论文

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.

本文对 LLM 的评估方法进行了全面综述，围绕三个关键维度展开：评估什么、在何处评估、如何评估，并为 LLM 评估领域的研究者提供了宝贵洞见。This paper presents a comprehensive review of these evaluation methods for LLMs, focusing on three key dimensions: what to evaluate, where to evaluate, and how to evaluate, and offers invaluable insights to researchers in the realm of LLMs evaluation.

在 BIG-bench 上对 OpenAI 的 GPT 模型、Google 内部稠密 Transformer 架构及 Switch 风格稀疏 Transformer 进行评估，模型规模跨越百万至千亿参数，结果显示性能与校准均随规模提升而改善，但绝对水平仍然欠佳。Evaluation of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters finds that model performance and calibration both improve with scale, but are poor in absolute terms.

对 SOTA LLM GPT-4 在医学能力考试与基准数据集上进行全面评估，并通过案例研究定性探索其行为，展示了 GPT-4 解释医学推理、为学生定制个性化讲解以及围绕病例交互式构造新反事实场景的能力。A comprehensive evaluation of GPT-4, a state-of-the-art LLM, on medical competency examinations and benchmark datasets and explores the behavior of the model qualitatively through a case study that shows the ability of G PT-4 to explain medical reasoning, personalize explanations to students, and interactively craft new counterfactual scenarios around a medical case.

分析表明，借助 LLM，美国约 15% 的工作任务可在保持同等质量的前提下显著提速完成，意味着 LLM 驱动的软件将对底层模型经济影响的规模化产生实质性作用。The analysis suggests that, with access to an LLM, about 15% of all worker tasks in the US could be completed significantly faster at the same level of quality, implying that LLM-powered software will have a substantial effect on scaling the economic impacts of the underlying models.

研究揭示了 LLM 空间能力中的清晰规律：尽管其仍落后于 SOTA 方法，但具有潜力并能同时处理多种空间约束，从而可扩展到异构场景。A clear pattern is revealed in LLM spatial capabilities: while they still lag behind state-of-the-art approaches, they are promising and can handle multiple spatial constraints simultaneously, enabling scaling to heterogeneous setups.

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；但这不应被解读为模型总体上优于合格法律专业人士的证据。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 detail, should not be interpreted as evidence that the models generally outperform qualified legal professionals.

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.

本文认为，可解释 AI 研究总体上有助于推动 AI/ML 在医疗领域的落地，并特别有助于增强透明性与信任。It is argued that research in explainable-AI would generally help to facilitate the implementation of AI/ML in the medical domain, and specifically help to facilitates transparency and trust.

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.

提出 DocOps，一种确定性可验证的评估框架，基于分层分类法，将受真实实践启发的文档操作分解为原子维度与逐级递增的工作流复杂度，从而揭示 Agent 在维护全局文档一致性方面的能力边界。DocOps is introduced, a deterministically verifiable evaluation framework underpinned by a hierarchical taxonomy that deconstructs document operations inspired by real-world practices into atomic dimensions and escalating workflow complexities that exposes the capability boundaries of agents in maintaining global document consistency.

提出 ProVisE（Protocolized Visual Evaluation），一个与基准无关的框架，通过受协议约束的视觉问答从图像生成模型中抽取答案，并将其解析为与原始指标兼容的结构化预测，揭示了像素空间表达与基于文本推理的互补优势。ProVisE (Protocolized Visual Evaluation), a benchmark-agnostic framework that elicits protocol-constrained visual answers from image-generation models and parses them into structured predictions compatible with original metrics, is proposed and revealed, revealing complementary strengths of pixel-space expression and text-based reasoning.

本文提出 K12-KGraph，一个从人民教育出版社官方教材中提取的、与课程对齐的知识图谱，覆盖小学、初中和高中阶段的数学、物理、化学和生物，并由此衍生出 K12-Bench，一个包含 23,640 道题的多选基准，涵盖五类任务族，表明文本监督与视觉监督具有互补性。This work introduces K12-KGraph, a curriculum-aligned knowledge graph extracted from official People's Education Press textbooks in mathematics, physics, chemistry, and biology across primary, middle, and high school, and derives K12-Bench, a 23,640-question multi-select benchmark with five task families, showing that textual and visual supervision are complementary.

