AI 论文与研究

Physics informed machine learning provides a framework to study and simulate fluid dynamics, potentially at a fraction of the computational cost of traditional numerical methods, with the goal of improving the explainability, accuracy and …

Inverse reinforcement learning (IRL) seeks to infer the reward function underlying observed behavior, but its predictive accuracy and interpretability depend on how that reward is computed from the underlying features. While prior IRL rewa…

Deep reinforcement learning (RL) is a powerful framework for sequential decision-making. Despite many high-profile successes in recent years, deep RL is known to be brittle. An emerging line of research attributes this brittleness to a los…

World action models (WAMs) learn latent variables from observations and use these to both predict the evolution of the scene and to generate actions. The prediction has to be of the whole scene, whereas only a portion of the scene causes a…

Arena 基于 2026 年 9 月 12 日至 10 月 23 日的 2.6 万多次 Agent Arena 会话（涵盖 30 个 Agent 和 13 家实验室）分析缓存对 Agent 任务成本的影响：若缓存 token 按全价输入计费，top-10 Agent 的典型任务成本将高出 260%（3.6 倍）。

Arena 基于 2026 年 9 月 12 日至 10 月 23 日 Agent Arena 的 26,225 条会话、30 个 Agent 和 13 家实验室数据，分析提示词缓存对 Agent 成本的影响。

Epoch AI 研究人员报告，开源权重模型 GLM 5.3 在无人工输入下，利用安全补丁公开到用户更新之间的补丁空窗，对 9 月 17 日至 10 月 26 日发布的 6 个 Firefox 版本构建完整漏洞利用链，每次成功攻击 API 成本约 $1,000，最快仅 12 小时。

Epoch AI 实验发现，Z.ai 八月发布的开源权重模型 GLM 5.3 能利用 Mozilla 已发布但未推送给用户的安全补丁（patch gap），自主构建网页攻击，仅访问页面即可读取访客电脑文件，攻破 9 月 17 日至 10 月 26 日发布的全部六个 Firefox 版本。

Google 论文《Thinking-in-Bets at Test Time Reduces Sycophantic Revisions in LLM Agents》提出一种免训练方法：Agent 在读取用户反馈前先私下记录主要论断、依据和置信度，被质疑时只复核有争议的部分。

Meta 新论文提出 WM-Basin，可从蒸馏模型的文本输出中黑盒检测其是否源自带水印的 API 输出。该方法把嫌疑模型的 API 输出转成固定探针模型的梯度，再与真实密钥比对；在 1,200 个蒸馏模型、0.1% 假阳性率下，对 n-gram=1 水印的检出率达 79%，而传统 token 计数仅 29%。量化、轻度剪枝和 RL 后训练未能消除该痕迹，但在干净数据上进一步微调会削弱它。

OpenAI 发布 GPT-Rosalind-6 系统卡附录，说明该模型与 GPT-6 Astra 使用相同权重，未新增训练数据或额外后训练。在 GeneBench Pro v13 最高推理预算下，单次尝试成功率为 37.7%，高于 GPT-Rosalind-5.5 的 16.7%。

Anthropic 最新研究显示，50 位因果推断专家在模拟农业合作社政策评估中，四项分析决策采纳 AI 偏好的比例达 86%，远超随机的 50%，多数人误以为结论出自自己。

谷歌研究团队发布“睡眠框架”（Sleep Framework）持续学习研究，将模型运行分为“清醒”与“睡眠”两阶段，睡眠时通过参数扩展、混合专家记忆块和知识蒸馏把短期上下文写入长期参数，并生成合成样本做强化学习练习。实验显示模型可处理最长约 1000 万 token 上下文，在未见语言的多阶段持续学习中保持接近零的性能下降，AIME 24/25 等数学基准准确率更高。

Turning an engineering request into a simulation requires consistent geometry, physics, boundary conditions, and numerical settings. Recent agent-assisted systems automate simulation setup, but successful execution alone does not establish…

Training language-model agents to play long-horizon strategy games is hard because deterministic feedback is sparse and often arrives only at the end of a game. In-game scores can offer dense signals, but they do not reliably predict futur…

Memory has become central to self-evolving LLM agents, which store and retrieve experience from past rollouts to guide future decisions. Existing systems represent this experience as free-form natural language and maintain it by rewriting,…

We study the general problem of estimating $D(\dde(S))$, the probability that an unknown distribution $D$ assigns to an event determined by a sample $S$ of $n$ i.i.d.\ draws from $D$ through a known mapping $\dde$. The estimation algorithm…

We develop a divergence-kernel method for jointly optimizing neural SDE drift and diffusion fields and the initial Gaussian distribution, using losses that depend directly on the terminal density. In two- and eight-dimensional Gaussian-mix…

