AI 论文与研究

RoboFollow 是一个诊断基准，指出具身智能体的高成功率掩盖了其薄弱的指令遵循能力，根源在于"低场景熵"。它通过高场景熵场景、L0–L3 四级诊断协议和混淆控制评估，将理解与动作执行分离。

研究者提出 PACT（Policy Aligned Critic Training），通过 Actor-then-Critic 更新顺序对 critic 训练施加重要性采样校正，使 critic 与更新后的策略更好对齐。

MemoryAthena 提出在直接 Engram 检索（E）之外，增加从检索线索生成（GE）和不查记忆表的因果主干状态生成（GH）两条路径，并用轻量因果路由头判断生成表示何时介入。

一篇 NLP 论文主张把校准（模型置信度与经验正确率的一致性）列为 LLM 评估的一等标准，而非小众子领域。作者指出，LLM-as-a-judge、合成数据生成和主动学习等方法都依赖校准置信度却未加验证，而现有大多数 benchmark 已同时提供置信度分数与正确性判断，校准可立即报告；开放生成场景下这两个输入的界定仍是未解难题。

Agensh 是一个无需中心编排器的自组织多智能体框架，通过并发 worker 执行协作循环，自主认领子任务、共享发现并异步合并进展。

SpeakerMem-R1 提出以说话人为中心的双轨记忆系统，同时存储带说话人标注的逐字消息与派生状态，并按人物级、群组级视图组织，查询时按实体、事件和时间融合两条轨道的证据。

研究者通过标准 API 注册自定义工具，成功诱导 GPT-6 Astra 等闭源前沿模型外化中间推理过程，提取出的推理在竞赛数学、科学和代码生成任务上与原生理性表现相当，并大幅超越无推理基线。分析显示 Astra 采用 token 高效的有向推理，更早选定正确路径，内部完成基础步骤，仅外化关键推理。

研究用自建分类器 BlameBERT（F1: 0.80）结合多层统计模型，分析丹麦议会 1997 至 2026 年的指责归因，发现呈"香蕉形"轨迹：指责先下降至 2016 年前后，2019-2026 年显著持续上升。反对党指责明显多于执政党，这一"政治对比"效应受意识形态调节，右翼执政的抑制效应更弱，近年右翼的意识形态极化进一步强化了指责修辞。

该论文提出 FoMo（Forking Moment），一种全自动数据生成管线，用于在无需人工标注的情况下生成图像对之间的点式感知距离标签。方法利用扩散模型的生成动态，认为早期时间步决定粗结构，晚期决定细节；图像在生成轨迹中“分叉”越早，感知距离越远。实验表明该方法能对齐人类视觉系统，并用于训练参考型图像质量评估指标，在多个基准上超越依赖人工标注的数据集。

研究者构建了 Taste-Bench，从智能体工程与研究任务轨迹中自动挖掘决策分叉点，用于衡量 LLM 智能体在长周期任务中的"品味"。评测显示最强模型仅答对 59.7% 的题目，决定证据出现越晚的分叉越难，增大推理预算也无法提升准确率。

论文提出 AIDE²，一个让前沿 AI 研究智能体递归自我改进的系统：智能体提出对自身代码的修改，在 AI R&D 任务套件上基准测试修改版本，只保留隐藏评估表现最好的改动。

论文提出 Question's Gambit，在智能体开始搜索前运行一次：将问题拆成线索、转为互补搜索、汇总并重排结果，让智能体以已排序的上下文启动循环。

TypeSafe CEO Jev 在播客中提出 System One Model，主张 AI 应做软件内的可靠决策而非聊天优先。他称 TypeSafe 拒绝公开 benchmark 和 API 层拒答，认为数据与任务选择比堆算力更重要，System One Model 或重塑编程智能体。即便有 10 亿美元，他也不会从头预训练模型。

作者更正此前对 Grok 4.7 的兴奋评价，称真正的惊喜是 MiMo-2.6，它比 Grok 4.7 更好且便宜得多。此前他曾称 Grok 4.7 在部分 benchmark 上可与 SOTA 竞争、甚至超过 fable 5.1，且比 GPT 和 fable 5.1 便宜得多。

如果你在用AI构建产品，你应该花超过25%的时间来做基准测试，并努力让模型实验室关注这些基准测试 这是加速公司进展的最简单路径

Hugging Face 联创 Thomas Wolf 称，发布高质量开源 RL 环境是当前推动开源前沿最有影响力的事，相当于在 RLVR 新范式下共享高质量预训练数据。引用推文提到，有团队将开源约 7k RL 训练数据及框架，该模型在 AA 排名前 6，且从启动最终 RL 训练到发布模型与技术报告不到 1 周。

