AI 论文与研究

Google 的 Logan Kilpatrick 发推称已有许多 Argon 的实际使用示例，来自 Google 内部、客户和合作伙伴，并附上博客链接 https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/。链接指向 Gemini 4 Argon 的介绍页面。

Artificial Analysis 发布 AA-Robotics v0.1，用 8 项实体任务评测具身智能，包括按红色按钮、把鳄鱼玩具放进矩形、倒饮料、整理积木、形状匹配、写 hello world、搭最高塔和测量盒子。

Cohere Labs 在 ProgramBench 200 个任务上系统测试编码 Agent 团队的设计选择，发现团队优于单个 Agent 但成本高 4 到 7 倍。

Multilingual dense retrievers place all languages in one shared space so that a question finds relevant documents, regardless of which language they are written in. We show that the same question, asked in different languages, retrieves su…

Task planning for robots in off-site construction prefabrication is primarily manual, leading to planning inefficiencies and reluctance to adopt robotic solutions. Existing approaches provide limited support for scalable, modular automated…

Reusable skills let LLM agents load task-specific instructions on demand, but they make execution depend on implicit routing over skill metadata. We test whether this design scales by holding SkillsBench tasks and their golden skills fixed…

Autoregressive language models produce, at every prediction step, a vocabulary-sized logit vector whose structure determines the conditional next-token distribution. We empirically identify that the ordered vocabulary-head logits follow a …

Knowledge distillation trains student models on teacher-generated responses to transfer task capability, but those responses may also carry behavioral tendencies of their source. We investigate whether the source teacher can shape student …

Post-training pruning enables efficient deployment, but conventional methods do not explicitly preserve internal-gradient alignment between sparse and dense models. We study novel connections between downstream performance of compressed mo…

Large language models (LLMs) increasingly act as personal assistants; however, little is known about how they capture national-level preferences on a global scale. Given that "time" is a universal and representative concept, we examine LLM…

Motivated by applications to cognitive radio networks, multi-player multi-armed bandits (MMAB) have attracted considerable attention in recent years. The asynchronous variant, in which players are active only at a fraction of timesteps, be…

Buildings cause nearly 40\% of carbon emissions in the US. Of total building emissions, HVAC systems are responsible for approximately 40-60\% and account for 15\% of global energy consumption. Despite advances in artificial intelligence, …

The safety of LLM-based agents currently rests on alignment training and safety classifiers, neither of which can rule out that a sophisticated indirect prompt injection (IPI) goes through and causes significant harm. Recent work therefore…

Clinical trial protocols serve as the source of truth for study design, but are mostly written and used as unstructured PDF documents. Digitizing protocols turns static content into structured, machine-readable data that clinical software …

Theory 发布 2026 AI in Practice Survey 结果，455 名资深技术从业者参与，安全与治理从第六位升至首要痛点，占 43%，评测以 22% 居次。

Recently, Morris et al. introduced TinyLoRA, a strategy for LLM fine-tuning which introduces arbitrarily few parameters, and demonstrated accuracy improvements using less than $10$ parameters rather than the ${\gtrsim 10,000}$ typically de…

Egocentric video understanding spans diverse capabilities, modalities, and evaluation formats, yet these are often studied in isolation despite their tight coupling in real-world interaction. We propose the Ego-Action-World interaction loo…

Autoformalization (AF) requires more than compilation: formal statements must preserve the meaning of their informal sources, a property known as faithfulness. An emerging frontier is the automated evaluation of this property, yet most exi…

Tool-call judges evaluate argument values proposed by LLM agents. We find that they accept a correct value less often when their input holds no reference for it, that is, no text that says what the value should be. This happens when the va…

In order to execute quantum algorithms on bosonic quantum computers, it is necessary to encode their inputs into a desired target quantum state, with a process called state preparation. In real settings, state preparation methods operate i…

Objective digital measures of depression severity could complement conventional symptom assessments and facilitate more timely clinical evaluation. Here, we investigated whether task-evoked eye movements provide interpretable behavioral ma…

Many related prediction problems, which we call questions, come with few labels each. What information can be shared across them, and what is actually necessary for good decisions? We answer both for recalibration, motivated by decision mo…

Membership Inference Attacks (MIAs) are currently the dominant approach to evaluating privacy in machine learning applications. Despite their significance, several concerns remain unexplored, particularly with regard to tabular data. Exist…

Terminal agents have shown strong abilities in complex coding tasks. However, they often struggle to recover from errors during long-running execution (e.g., command errors, faulty implementations), eventually leading to task failures. To …