人工智能语言公平研究联盟上月发布研究发现，AI 模型处理不同语言时 Token 消耗量差距悬殊，最常用 20 种语言间相差 1.5 至 5 倍，某模型中最多相差 17 倍。这种隐性的分词器税推高低资源语言用户的成本，还会拖慢响应、挤占上下文窗口；文字系统而非使用人数是关键因素之一，专家呼吁实验室公布各语言评估结果并改进下一代分词器设计。

Sparse MoE mechanism significantly improves the trade-off between model capacity and inference efficiency. However, practical deployment still requires storing all experts, resulting in substantial memory overhead. We show that expert outp…

As Large Language Models (LLMs) become increasingly multilingual, distinguishing human-written from AI-generated text is essential for maintaining information integrity. However, research remains English-centric, with limited benchmarks fo…

Coarse-to-fine 3D-native generation relies heavily on the first-stage sparse structure, which determines the global orientation and overall geometry for subsequent refinement. However, existing coarse generators suffer from two key limitat…

Large language models offer a promising route to register-transfer-level (RTL) design, but reliable implementations require checking and revision based on electronic design automation (EDA) tools' feedback. Existing RTL-specialized models …

In pattern analysis, models trained independently for the same task often develop complementary capabilities through differences in architecture, data, and optimization. Exploiting complementarity while retaining independent deployment is …

Why does length generalization (LG) depend so sharply on the positional encoding (PE)? Relative schemes extrapolate fixed offsets, absolute schemes fix positions, and neither handles growing offsets. We give a realizability theory for a de…

Transformer-based language models have achieved remarkable success in automated short answer grading (ASAG). However, current methods in ASAG struggle to comprehend, reason about and evaluate complex mathematical answers involving symbols …

Long-tailed recognition often learns a shared representation before correcting the classifier for class imbalance. Once the classifier has been rebalanced, however, final accuracy does not reveal whether further encoder adaptation remains …

Real-world image editing requires coordinating multiple visual transformations with content preservation and reference constraints, yet aggregate scores can readily obscure unmet local requirements. Effective evaluation must therefore iden…

RGB-thermal (RGB-T) object detectors combine visible and thermal imagery for robust detection, but signal corruption in either stream can substantially degrade accuracy. Restoration-based methods incur significant latency and do not consis…

Cameras and vision sensors are widely used for product and process inspection in manufacturing environments. Traditionally, this process required thousands of images of both nominal and defective products to train such models. Recent resea…

Multimodal federated learning (MFL) trains a shared model by aggregating client updates without exchanging raw data. Under missing modalities, clients adapt active modality encoders and a shared prediction block (fusion and classifier) to …

LLM agents increasingly triage security alerts by combining the outputs of several detection tools, implicitly treating those tools as trustworthy. Unlike prior work on conflicts between a model's knowledge and its context, we study confli…

Neuralink 于 10 月 28 日公布语音恢复项目 VOICE 的早期可行性研究，让 ALS（渐冻症）参与者将大脑信号转换为文字和语音输出。一名参与者在使用第一周达到每分钟 75 词，在 12.5 万词词库下词错误率为 5.8%；研究计划招募约 6 名因 ALS、脊髓损伤或中风失去沟通能力的参与者。目前结果来自单一参与者，尚未发布经同行评审的完整论文。

End-to-end (E2E) autonomous driving models achieve strong planning performance but remain largely opaque, limiting observability and trust. We apply mechanistic interpretability to E2E driving by training Sparse Autoencoders (SAEs) on the …

Google DeepMind 发布论文《The Economics of Intelligence》，由 Alex Imas、Andrew Koh 与 Shane Legg 合著，探讨 AGI 如何以及何时改变经济。

LLM-guided program evolution can produce mathematical constructions, but validity, generality, and reproducibility require distinct evidence. We develop a researcher-in-the-loop discovery pipeline that combines program evolution, exact eva…

这真是下一个级别的东西。 RL with Neural Feedback（RLNF，神经反馈强化学习） 一个AI图像模型通过读取你看图时大脑的反应，学习你个人觉得有吸引力的东西，然后修改图片以获得更高的分数，全程无需你打分、点击或写提示词。

UC Berkeley 与华盛顿大学的新研究发现，截至 2026 年 8 月约 24% 的可爬网页含 AI 生成或 AI 编辑文本，而 2022 年前几乎为零。

Princeton 与 Meta FAIR 的论文提出 Self-Direction，让智能体在失败后反思轨迹并将经验训练成约占基础模型 0.02% 参数量的 LoRA 适配器，之后由路由器在遇到类似任务时加载。用 Qwen3.6-27B 在 GAIA 上任务成功率从 38.8% 提升到 55.8%；185 个早期错误中 67.3% 的修复在第三次遇到时仍有效，而简单重试为 42.9%。

Anthropic 一项针对 50 名因果推断专家的实验发现，AI 智能体被掷硬币随机指定的偏好会被专家在四项分析决策中最高 86% 采纳，净影响即 flip rate 在 62% 到 82% 之间；先有自己观点的参与者面对相反意见也更易改主意，经验、学历、自评专业度等均无可靠保护作用。