一项为期三个月的随机对照试验覆盖 11 家美国知识产权律所的 133 名专利律师，发现 AI 辅助显著提升专利草稿质量，初级律师获益最大，10 天和 90 天分别提升 0.58 SD 和 0.60 SD，高于资深律师的 0.30 SD。

Salesforce AI 用 Qwen3-Coder-30B-A3B 在 7 项企业任务上测试发现，把 Qwen 用 Gemini 完整轨迹微调后，成功率从 78.0% 降到 63.1%，7 项任务全部下滑，原因是 Qwen 照搬了 Gemini 的规划方式，与围绕它原有行为搭建的智能体设置不再匹配。

Modern video games provide a measurable testbed for AI models, combining abilities of visual understanding, instruction decomposition, goal planning, and precise action control over multiple temporal horizons. Existing datasets and benchma…

Multi-turn tool-use failures can hinge on a single model call, yet reward variation alone does not reveal which call would benefit from training. When rewards depend on later interactions, their variation can reflect downstream randomness …

Video world models enable interactive exploration of dynamic environments, yet struggle to respect prior observations over long horizons and across viewpoints. We present WorldCrafter, a video world model that learns a camera-queryable imp…

We present onPanda, an interactive tool for efficiently annotating LLM alignment data and agent trajectories. onPanda adopts token-level correction as its core interaction: while reading a model response, the annotator locates the first in…

On-device large language models (`LLMs'), e.g. running on mobile phones, are ripe for improvement via personalization. The limited compute resources of mobile devices impose limits on model scale and thus model quality, making any realizab…

Dexterous manipulation depends on contact dynamics that are often only partially observable from vision. Recent World-Action Models (WAMs) couple predictive video world modeling with action generation, but remain largely vision-centric and…

Agent harnesses, the external systems that mediate model-environment interaction, can substantially improve agent performance, but their gains remain tied to the harness at deployment. Because the best harness varies across domains, instan…

An LLM agent's capability is largely magnified by its harness, namely the prompts, control flow, tooling, memory, and context management surrounding the frozen backbone model. Recent methods increasingly automate this process by iterativel…

Agents today often take real-world actions that depend on long-term memory and context recall over time. However, most current memory benchmarks are built for a conversational question-answer format, where the question itself signals that …

As agents are deployed with increased autonomy, even extremely rare events along their stochastic output trajectories can occur and prove catastrophic. Safe deployment therefore does not depend on whether these events can occur, but on how…

LLM agents are increasingly deployed in collaborative settings, yet long-term interaction may give rise to undesirable coordination. We study the emergence of collusion in a long-horizon multi-agent environment: two agents repeatedly compl…

Scientific workflows often require choosing among known relations before a deterministic calculation can proceed. Whether observations share a culture, treatment or reference standard can change the scientific meaning of the resulting coun…

OpenAI 正与一个由数学家组成的独立顾问小组合作，以负责任地分享 AI 与数学领域的进展。该小组将就如何评估和传达新的数学成果、维护学术与职业标准，以及构建支持数学研究和学习的工具提供建议。OpenAI 希望让数学家处于核心位置，共同塑造 AI 如何支持数学理解并让成果惠及更广泛社区。

Visual instructions for physical tasks are typically authored in one context and followed in another, requiring users to translate demonstrated tools, materials, and spatial relationships into their own environment. We introduce Generative…

Pharmaceutical process characterization is central to Quality by Design because it defines how variations in process parameters affect the ability to meet product quality specifications, thereby supporting proven acceptable ranges and robu…

Neural operators evaluate parametric partial differential equations cheaply but degrade sharply outside their training distribution. Physics-informed neural networks avoid dependence on labeled data, yet their optimization can be basin-fra…

A model generalizes outside its training distribution only when it computes a representation structurally equivalent to the generating mechanism, not an approximation fitted to it. Such equivalence is necessary for exactness in and out of …

Despite the success of neural models in natural language processing, their black-box nature limits interpretability and conceals the linguistic phenomena underlying their predictions. We present SLITE, an explainable hybrid model for Recog…

In this paper, we study nonasymptotic $L^p$ error bounds for interval length and conditional coverage in split conformalized quantile regression (CQR). Our bounds rely on local regularity conditions and accuracy guarantees for the estimate…

Personal AI agents make recommendations and take actions on people's behalf in high-stakes economic contexts, e.g., buying a flight, choosing health insurance, or selecting a graduate program. The agent is given access to the user's person…

Scientific weak signals are early, low-visibility research directions that later become central to mature scientific topics, yet existing resources such as trend tracking, citation forecasting, and foresight reports rarely provide validate…

LangSmith 现已支持将 Jev 用作评估裁判（Jev-as-a-Judge），可对智能体 trace 进行结构化反馈评估，覆盖生产运行、数据集与回归测试场景。官方称该方式更快、更便宜。