AI 论文与研究

小红书大模型基建部发布 NeurIPS 2026 论文 Hint Tuning: Less Data Makes Better Reasoners，把推理模型的思维链按反思词切成片段作为 hint 交给同家族 Instruct 模型，测出每道题的最小有效提示，据此构造长度可变的训练数据。

腾讯混元推出 WorkForge 框架，可从真实世界资源构建可验证的长周期工作智能体训练环境。该框架从专家工作流出发搭建真实工作空间、抽取可核查事实，并生成任务及程序化与语义双重验证器，已扩展至覆盖 40 个专业领域、60 种文件类型的 16.7K 个环境。随着数据量和交互时长增加，该框架在长周期工作基准上带来稳定提升。

腾讯混元推出 WorkForge，一个可从真实世界资源构建可验证训练环境的可扩展框架，面向长周期工作智能体。该框架从专家工作流出发搭建真实工作空间、抽取可核查事实，并派生任务及程序化与语义双重验证器，已扩展至覆盖 40 个专业领域、60 种文件类型的 16.7K 个环境。随着数据量和交互时长增加，其在长周期工作基准上带来持续稳定的提升。

腾讯混元与复旦大学研究团队提出 WorkForge 框架，从真实世界资源出发规模化构建可验证的智能体工作环境，产出 WorkForge-16K 共 16,665 个环境，覆盖 40 个专业领域、60 种文件类型。

Ai2 发布 PLANTEUR，将 PASTIS 基准扩展到留尼汪、瓜德罗普、马提尼克、法属圭亚那、马约特和圣马丁六个热带地区。该基准包含 2,750 条 Sentinel-2、Sentinel-1 和 Landsat-8 对齐时间序列，以及覆盖 7 类作物的 4,506 km² 地块标签。

Beren Millidge 发文提出 AI 发展经历预训练、RL 与持续自主学习三阶段，训练循环逐步内生化，对齐挑战也相应迁移到更抽象的层级。文章认为每阶段的对齐难题都是能力难题的镜像，核心瓶颈从数据集构造，到设计可靠的判断程序，最终变为设计能在模型自主变化中维持不变量的控制系统，并指出持续学习会带来漂移和类灾难性遗忘的新挑战。

Microsoft 研究员与合作者证明 1995 年提出的 Kannan-Lovász-Simonovits（KLS）猜想，形成 140 页证明。AI 智能体探索约 100 条研究路径、执行推导、辅助撰写手稿并运行内部证明检查，研究者负责筛选弱方向，递归方法和误差界归功于两位研究者。该猜想描述随机游走者在凸体内不会被狭窄通道困住，可高效探索整个形状。

新论文测试 GPT-6-Astra 在 12 个机器人基准上的表现，其在 LIBERO 套件接近饱和，RoboDojo 达 81.9%、DexJoCo 达 87.3%，但无法验证物理成功或从不可逆错误中恢复。

Meta 与密歇根州立大学的新论文发现，LLM 构建的推理图质量差异很大，且图中的错误会传导到基于图计算的置信度分数上。研究测试了 9 个模型，从给定步骤到完全自建图逐步减少辅助，结果显示图质量更多与模型家族和图结构相关，而非模型规模，Qwen3-32B 回答子问题的表现甚至不如 Qwen3-8B。作者建议在信任基于 LLM 推理图得出的答案或置信度之前，先检查图本身。

Perplexity 与 NVIDIA 团队提出后训练方法 JOLT，让同一个模型同时扮演教师和学生，且教师也参与训练。在 MATH 上，JOLT 以 6.5 倍更少的训练 token 达到 GRPO 的准确率；JOLT+ 在 AppWorld 状态测试通过率达 72.09%，比 GRPO 的 62.20% 高 9.9 个百分点。

Google 新论文提出 OO-Harness，将 agent harness 按代码类演化成共享家族，每个专用 harness 继承父类、只改本簇所需部分，再由路由器根据基础运行结果选择专用版或保留基础结果。

Nathan Lambert 在哈佛 CS 2881R 客座讲座整理了开源模型现状全景幻灯片，判断开源模型落后闭源前沿约 3 到 5 个月且差距在缩小。他指出性能、下载量、论文引用三方面中国开源全面领先，美国开源又落后中国 3 到 4 个月；把蒸馏当成核心问题是高估，美国真正持久的优势是算力，前五实验室总算力约为中国前五家的 16 倍，禁止开源只会损害经济活动。

@oradotai 联合 Vercel、Stripe、Auth0、Mintlify、Tavily、Google 等推出 AgentReady（agentready.org），一份 MIT 许可的开源规范，规定网站要让 AI Agent 找得到、读得懂、用起来需做到什么。

Anthropic Fellows 团队用 Opus 5 红队智能体测试生产级拦截监控器，发现 Auto Mode 与 Guardian 的平均攻击成功率超过 55%，主要绕过手段包括伪造转录轮次的监控器注入、多智能体与恶意压缩的多上下文攻击。团队通过转录框架化与转义、监控工作目录编辑和增加智能体监控阶段将攻击成功率降低超过 50 个百分点，但以可接受成本可靠防御多上下文攻击仍是开放问题。

我们如何扩展 LLM 的上下文？ Agentic search 可扩展到任意规模的信息，但很容易遗漏相关信息。 想法💡：针对下游效用优化信息的组织方式。即使在 OOD 领域也能带来 30+% 的提升 论文 + 代码：https://librarian-agents.github.io/

