论文

本文命名并研究这两条研究脉络之间的模式：其递归单元是配备文件系统工具、代码执行与规划的完整 Agent harness，而非无工具的模型调用，并给出针对长上下文推理的受控评估。This work names and studies the pattern between these two lines of work, where the recursive unit is a full agent harness with filesystem tools, code execution, and planning rather than a model call with no tools, and provides a controlled evaluation on long-context reasoning.

一个两级框架将手动 harness 工程转变为自动化 harness 工程，并更进一步——将"自动化本身的设计"也自动化。A two-level framework shifts manual harness engineering into automated harness engineering, and takes one step further --automating the design of the automation itself.

提出 Agentic Harness Engineering（AHE），一个通过三个相互匹配的 observability 支柱应对 harness 工程挑战的闭环，将每一次编辑转化为可证伪的契约，使 harness 演进能够自主进行而不退化为试错。Agentic Harness Engineering (AHE) is introduced, a closed loop that addresses harness engineering challenges through three matched observability pillars that turn every edit into a falsifiable contract, so harness evolution proceeds autonomously without collapsing into trial-and-error.

本文提出 Natural-Language Agent Harnesses，即可编辑的、描述运行级 harness 策略的文档，以及 Intelligent Harness Runtime（IHR），一个将上述文档解释为 agent 调用、交接、状态更新、验证门控与 artifact 契约的共享运行时。This paper introduces Natural-Language Agent Harnesses, editable documents that describe run-level harness policy, and Intelligent Harness Runtime (IHR), a shared runtime that interprets these documents into agent calls, handoffs, state updates, validation gates, and artifact contracts.

本文通过大规模超参数搜索，系统地重新评估了 Vanilla LoRA 以及九个代表性 LoRA 变体，发现不同 LoRA 方法偏好的学习率区间各异，并将最优学习率区间的差异归因于最大 Hessian 特征值的变化，与经典学习理论相吻合。This work systematically re-evaluate nine representative LoRA variants alongside vanilla LoRA through extensive hyperparameter searches, finding that different LoRA methods favor distinct learning rate ranges and attributes the differing optimal learning rate ranges to variations in the largest Hessian eigenvalue, aligning with classical learning theories.

结果表明，强大的 LLM 能够产出一系列合理思路，但其范围仍比人类研究品味更窄，并存在系统性偏移。It is suggested that strong LLMs can produce a range of reasonable ideas, but that range remains narrower than, and systematically shifted relative to, human research taste.

介绍其潜在原因的理论基础，阐明强化学习算法的渐近性能在性能排名与数据规模之间不存在单调关系。The theoretical foundations of the underlying causes outlining that the asymptotic performance of reinforcement learning algorithms does not have a monotone relationship between performance rankings and data-regimes are introduced.

压缩可减少推理 token 并保持大部分选择题准确率，同时提示影响接近基线——在一个前沿条件下，降低推理成本移除的证据比单纯缩短轨迹所预期的更多。Compression reduces reasoning tokens and preserves most multiple choice accuracy, while hint influence remains near baseline, in a frontier where reducing reasoning costs removes more evidence than shorter traces alone would predict.

本文提出 UniVR，这是首个从纯视觉演示中同时学习复杂推理、细粒度物理动力学和长期规划的研究，并配套首个在纯视觉协议下评估这些异质能力的综合评测套件。UniVR, the first investigation into simultaneously learning complex reasoning, fine-grained physical dynamics, and long-term planning from pure visual demonstrations, is introduced, the first comprehensive suite to assess these heterogeneous capabilities under a purely visual protocol.

发现对 GPT 语言模型进行重复采样，是为困难 prompt 生成可用解的一种出乎意料的有效策略，并讨论了部署强大代码生成技术在安全性、安全保障与经济等方面的潜在更广泛影响。It is found that repeated sampling from the GPT language model is a surprisingly effective strategy for producing working solutions to difficult prompts, and the potential broader impacts of deploying powerful code generation technologies, covering safety, security, and economics are discussed.

