality prompting can mitigate this imbalance. Drawing on the Big Five model, we study agreeableness and extraversion as behavioral interventions applied to either the strong or weak side. We find that their effects are trait-specific. Influence consistently shifts in the direction of lower agreeableness
标题:Pretext: Defeating Malicious Skill Detection Frameworks for AI Agents
来源:arXiv
链接:https://arxiv.org/abs/2609.39607
摘要:Skills extend an agent's capabilities by injecting instructions and information into the context, and are widely used by agents such as OpenClaw and Claude Code. Prior work shows third-party marketplaces host malicious skills that give attackers direct influence over the victim's agent. The emerging defense scans skills before installation, pairing deterministic static checks with an LLM-based semantic judge, as in NVIDIA's SkillSpector. We show that such defenses fall to an attacker who knows the detector. Our white-box LLM attacker, Pretext, iteratively crafts skills that evade detection while still delivering the payload and performing the benign task: moving the payload from code into natural language leaves static analysis inert, while framing it as the skill's legitimate purpose and
标题:Self-Repulsive Sampling for Diffusion Language Models
来源:arXiv
链接:https://arxiv.org/abs/2609.39560
摘要:Sampling several responses and voting over their answers can improve a language model's accuracy, but repeated answers limit the benefit of additional samples. Raising temperature increases diversity at a potential cost to per-sample accuracy. We introduce Self-Repulsion (SR), a sampler for masked diffusion language models that uses peer commitments to diversify the pool. At each penalized denoising step, each path lowers a token's logit according to how many peers have committed that token at the same position. Paths share a batched forward pass and then commit in sequence, so later paths observe choices made earlier in the same step. This coupling requires no training or additional forward or backward pass and can produce distinct paths even at temperature zero. When all paths commit a p
标题:Also Small Models Can Reasonably Self-Evaluate Their Confidence
来源:arXiv
链接:https://arxiv.org/abs/2609.39478
摘要:This study systematically evaluates self-evaluation-based uncertainty quantification across different language models of varying sizes on question-answering tasks spanning general to specialized knowledge domains. Using various self-evaluation methods where models judge their own predictions, we examine how model scale and domain specificity affect the quality of self-assessed confidence signals. Our results reveal that while accuracy predictably declines with smaller models and more specialized domains, the reliability of self-evaluated confidence remains largely stable across both dimensions. This independence means the most capable model is not necessarily the best at self-assessing prediction reliability. These findings suggest that smaller models can achieve reasonable self-assessed c
标题:Alleviating Hallucination in Reasoning Tasks with Training-Free Uncertainty-Guided Steering
来源:arXiv
链接:https://arxiv.org/abs/2609.38962
摘要:Recent work on hallucination detection in large language models has shown that, for a fixed pre-trained model and reasoning task, it is possible to estimate the model's confidence in the correctness of its outputs. Such uncertainty estimates have primarily been used to improve truthfulness by detecting or filtering confabulations. In this work, we ask whether these signals can instead be used more proactively to directly improve the accuracy of model-generated answers. We propose USteer, a simple, training-free steering mechanism that adjusts a model's layer-wise activations during inference using the gradient of a confidence measure with respect to the activations. This procedure nudges generation toward outputs with lower uncertainty at inference time, without modifying model parameters
标题:Can Computation from Earlier Problems Help LLMs Solve New Ones?