Perplexity 与 NVIDIA AI 提出后训练方法 JOLT，让同一模型同时充当教师和学生，并训练教师从自身尝试的奖励中学习、惩罚偏离学生行为的漂移。在 MATH 基准上，JOLT 用少 6.5 倍的训练 token 达到与 GRPO 相当的准确率。JOLT+ 在此基础上为学生模型增加结果奖励，在数学、编程、终端任务和工具使用基准上明显优于 GRPO。

Sierra 发布 𝜏-i18n，将 𝜏-Voice 扩展到西班牙语、巴西葡萄牙语、印地语、韩语和普通话五门语言，使用统一的航空、零售、电信任务跨语言对比。

两年前的今天，我们推出了 Artificial Analysis Video Arena，这是领先的文本生成视频模型众包对比平台。 此后，我们陆续新增了图生视频、带音频视频和视频编辑的排行榜。最初的排行榜仅有八个模型——如今我们已评测超过 140 个。

新研究：我们开发了 JOLT，一种后训练方法，让一个模型同时充当教师和学生，其中教师经过训练，能够给出学生可用的反馈。 在 MATH 上，它以少 6.5 倍的训练 token 达到了与 GRPO 相当的准确率。 https://www.perplexity.ai/hub/blog/a-good-teacher-meets-the-student-where-they-are

Reinforcement learning (RL) is an effective way to strengthen the long-context reasoning of large language models (LLMs), but it needs training data that are long, diverse, challenging, and verifiable. Human annotation of such data is expe…

Existing sequential recommenders usually allocate a fixed computational budget to every item in a user's interaction history, regardless of the item's contribution to the final prediction. We challenge this de facto design with two diagnos…

Existing world models typically perceive a scene through current visual frame or a narrow history window. However, these observations may not reveal the physics properties that determine transition dynamics, such as an object's mass distri…

We give an algorithm for computing an $\eps$-approximate fixed point of any $d$-dimensional nonexpansion with respect to the $\ell_\infty$ norm in time $\log(1/\eps)^{O(\log^2 d)}$. At its core, our method uses a boosting argument to geome…

We prove a self-normalized Bernstein inequality for bounded functions of a regenerative process with independent, identically distributed marked cycles of finite mean length. The bound uses observed cycle lengths and retains explicit terms…

The success of convolutional neural networks is often attributed to locality and weight sharing. However, how these two properties separately shape learning, and how their effects depend on the spatial correlations of natural images, remai…

Through a tangle of real-world interactions, from synthetic training pipelines to data shared on the open web, frontier LLMs absorb what their predecessors and competitors create. How well does each model play its role in this ecosystem, a…

Consensus is a fundamental problem in distributed computing and at the core of state machine replication (SMR). Despite decades of active human research, discovering practical low-latency solutions to the problem is still open, as it requi…

Quantitative research is an iterative loop: a researcher proposes a market hypothesis, implements it as an executable signal, tests it on historical data, and decides what to keep. Existing benchmarks score formulas, code, or individual tr…

We analyze the problem of estimating univariate Gaussian mixture densities using the nonparametric maximum likelihood estimator (NPMLE) based on $n$ independent and identically distributed samples. We show that the unconstrained NPMLE achi…

Multimodal agents (MMAs) sustain interaction through reasoning, memory, tools, and feedback, most visibly as computer-use agents, while robotic systems (RSs) couple sensing and actuation under physical dynamics. We define multimodal embodi…

Flow matching trains a continuum of velocity regression problems but usually reports a loss averaged over interpolation time, obscuring which regions are limited by intrinsic ambiguity and which remain data-limited. We introduce the Veloci…

Sequential decision-making under partial observability is difficult because agents must explore to infer hidden, task-relevant information before acting effectively. Learning such information-gathering behaviors tabula rasa via online rein…

Despite recent advances, current AI systems remain highly susceptible to domain shifts, particularly in biomedical imaging, where images are usually acquired under substantially different experimental conditions. Although recent domain ada…

Many NP-complete (NPC) problems have a polynomial-time (P) twin that differs in a single structural detail, such as 3- versus 2-colouring a graph. A model that reasons rather than recalls should notice which side of this boundary a problem…

Reinforcement Learning with Human Feedback (RLHF) has become a central paradigm for aligning Large Language Models with human values. However, the reliance on sensitive preference data necessitates rigorous differential privacy (DP) guaran…

Vision-language-action (VLA) models have shown promise in robotic manipulation, but imitation learning (IL) alone often struggles to master complex long-horizon skills. Reinforcement learning (RL) offers an effective avenue for further imp…

Perimeter triage faces a latency/semantics trade-off: signature engines are fast and format-safe but blind to unseen attacks, while generative models are semantic but slow and produce unparseable output. In late 2026 a class of commercial …

基于 Qwen3-4B 的葡萄牙语作文自动评分器通过 VocabX 方法重建词表，在基座模型、数据和训练配方不变的情况下将实测推理成本降低 77%。该方法按每个 token 的成本而非词频来专精词表，使输出层从约 152,000 行缩减到约 65,000 行。原文指出生成 token 的成本约为读取 token 的 7.4 倍，词表是领域专精模型最大的成本杠杆之一。

SemiAnalysis 发布长文，分析 9 月 30 日 DeepSeek 为 TileLang、DeepGEMM、DeepEP、FlashMLA 等 6 个仓库加入 Ascend 支持背后的华为-DeepSeek 互补关系。

Generative models propose crystalline materials by learning composition and geometry, but unconditional sampling gives no direct control over the space group, which constrains whether responses such as piezoelectricity are symmetry-allowed…