研究发现，ChatGPT 最成功地被用作数学助手，用于查询事实、充当数学搜索引擎和知识库接口；GPT-4 还可被用于本科水平的数学问题，但在研究生难度的题目上表现不佳。It is found that ChatGPT can be used most successfully as a mathematical assistant for querying facts, acting as a Mathematical search engine and knowledge base interface, and GPT-4 can additionally be used for undergraduate-level mathematics but fails on graduate-level difficulty.

EvalPlus——一个用于严格基准测试 LLM 生成代码功能正确性的代码合成评估框架，通过 LLM 与基于 mutation 的策略驱动的自动测试输入生成器，为给定评估数据集补充大量新生成的测试用例。EvalPlus -- a code synthesis evaluation framework to rigorously benchmark the functional correctness of LLM-synthesized code and augments a given evaluation dataset with large amounts of test-cases newly produced by an automatic test input generator, powered by both LLM and mutation-based strategies.

提出 OpenForgeRL，一个用于在多样化环境中端到端训练基于 harness 的 Agent 的开源框架，并在多种复杂的 harness 和环境中得到验证，涵盖工具/爪型 Agent 以及多模态 GUI 浏览器和计算机使用 Agent。OpenForgeRL is presented, an open-source framework for training harness-based agents end-to-end in diverse environments, and validated across diverse, complex harnesses and environments, spanning tool/claw-based agents and multimodal GUI browser- and computer-use agents.

结果表明，通过结合基础 LLM 改进（PaLM 2）、医学领域微调以及包括新颖集成精化方法在内的提示策略，医学问答正快速接近医生水平的表现。Results highlight rapid progress towards physician-level performance in medical question answering by leveraging a combination of base LLM improvements (PaLM 2), medical domain finetuning, and prompting strategies including a novel ensemble refinement approach.

本研究在 Reachy Mini 机器人平台上，通过系统提示、检索增强生成 (RAG) 和有状态的提示编排，将选定的 KBD 需求落地实现，表明 KBD 可以塑造负责任的机器人行为，并有望提升机器人辅助学习中的学习效果。This study operationalized selected KBD requirements in the Reachy Mini robot platform through system prompting, retrieval-augmented generation, and stateful prompt orchestration, indicating that KBD can shape responsible robot behavior and potentially increase learning effectiveness in robot-supported learning.

即使 NVIDIA 控制流 ISA 与重汇聚（reconvergence）机制持续演进，Divergence 仍能保持稳定且可预期的性能开销。Divergence retains a stable and predictable performance cost even as NVIDIA's control-flow ISA and reconvergence mechanisms continue to evolve.

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.

本文提出 InMind，一项包含 125 个任务、经专家校验、跨越十个生活领域的基准，其中 113 个任务基于可引用的公开来源；本文将此类失效模式命名为 implicit-association blind spot，并提出一种极简诊断探针（diagnostic probe），在 query 到达前保持 memory 可见即可恢复大部分性能差距。This work introduces InMind, a 125-task, expert-verified benchmark spanning ten life domains, with 113 tasks grounded in citable public sources, and calls this failure mode the implicit-association blind spot, and introduces a minimal diagnostic probe that keeps memory visible before the query arrives recovers most of the gap.

PerceptionBench 通过诊断前沿 MLLMs 在 42 项现有基准上响应中的最早失效点，并构建一个感知分支定义十种原子感知能力的错误分类法，为衡量与诊断 MLLMs 视觉感知边界提供了一项能力级（capability-level）标准。PerceptionBench provides a capability-level standard for measuring and diagnosing the visual perception boundaries of MLLMs, by diagnosing the earliest failure points in the responses of frontier MLLMs across 42 existing benchmarks and constructing an error taxonomy whose perception branch defines ten atomic perceptual capabilities.

介绍 RepoReasoner，一个用于评估仓库级代码推理的 benchmark，评估两种互补能力：Output Prediction，衡量跨文件的细粒度、有状态执行推理；Call Chain Prediction，在噪声上下文下评估高层架构依赖理解。RepoReasoner is introduced, a benchmark for evaluating repository-level code reasoning that assesses two complementary abilities: Output Prediction, which measures fine-grained, stateful execution reasoning across files, and Call Chain Prediction, which evaluates high-level architectural dependency understanding under noisy context.

