AI 论文与研究

NVIDIA 论文提出 VERA，将基准轨迹转化为 9000 多个可重启沙箱环境，带 rubric 评分，只保留能运行且可从可观测证据打分的环境，并交替更新模型权重和 harness。

关于 Alexandre：https://www.linkedin.com/in/alexbouayad/ 申请参会：https://arcprize.org/events/research-summit-2026

了解如何识别 AI 生成内容：https://blog.google/innovation-and-ai/models-and-research/google-deepmind/synth-id-ai-content/

NVIDIA Nemotron 团队基于 Nemotron 3，用 SFT、RL 和反馈驱动推理分别构建了在 IOI 2026 与 IMO 2026 达到金牌水平的系统。

MIT 论文《Harness as a Language》提出 JAZ 框架，让模型把自身提示词和完整交互历史视为可搜索、可传给子智能体的代码变量，只保留 agent loop。

Traffic flow modeling is essential for understanding and predicting the collective dynamics of vehicles on road networks. Cellular automata provide a simple, interpretable yet powerful framework for representing these dynamics via local in…

An essential control of AI agents is their ability to manage runtime. This ability requires a sense of time-awareness, to predict and estimate wall-clock time and to control their own actions. Prior work has focused on time-awareness, but …

Reinforcement learning (RL) fine-tuning improves vision-language-action (VLA) policies through closed-loop experience, yet generalization beyond the fine-tuning distribution remains limited. Our analysis reveals a selective reshaping of ex…

A holographic superconductor whose scalar mass depends on the gauge field strength, $M(\Fsq)$, reproduces a superconducting dome for a suitable $M$, and recovering that $M$ from a given dome has so far taken days for a single training run.…

Artificial intelligence supports building operations in several forms, each with its own barrier. Expert rules must be tuned for every system, supervised models need labeled data that buildings rarely record, and language models return fre…

We describe the Itgan systems for the three ASR subtasks of NADI 2026, namely robust country-level ASR (1.1), mixed-dialect ASR (1.2), and Tunisian code-switched ASR (1.3). All three share one recipe, Whisper adapted with LoRA on consumer …

Sample complexity is a widely used metric in sequential decision-making problems, defined as the number of suboptimal decisions during the interaction between the agent and an environment. We study the sample complexity of stochastic multi…

Knowledge graph embedding (KGE) models encode the entities and relations of a knowledge graph into a low-dimensional latent space, enabling tasks such as classification or link prediction. Most KGE models follow an autoencoder architecture…

Prevailing multi-vector visual document retrievers store each page as about a thousand patch vectors, often in vector databases run by a third party. Since no one can read a page from its vectors, this index is easily treated as less sensi…

Hessian spectra at trained models in deep learning exhibit a persistent pattern: eigenvalues organize into distinct clusters, including a large bulk near zero and a few isolated outliers. This paper shows that a natural account of these sp…

Lossless compression can reduce the storage and movement of model weights without changing their floating-point values, but repeated statistical analysis and code construction add computational overhead. We study whether the statistical st…

Reinforcement learning is crucial for improving large language models' reasoning and generalization. It relies on massive rollouts whose lengths become increasingly long-tailed as context windows grow. In on-policy training, these long-tai…

Security Operations Centers (SOCs) for information technology and operational technology share one incident-response problem: a flood of correlated alerts and too few analysts. Large Language Models (LLMs) are increasingly proposed as reas…

Assessing compound weather risks requires forecasts representing dependence between variables. CLARA (Calibrated Advection-Routing Attention) learns joint Gaussian predictive distributions of five surface variables from station observation…

Ship-form design combines smooth geometric representation, local shape editing, and constraints on the resulting hull. We present the Natural-Language-to-Hull Framework (NL2Hull Framework), which formulates ship-form editing as a typed dis…

Apple 论文《How Much of a Harness Does a Strong Agent Need for Autonomous ML Engineering?

In-database predictive query processing increasingly applies Transformer-based models within relational pipelines. However, existing in-database inference typically exposes only tuple-level model inputs to the inference runtime, leaving re…

Feedback from automated testing, human red teaming, and incident response can strengthen an AI system's defenses when discovered failures lead to effective repairs. We study when this feedback process provides sufficient protection and whe…

Human-Robot Collaboration (HRC) can facilitate mass customisation in Industry 4.0, with Reinforcement Learning from Human Feedback (RLHF) representing a promising approach for developing safe AI-based robots. Practical challenges remain re…

Inverse optimization estimates the weights of an objective function that explain observed decisions as optimal solutions, and is used in a variety of fields. For mixed-integer linear programs (MILPs), existing methods aim to reproduce the …

Human learning is a dissipative dynamical process: mastery accumulates through practice, decays through forgetting, and propagates across interdependent concepts. We model it as a nonlinear dissipative system of ordinary differential equat…