来源:arXiv
链接:https://arxiv.org/abs/2609.39394
摘要:Large language models often solve independent problems in the same conversation. Can computation from earlier problems help them solve new ones? To answer this question, we first conduct preliminary experiments showing that retained history can raise or lower later-turn accuracy, even within the same domain. To understand these effects, we use controlled replay to isolate internal state changes specific to each problem-history pairing. Across different histories, these changes preserve similar relationships among current problems. To improve reasoning under retained history, we introduce STAIR (Stale-Token Attention for Inter-query Reuse). STAIR captures keys and values from earlier response generation in a fixed bank. It learns to redirect current queries when they read this bank during p
CORE不再仅重启搜索或修订最后一步,而是请求验证器返回经过认证的冲突核,并回跳至其中最近的相关决策。系统会缓存已发现的冲突,避免后续搜索重复进入相同的失败分支;在可靠验证、有限分支与深度等条件下,无上限搜索具有完备性且不会剪除有效解。在2,000个植入解的图着色实例上,CORE相较按时间顺序修复将验证器调用中位数降低35.0%至39.8%。在五项推理任务中,其平均成功率在Qwen2.5-7B-Instruct和Qwen3-8B上分别达到75.9%和84.2%,同时优于Tree of Thoughts并减少验证调用及生成词元。
用途与启示
将可认证的失败解释转化为搜索控制信号,比简单重试或局部修订更准确地定位根因。
冲突驱动回跳与缓存机制可用于降低测试时推理中的验证调用量和词元成本。
为结合语言模型生成能力与符号验证器的完备搜索提供了具有理论保证的设计范式。
可借鉴SAT/CDCL求解中的冲突分析思想,改进代码生成、数学推理和约束规划智能体。
📝 原始笔记(逐字保留)
标题:CORE: Conflict-Oriented Reasoning Elimination for Verifiable Language-Model Search
来源:arXiv
链接:https://arxiv.org/abs/2609.39069
摘要:Test-time reasoning systems often respond to failure by restarting or revising the latest step, even when an earlier decision caused the error. We introduce CORE, a search controller that requests a certified conflict core from a verifier, backjumps to the latest decision in that core, and caches the conflict to avoid repeating it. Under sound verification, finite branching and depth, and exhaustive proposals, the uncapped search is complete and never prunes a valid solution. On 2,000 planted graph-coloring instances with matched proposals and an exact verifier, CORE reduces median verifier calls by 39.8% at 30 variables and 35.0% at 36 variables relative to chronological repair; caching further improves on backjumping alone. Across five reasoning tasks, CORE achieves 75.9% mean success wi
标题:SkillFM: Generating Skills for LLM Agents via Latent Flow Matching
来源:arXiv
链接:https://arxiv.org/abs/2609.39382
摘要:Textual skills provide reusable guidance for large language model agents, but existing approaches often rely on manually curated skill banks or reinforcement learning with indirect and delayed feedback. We introduce SkillFM (Skill Flow Matching), a generative framework that synthesizes task-conditioned textual skills directly without test-time skill retrieval. Our framework combines a codec for encoding and reconstructing textual skills in a continuous latent space with a conditional flow model trained using improved MeanFlow. At inference time, the learned velocity field enables single-step latent sampling, and an LLM-based decoder converts the sampled representation into textual guidance for a frozen downstream agent. We evaluate the framework on embodied tasks, question answering, and w
标题:Wavelet Flow Matching for Time Series
来源:arXiv
链接:https://arxiv.org/abs/2609.39374
摘要:Synthetic time series are increasingly used for data augmentation, privacy-preserving data sharing, and downstream model development, yet faithfully reproducing both multi-scale temporal structure and cross-channel dependencies remains challenging. We study multivariate time-series generation through flow matching in the wavelet domain. By operating on multilevel discrete wavelet coefficients rather than directly in the time domain, the model represents coarse structure and progressively finer details at separate scales. Their naturally different variances further induce an implicit coarse-to-fine generative process without requiring an explicit multi-scale schedule. Since the transform acts independently on each channel, we pair it with a channel-token transformer whose attention directly
标题:Offline Guidance, Online Reasoning: Reusing LLM Feedback for Small Language Models
来源:arXiv
链接:https://arxiv.org/abs/2609.39346
摘要:Large language models (LLMs) offer strong reasoning capabilities but are often costly to access through commercial APIs, while small language models (SLMs) are easier to deploy locally yet remain weaker in reasoning. This capability-deployment gap has motivated LLM-SLM collaboration, which aims to improve SLM reasoning using LLM capabilities while preserving the deployment advantages of SLMs. Existing approaches mainly follow two paradigms. Knowledge distillation uses LLM-generated answers and reasoning trajectories to train SLMs offline, but requires parameter updates and additional training. Alternatively, online collaboration routes difficult problems to an LLM or leverages LLM-generated guidance and corrections when an SLM encounters difficulties. Although effective, online collaborati