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.

提出 PALATE（Person-Aligned LLM-Simulated-User Assessment with Tailored Evaluation），一个基于用户模拟器的可扩展 RPA benchmark，可针对具体 user-RPA 对给出可解释的评估，避免将系统压缩为单一、与用户无关的排名。PALATE (Person-Aligned LLM-Simulated-User Assessment with Tailored Evaluation with Tailored Evaluation), a scalable RPA benchmark built on user simulators, produces interpretable evaluations of specific user-RPA pairs rather than compressing systems into a single user-independent ranking.

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.

LlamaExtract Agentic Plus在三项指标上均排名第一，准确度可与coding agent相媲美而成本仅为其一小部分，是首个同时在大规模下对数值准确性、记录完整性、grounding与实测成本进行打分的方法。LlamaExtract Agentic Plus ranks first on all three metrics, with accuracy comparable to coding agents at a fraction of the cost, and is the first to score value accuracy, record completeness at scale, grounding, and measured cost together.

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.

CAPA通过六种机制刻画个性化编码歧义，并使用受控的三阶段生成流程将这些机制注入无歧义的可执行任务，为开发长期编码助手奠定基础，使其生成的代码更好地对齐用户意图并减少反复澄清。CAPA, which characterizes personalized coding ambiguity through six mechanisms and injects these mechanisms into unambiguous executable tasks using a controlled three-stage generation pipeline, provides a foundation for developing long-term coding assistants that better align generated code with user intent while reducing repeated clarification.

RecHarness 将优化过程分为两步：bandit 路由器根据历史验证反馈选择下一步修改方向，LLM 在选定方向内生成具体优化假设与可执行代码编辑RecHarness separates the optimization process into two steps: a bandit router selects the next modification direction according to historical validation feedback, while the LLM generates a concrete optimization hypothesis and executable code edit within the selected direction.

本文提出以交互为中心的分类法，将失败定位到其起源的交互并识别责任组件，将 41 种失败模式归到两个组件之间的边及指示修复归属的故障侧This work introduces an interaction-centric taxonomy that localizes failures to the interactions in which they originate and identifies the responsible component, and organizes 41 failure modes by assigning each to an edge between two components and a fault side indicating where the repair belongs.

在后训练中教授删除操作可减少删除回避行为并提升更广泛的代码编辑性能，表明该行为是训练不足而非不可达成Teaching deletion during post-training reduces deletion avoidance and improves broader code-editing performance, suggesting the behavior is undertrained rather than beyond reach.

提出 AntiSkillBench，一个端到端的基准，用于评估 persona-skill 流水线中的风险与防御；实验表明 persona-skill 风险在不同的 agent backbone 和蒸馏协议下持续存在，从显式属性扩展到沟通风格与个性特征。AntiSkillBench is introduced, an end-to-end benchmark for evaluating risks and defenses across the persona-skill pipeline, and experiments show that persona-skill risks persist across agent backbones and distillation protocols, extending from explicit attributes to communication styles and personality traits.

FinIndices 是一个大规模基准，在未裁剪的财务报表（最长 32K tokens）上评估数据处理保真度，带来显著的零提示增益，验证通过以数据为中心的对齐可部分恢复结构化逻辑。FinIndices is a large-scale benchmark evaluating data-processing fidelity over uncropped financial statements (up to 32K tokens) and yields substantial zero-hint gains, validating that structured logic can be partially restored via data-centric alignment.

提出 ChronoLens，结合冻结的多语言语言模型、特征对齐的 crosscoder 与事后语言学干预，应用于来自五个议会传统、跨越 1803–2026 年的 4498 万篇文档和约 172 亿 tokens，表明历史语言变化是一个结构化的、多维度的过程。ChronoLens is introduced, a framework that combines frozen multilingual language models, feature-aligned crosscoders, and post-hoc linguistic interventions, and applies it to 44.98 million documents and approximately 17.2 billion tokens from five parliamentary traditions spanning 1803--2026, showing that historical language change is a structured, multidimensional process.