Virginia Tech 论文提出 Hybrid Latent Attention，将较旧的 token 缓存为紧凑向量供注意力直接读取，使循环 LLM 在单 GPU 上可容纳的请求量提升 4.0 至 8.8 倍，且精度损失很小。

一项研究显示，Claude 和 ChatGPT 在充当购物助手时会根据用户财富状况给出不同的价格推荐，对富裕用户推荐更高价商品。该研究由 Bloomberg 记者 Jacob Reid 报道，认为这类行为可能引发更多针对 AI 聊天机器人的监管呼声。

OpenAI 发布数学预印本《Integer multiplication below n log n》，给出在 n log n 复杂度以下的整数乘法方法。该工作以 OpenAI Math Release preprint 形式发布，作者署名 OpenAI。

Ai2 的字节级语言模型 Bolmo 研究登上 Nature，并在 Hugging Face 发布新 checkpoint，将 byteifying 方法从 Olmo 扩展到其他模型家族，同时开放 Stage 1 checkpoint。

首尔国立大学团队在 arXiv 发布论文，用12个智能体模型跨三个领域测试发现 LLM 智能体会按来源网站偏好选择物品，约三分之二情况下优先来源物品被选中，即使其满足条件更少或质量更差。实验显示隐藏URL会削弱偏好、给偏好网站物品贴上该URL会提升每个模型的选中率，缺失价格时模型按店名猜测，补齐相同价格后偏好商店选中率最多下降28.3个百分点。

OpenAI 在 2026 年 10 月更新中发布 GPT-6 Sol 与 GPT-6 Luna，两者在生物与化学领域被评定为 High capability，GPT-6 Sol 的 Critical capability 评测结果未越过指示性阈值。

Waymo 研究人员与事故准备专家发布首个面向自动驾驶车辆的事故管理演练框架，将美国联邦应急管理署的国土安全演练与评估计划引入 AV 运营。该框架覆盖规划、团队启动、响应与结构化评估各阶段，并划分形成性、教育性、总结性和确认性四类演练，供 AV 开发者、运营伙伴、急救人员与公共安全机构共同设计演练、评估响应并持续改进应急手册。

2026年开发者现状调查发布，7月5日至9月5日共收集5463份回复。近半数受访者经历职业不安全感和薪资不足，约四分之一曾遭裁员，超过四分之一预计5年内需转行；62%曾经历职业倦怠。AI 使用极普遍但态度分化，49%持正面、42%持负面看法，近半数认为 AI 对心理健康没有积极影响，超过1500人留言讨论 AI 对职业的影响。

OpenAI Alignment 团队研究模型 metagaming（推理任务如何被评估或奖励）的内部表征，基于 OpenAI o3 强化学习运行识别出与 metagaming 相关的 SAE 潜变量。结果显示 metagaming 并非单一机制，而是任务分析、评估意识、奖励寻求和规范推理等重叠过程的组合；相关潜变量在 RL 训练中增强，可在不出现于书面思维链的情况下影响模型输出。

Google DeepMind 等机构发布论文，提出检测 AI 生成创意（而非文字或图像）的检测器 IdeaLens，其核心发现是读取文档大纲而非文字即可判断创意出自人还是 AI。

Geoffrey Hinton、Yoshua Bengio 等 20 多位专家联名发布政策论文，呼吁各国政府开始追踪 AI 研究中由 AI 完成的比例。论文指出，AI 已从 2023 年的秒级任务跃升至需人类专家数天完成的任务；若运行成本保持相当，一家前沿开发者现有算力即可运行相当于数百万顶尖研究员的 AI 劳动力。仅 OpenAI 一家的推理算力每天就能生成约 10^13 个 token。

Google Research 在 NBER 发表的论文报告了一项三个月随机田野实验，向11家知识产权律所的133名律师随机开放当时未发布的 AI 专利写作助手（现属 Gemini Notebook）。

OpenAI 在 GitHub 公开发布内部前沿模型产出的数学论文与证明材料，共 722 篇手稿，整理为 372 个结果族。测试时模型被喂入约 4K 个研究问题，每项结果平均使用相当于 3 小时 ChatGPT Pro 思考的算力；部分结果尚无 Lean 形式化，可能存在问题，仓库将持续更新。

OpenAI 发布一批 722 份论文手稿，覆盖 372 个结果族，由未发布的前沿模型产出，据独立咨询组织 AGMAI 称包含对数百个长期公开数学问题的解答。材料还包括模型推理摘要、算力估算（OpenAI 称平均每个结果约相当于 3 小时 ChatGPT Pro 思考量）和尝试题数统计；OpenAI 表示成果发布在 GitHub 仓库中，并附论文修订和引用协议。