标题:RAIM: Robust Aggregation of Inexpensive Models for Hallucination Detection
来源:arXiv
链接:https://arxiv.org/abs/2609.39229
摘要:Automatic evaluation of faithfulness increasingly relies on a large language model acting as a judge, yet the most reliable judges are proprietary frontier models, costly and ill-suited to high-throughput monitoring. We investigate whether a panel of cheap open-weight judges (4--9B) can be aggregated to stand in for a frontier one, what the substitution sacrifices, and when it is worth making. We propose RAIM, an aggregation scheme robust to the members' correlated errors, coupling a cross-fitted stacked logistic regression with an admissibility test that, read from the members' own outputs, identifies when aggregating them improves on their best member and stays within reach of the frontier judge. We instantiate RAIM with ten judges from disjoint families across eight faithfulness benchma
标题:Fyan: A Human--AI Harness with Semantic Auditing for Document-Level Formalization
来源:arXiv
链接:https://arxiv.org/abs/2609.39228
摘要:We present FYAN, a human--AI harness for document-level mathematical formalization. Rather than treating theorems in isolation, FYAN coordinates an end-to-end workflow spanning specification, proof planning, logical review, Lean proof construction, knowledge curation, and validation, with support for independent supervision and human guidance. A central component is evidence-grounded semantic auditing, which assesses whether formal statements faithfully preserve their informal specifications. A language model constructs structured evidence over local correspondences, omissions, scope, and logical relations, while a deterministic validator checks this evidence and produces reproducible judgments. When a substantive but admissible deviation is accepted, FYAN requires an explicit proof-transf
标题:Learning Process Rewards via Reasoning State Propagation
来源:arXiv
链接:https://arxiv.org/abs/2609.39220
摘要:Process reward models (PRMs) have demonstrated notable effectiveness in test-time scaling and reinforcement learning by providing fine-grained signals for evaluating intermediate reasoning states, but their training relies heavily on costly process annotations. A natural way to alleviate this dependence is to complement limited process supervision with scalable outcome supervision. However, existing PRMs often model reasoning prefixes independently, providing no explicit mechanism for effectively using final outcome to guide the learning of intermediate reasoning states. We introduce Reasoning State Propagation (RSP), which represents each reasoning prefix with a binary validity state and models transitions between successive states across the reasoning trajectory. Specifically, RSP predic
标题:BELIEFRAG: Making Adaptive RAG State-Aware under Evolving Evidence
来源:arXiv
链接:https://arxiv.org/abs/2609.39139
摘要:Adaptive RAG uses signals such as confidence, relevance, support, and retrieval quality to decide when to search or correct evidence. In multi-step retrieval, however, these local signals must be combined into a persistent view of what the current evidence supports, what remains missing, and which action should follow. Existing methods often use such signals as separate triggers, making it difficult to preserve a coherent evidence state across a trajectory; we call this problem evidence-state fragmentation. We introduce BELIEFRAG, a closed-loop controller that updates an explicit state over sufficiency, reliability, conflict, uncertainty, evidence gaps, and acquisition cost, then chooses among retrieval, query rewriting, verification, answering, stopping, and abstention. Across six QA benc
标题:Consistent Plan-Act for Long-Horizon Agentic Tasks
来源:arXiv
链接:https://arxiv.org/abs/2609.38891
摘要:Long-horizon agentic tasks demand strong reasoning and efficient execution across successive interactions with dynamic environments. A common approach decouples high-level planning from low-level execution through separate planner and actor roles. To investigate coordination failures in these tasks, we prompt both agents for structured state assertions and compare their reports programmatically to detect explicit contradictions. Our analyses reveal systematic disagreement about the same task-relevant state facts, a phenomenon we term planner-actor state mismatch. We further find that providing agents with task-relevant state information reduces mismatch and improves coordination and task performance. Based on the systematic analysis of the state mismatch, we propose Consistent Plan-Act (Co
Bongard:训练机器直觉的开放权重 System One 模型arXiv · 认知推理 · 论文Bongard 通过表征学习与动作结果反馈训练无需展开完整中间推理的机器直觉,在决策准确率和高吞吐推理方面取得良好表现。
概率化候选输出便于接入不确定性控制、候选排序及后续慢速推理模块,构建 System One 与 System Two 混合架构。
📝 原始笔记(逐字保留)
标题:Bongard: Training Machine Intuition
来源:arXiv