在该研究的评估协议下，部分 AI 生成回答获得了高于可用参考答案的评分，尤其当其包含准确且相关的支撑细节时尤为明显。Under the study's evaluation protocol, some AI-generated responses received higher ratings than the available reference answers, particularly when they contained accurate and relevant supporting details, particularly when they contained accurate and relevant supporting details.

本文提出 Verbatimize 这一新任务，能够以高质量规范化逐字转录对语音语料库进行可扩展的创建与扩充，并通过有监督的 cross-attention 微调，将不流畅语音上的词级时间戳表现提升至超过强制对齐 baseline。Verbatimize is proposed, a new task enabling scalable creation and enrichment of speech corpora with high-quality canonical verbatim transcriptions and supervised cross-attention fine-tuning that improves word-level timestamps on disfluent speech beyond forced-alignment baselines.

FinMMEval 2026 Task 2 围绕多语言证据评估金融领域的短答问答，采用文档 RAG、跨语言证据处理、结构化提示、答案压缩与验证策略。FinMMEval 2026 Task 2 evaluates short-answer financial question answering over multilingual evidence over multilingual evidence using document retrieval-augmented generation, cross-lingual evidence handling, structured prompting, answer compression, and validation strategies.

本文提出 PAJAMA，该系统将程序合成为 judge，将其决策聚合为联合裁决（joint verdict），并通过 fallback 机制选择性地将低置信度用例升级交由 LLM 处理。PAJAMA is introduced, a system that synthesizes programs as judges, aggregates their decisions into a joint verdict, and incorporates a fallback mechanism to selectively escalate low-confidence cases to an LLM.

介绍 DecoEvo（Decoupled Co-Evolution），在解耦目标下协同进化一个求解器 skill 和一个 rubric 生成器 skill，优化过程中不使用 gold rubric。DecoEvo (Decoupled Co-Evolution), which co-evolves a solver skill and a rubric-generator skill under decoupled objectives without using gold rubrics during optimization, is introduced.

本文提出 CoRT，一种用于 rubric 条件化 GRPO 的 token 级 credit 加权方法，通过反事实回放在原始 rubric 条件化 prompt 和匹配的免准则 prompt 下对同一样本响应重新打分，并表明策略内部反事实似然对比为响应内 credit 分配提供了有效的训练信号。CoRT is proposed, a token-level credit weighting method for rubric-conditioned GRPO that uses counterfactual replay to rescore the same sampled response under the original rubric-conditioned prompt and a matched criteria-free prompt, and suggests that policy-internal counterfactual likelihood contrasts provide an effective training signal for within-response credit allocation.

对 Fairness Pruning 的实证评估表明，群体偏置处理与模型能力运行在可分离的电路上，奠定了从盲目零化向定向行为调制过渡的方法论基础。Empirical evaluation of Fairness Pruning empirically confirm that demographic bias processing and model capabilities operate on dissociable circuits, establishing the methodological foundations for transitioning from blind zeroing toward directional behavior modulation.

多维评估-验证奖励(EVR)将评估分解为独立视觉准则；针对每个准则，MLLM Evaluator生成多个候选假设，Verifier在具体视觉证据中grounding每个claim以接受或拒绝，产生可靠且细粒度的奖励信号。A Multi-dimensional Evaluation-Verification Reward (EVR) decomposes evaluation into distinct visual criteria; for each criterion, an MLLM Evaluator generates multiple candidate hypotheses, and a Verifier grounds each claim in concrete visual evidence to accept or reject it, producing reliable and fine-grained reward signals.

本工作提出 BERTScore——一种文本生成自动评估指标，与人类判断的相关性更强，并在模型选择性能上优于现有指标。This work proposes BERTScore, an automatic evaluation metric for text generation that correlates better with human judgments and provides stronger model selection performance than existing metrics.