IBM Research 演示用开源前沿模型结合 llm-d 分布式推理构建 agent factory，对量子纠错解码算法 Relay-BP 做 HLS 硬件优化，在 100 个假设预算下实现约 52 倍加速、推理成本约 26 美元；同预算下全专有模型配置成本 299 美元、加速 204 倍。

Modal 工程团队展示在 Qwen3.6-35B-A3B 的 agentic coding RL 训练中，用跨六种加速器（B200、B300、H100、H200。

Anthropic 研究人员测试发现，小型 AI 模型已能在廉价消费设备上部分完成自主复制，其中 GLM-5.2 在五次运行中有四次可靠感染全部八个测试节点。

小红书 AllSpark 团队联合社区审核算法团队提出 Agent 学习框架 AdaptEvo，将置信度自适应强化学习 CA-GRPO 与决策知识、评价准则的演化交替进行。

Composo 开源 DishwasherBench v0.1.0，基于一台 Bosch SMV4HAX40G 的物理模拟，让模型和人类在浏览器中通过 MCP 比拼洗碗机装载，按 Eco 50 洗后干净程度计分。

Anthropic 于当地时间周一发布实验结果称，再过数月小型 AI 模型或将具备在商用物联网设备上自我复制的能力。实验发现参数量 300 亿以下的小模型可完成复制流程中的多项步骤，但难以独立跑完整链路；在移除部分安全校验的简化环境下，GLM-5.2 及更强模型能实现端到端完整自我复制。

NeurIPS 2026 接收的 arXiv 论文（2610.13928）构造后门攻击，在宿主模型权重中嵌入加密神经网络，仅在收到秘密触发器时解密执行。这些隐藏电路在密码学假设下可证明难以被诱导和解释，即使完全访问模型权重。

YouTube 频道 Jun Yamog 测试将 4 张英伟达 Tesla V100 装于 NVLink 载板，在 4-bit 量化 Qwen3.8-Flash-Next 基准中长提示处理速度较 2 台 DGX Spark 快 40%、较 M5 Ultra 快 230%。

Apple 研究团队在 150M 到 1.43B 四种模型规模、约 1000 次预训练实验中对比发现，混入高资源语言数据比超参数调优更有效，且优势随模型增大而扩大。混合数据带来的提升相当于 2–3 倍目标语言独有数据（验证损失）和 2–13 倍（下游任务准确率）。团队建议优先调整混合比例，并通过 μP 从小代理模型迁移超参数。

Preferred Networks 旗下 AI 面试服务「タレスカ」对 308 名 2027 届就活生的调查显示，认为 AI 面试公平的比例仅 51.9%，远低于面对面面试的 85.4% 和在线面试的 76.6%，AI 面试受检者中该比例为 57.4%。

LG AI Research 在 ICML 2026 发布 GEAR 与 Sim-GEAR，用黎曼几何连接不同物理性质预测模型的隐空间，解决材料研发中部分物性仅有数十条数据、而另一些有上亿条的数据失衡问题。该方法不再假设隐空间是平坦的欧氏空间，而是把模型间的差异视为空间的几何差异，从而在源任务与目标任务之间迁移知识。

过去几个月，我们在降低针对 Claude 的蒸馏攻击有效性方面取得了重大进展。 在这篇文章中，我们解释了具体做法：https://www.anthropic.com/research/distillation-defenses

Artificial Analysis 推出 AA-3D 基准，让 21 个模型作为智能体在离线沙箱中把书面简报转化为 three.js 网页，由超 1000 名评委盲评产出逾 10000 票的 Elo 排名。

阿里 Qwen 团队推出 QwenGyre，一个面向超长时程智能体训练的弹性强化学习框架，通过在线 RL 在未经修改的黑盒 harness 中训练编码智能体，单次 rollout 可运行数小时、处理近 1M tokens。

AI 在承认需要澄清时表现更好；提升机器人在视觉-语言任务上的表现；如何从 AI 的失败中学习；改进 AI 辅导；理解我们的嗅觉；更高效的 KV-cache 内存存储 https://msft.it/6019aSVIP

Perplexity CEO Aravind Srinivas 转发 Meta 新前沿模型发布公告并祝贺 Meta 训练出前沿级模型。其引用内容称该模型在个人智能体和专业智能体上是 SOTA，是其最大、最强的模型；作者本人补充称 Watermelon 在 WANDR 评测中与 Fable 5.1 相比表现不差。