链接:https://arxiv.org/abs/2609.39111
摘要:Human intelligence relies heavily on learned intuition: recognising patterns and judging situations without explicitly unfolding every intermediate step. We introduce Bongard, an open-weight System One model that treats machine intuition as an independent capability to design and train. A T5Gemma 2 4B-4B encoder-decoder separates reading the evidence from making judgments. The encoder reads the state bidirectionally together with the question instructions, and separate decoder branches share this encoding, so many judgments about the same situation require only one reading of the state. A trained head returns probabilities over the supplied candidates without generating text. Training proceeds in three stages, from supervised judgments to semantic relationships to action outcomes, and each
标题:Trustworthy Runtime Error Healing in Real-World Repositories: A Benchmark and Guardrail
来源:arXiv
链接:https://arxiv.org/abs/2609.39086
摘要:Runtime error healing lets a crashed program continue by generating code that repairs its live runtime state. Recent work shows that LLMs can generate such healing code, but it is evaluated only on small competition programs, and executing LLM-generated code inside a live process raises safety concerns that remain unaddressed. In this paper, we take LLM-based runtime healing toward practical use in real-world repositories. We first build HealBench, a benchmark of 265 runtime errors from 18 real-world repositories, each paired with a reference execution on the patched version. HealBench also provides a unified framework that lets LLM agents heal with cross-file context and live runtime state. We then design HealGuard, which requires healing code to be written in HealCore, an analyzable subs
标题:Coding Agents for Coding Theory
来源:arXiv
链接:https://arxiv.org/abs/2609.39081
摘要:We spent five weeks using an LLM coding agent on open problems in coding theory: finding large sets of four-letter words, such as DNA barcodes, that stay far apart in edit distance. The agent wrote the verifiers and search code; a human chose the problem and set the verification protocol. Restricting the search to codes with a prescribed symmetry, a classical technique, shrank the problem about fourfold and raised the best known code of length 6 and minimum edit distance 3 from 114 to 120 words ($E_4(6,3) \geq 120$). The same pipeline improved twelve further lower bounds at lengths 6 to 9 and distances 3 to 6. We give the failures equal space. Our own search stopped at 116 and recorded the last symmetry class as topping out at 112; a second agent session, running the same search with a bet
从验证失败中提炼编码智能体的可复用指导arXiv · 形式化验证 · 论文研究结合 K 框架的可执行语言语义与规格构造、证明修复和充分性审计流程,将专家对验证失败的诊断转化为编码智能体可复用的指导。
研究提出一套基于 K 框架的程序验证工具包,涵盖规格构造、证明修复以及规格充分性审计,帮助编码智能体产出可检查的正确性论证。在 HumanEval 的 164 个 Python 任务上,经过两轮针对性修复后,人工引导流程取得 164/164 的最终 AI 审计通过率。作者还构造了 12 对经人工复核的正常与缺陷验证包;尽管所有包均能通过 K 证明,完整审计仍成功识别全部缺陷并接受全部正常包。KleverBench 上针对 31 个运算符语义发生变化的程序所做实验显示,可复用指导相较完整验收规则和等长通用建议的效果受模型与预算影响,仍需研究资源约束下的指导选择机制。研究亦对 Optimism 的六项操作给出了人工复核证明,在 London 语义、无界 gas 和声明输入范围内验证了预期的暂停回滚行为。
用途与启示
展示“证明通过”并不等同于规格正确,编码智能体需要同时审计规格是否完整覆盖用户需求。
可将专家对验证失败的诊断沉淀为结构化流程,用于后续任务的规格生成、证明修复与质量审计。
为构建具备可检查正确性论证的编码智能体提供了基于可执行语义的工程路径。
提示在模型能力和推理预算受限时,应对指导信息进行选择、排序或按需加载,而非简单追加更多建议。
📝 原始笔记(逐字保留)
标题:From Verification Failures to Reusable Guidance for Coding Agents
来源:arXiv
链接:https://arxiv.org/abs/2609.39022
摘要:Coding agents need to establish that a program satisfies a specification and that the specification captures the requested behavior. We study how expert diagnosis of verification failures can become reusable guidance for this work. Our approach combines executable language definitions in the K framework with a kit of procedures for constructing specifications, repairing proofs, and auditing their adequacy. A human-guided development campaign on HumanEval, a benchmark of 164 Python programming tasks, achieves a 164/164 success rate with the semantics and the kit, measured by final AI audit Pass verdicts after two targeted repairs. To examine whether auditing detects problems that successful proofs leave unresolved, we construct 12 author-reviewed pairs of clean and defective packages. Every
标题:Smaller Models, Better Rejects: Preference Distillation Scaling
来源:arXiv
链接:https://arxiv.org/abs/2609.38987
摘要:Preference distillation typically treats a teacher response as preferred and the student's own response as rejected. This assumes that self-generated failures are the most informative negatives and that rejects must come from a model at least as large as the student, making generation costly at scale. We find neither assumption holds: across students from 7B to 72B, smaller frozen models generate rejects with less inference compute yet train stronger students than self-generated rejects, before and after sequence-level knowledge distillation, on code generation and mathematical reasoning. To explain this result, we derive a finite-horizon utility bound for Direct Preference Optimization in a linearized feature model. The bound characterizes favorable reject distributions and motivates thre
标题:PrefPI: Preference-Guided Steering into Out-of-Distribution Behaviors
来源:arXiv
链接:https://arxiv.org/abs/2609.40165
摘要:We present PrefPI (Preference-Guided Policy Iteration), an iterative framework for steering pretrained generative robot policies using only relative preferences over self-generated trajectories. Unlike prior preference-learning methods that primarily sharpen modes already represented by the policy, we study steering beyond the initial effective support, where desired behaviors are rarely or never observed under the initial policy. Our key idea is to formulate preference learning as preference-conditioned generative modeling: preferred trajectories define a conditional distribution, whose density ratio with the broader behavior prior provides an implicit preference signal amplified by classifier-free guidance (CFG). Repeating this preference-conditioned modeling and guidance step yields a f
标题:Adaptive Self-Consistency: From Black-Box Sampling to Distribution-Valued Feedback
来源:arXiv
链接:https://arxiv.org/abs/2609.38931
摘要:Self-consistency samples many reasoning trajectories and aggregates their final answers, treating the LLM as a black box that returns one answer per trajectory. Yet the final answer of each trajectory is sampled from a softmax vector that is available from the model's log-probabilities. We refer to this as the grey-box setting in which each trajectory reveals this answer distribution rather than a single draw from it. We formulate efficient inference in this setting as sequential mode identification with distribution-valued observations: sample trajectories one at a time and stop as soon as the LLM's modal answer is identified at a prescribed confidence level. We characterize the asymptotic stopping rate of mode identification with distribution-valued observations exactly and show that it
标题:Does Learning Protein Folding Generalize to Broader Reasoning?
来源:arXiv
链接:https://arxiv.org/abs/2609.38879
摘要:Large language models rely heavily on human text, which often conveys surface answers rather than the spatial and structural logic behind them. Protein folding is a natural testbed, because one solved structure yields thousands of exactly checkable spatial and topological statements. We ask: can learning to fold proteins teach general models reusable reasoning capabilities? To answer this, we build FoldingCorpus, a protein-derived question-answer dataset, and Fold2Reason, a recipe that post-trains on it through two complementary signals: discrete structural answers predicted via the model's native language head, and continuous 3D geometry decoded from the same shared representations. On FoldBench, Fold2Reason achieves structure prediction scores 2.7 to 3.5 times those of Qwen3.5-9B. Beyond
[合并自 2026-10-08 item-53]
[学习蛋白质折叠能否推广到更广泛的推理任务?(98 ▲)](https://huggingface.co/papers/2609.38879?utm_source=digest-papers&utm_medium=email&utm_campaign=2026-10-05)
标题:TRACE: Target-Aware Retrieval, Attributed Evidence, and Contract-Constrained Extraction for LitTraceQA
来源:arXiv
链接:https://arxiv.org/abs/2609.38861
摘要:Finding a relevant paper is not the same as producing a verifiable answer from it. LitTraceQA requires canonical paper identifiers, exact evidence at the page or object level, and typed answers that match the evaluator. We call the separation between source access and scorer-visible correctness the grounding contract gap. TRACE - Target-Aware Retrieval, Attributed Evidence, and Contract-Constrained Extraction - addresses this gap with target-grouped retrieval, independent typed evidence localization, multimodal table extraction, schema-driven table construction, and fail-closed validation. It indexes 27,487 papers through passage, object, alias, citation, and dense representations while retaining the question target behind each signal. For tables, TRACE predicts the observation unit before
标题:Optimal Design for Active Preference Learning with Biased LLM Judges
来源:arXiv
链接:https://arxiv.org/abs/2609.38860
摘要:Learning from human preferences is central to large language model (LLM) alignment, but human preference annotation is costly. Active preference learning reduces this cost by selecting informative comparisons, and LLM judges can provide additional scalable feedback. However, the preferences of the judges may deviate from those of the target human population. Even after calibration on trusted reference data, active acquisition can shift the comparison distribution and expose residual judge bias. We therefore incorporate judge deviations into the acquisition design rather than relying on a separate calibration stage. Under joint estimation, comparisons that appear highly informative about the reward may also reflect judge bias and therefore provide less information about human preferences. T
标题:Scoring Higher, Answering Worse: Mitigating Reward Hacking in Rubric-Based RL via Protocol-Level Rubrics
来源:arXiv
链接:https://arxiv.org/abs/2609.38847
摘要:Rubric-based reinforcement learning (Rubric-RL) trains language models where no verifier exists. A judge checks each criterion of a rubric, and the verdicts are aggregated into a reward, most often by a weighted sum. We show that this additive aggregation is the weak point. Under a sum, criteria compensate for one another: a policy that misses the one decision that matters can buy the points back with advice nobody asked for. On clinical consultation, such a policy scores higher and answers worse. Rubric coverage rises while appropriateness on held-out physician criteria falls below the untrained model. The medical criteria are not to blame. Grouped so that they must hold together, the same criteria, unchanged to the word, recover a third of the loss; shorter answers recover almost none. W
标题:Explore-on-Graph: Hybrid Embedding-LLM Reasoning for Knowledge Graph Question Answering under Incompleteness
来源:arXiv
链接:https://arxiv.org/abs/2609.39786
摘要:Large language models (LLMs) are increasingly combined with knowledge graphs (KGs) to ground reasoning in structured evidence. However, most LLM-based KGQA methods rely on traversing existing graph edges and become unreliable when reasoning paths are broken by missing facts. Alternatives that ask LLMs to generate missing knowledge risk introducing hallucinated evidence. We introduce XoG (eXplore-on-Graph), a framework for multi-hop question answering over incomplete KGs that recovers missing reasoning paths from learned graph structure rather than LLM parametric knowledge. XoG combines type-level entity-relation statistics to identify candidate relations with KG embeddings to retrieve plausible missing entities, using the LLM as a semantic selector and reasoner. These mechanisms are integr
标题:Hiding in Plain Sight: Decoupling Pretext from Actuation for Skill Poisoning in LLM Agents
来源:arXiv
链接:https://arxiv.org/abs/2609.39352
摘要:LLM agents increasingly rely on reusable Skills for complex, multi-step tasks, creating a critical supply-chain attack surface where poisoned Skill content steers agent decision loops under benign requests. Existing skill poisoning attacks either colocate actuation with its contextual pretext or distribute actuation across multiple Skills, but do not explicitly separate the rationale for execution from the operation itself. In this work, we reveal that untrusted agent decisions fundamentally depend on two conceptually distinct Risk-Realization Factors (RRFs): an actuation factor (specifying what concrete operation is performed) and a pretext factor (providing the situational rationale for why the agent must perform it). Guided by this abstraction, we propose a coordination-based attack par
标题:Who Said What, and Will It Be Remembered? Evaluating Persistent Speaker Attribution Across Meetings
来源:arXiv
链接:https://arxiv.org/abs/2609.39344
摘要:Speech transcripts used as long-term memory must preserve both words and stable speaker identities. Existing meeting-transcription metrics either ignore speakers or remap anonymous speakers independently in each recording, so they cannot measure whether the same person retains one identity across meetings. We evaluate persistent speaker attribution with Speaker Identified cpWER (SI-cpWER), which scores a corpus under one global speaker-ID assignment. The benchmark covers five commercial diarize-then-identify cascades, two open academic baselines, and ThyVoice on the full 129-meeting CHiME-8 NOTSOFAR evaluation set in clean and noiseaugmented form, plus CHiME-6. ThyVoice is our end-to-end reference system; it repairs overlap and gates the evidence used to create and update voiceprints. Requ
标题:SCB: SpeechConversationBench for Evaluating Multi-Turn Reasoning in Speech-to-Speech Models
来源:arXiv
链接:https://arxiv.org/abs/2609.40198
摘要:Speech-to-speech systems must solve tasks whose requirements emerge across conversational turns. We introduce SpeechConversationBench (SCB), a focused evaluation of spoken mathematical reasoning using 103 sharded GSM8K problems. The framework compares the original problem delivered in one turn (full), its concatenated information shards delivered together (concat), and incremental spoken disclosure across turns (sharded). We report final-answer accuracy for four commercial speech systems and LEGO, a proprietary speech pipeline developed internally by the SCBX Innovation Lab team with explicit conversational context management. Relative to concat, sharded accuracy decreases by 5.0-25.3 percentage points across the four commercial systems. LEGO achieves 77.5 percent accuracy in all three con