AI 论文与研究

METR 发布基于 Anthropic 内部信息的风险评估报告，评估 AI 是否在加速 AI 研发。团队历时 11 周分析内部研发数据并访谈核心研究员，估计 Anthropic 当前模型改进速度约为 2024 至 2025 年节奏的 2 倍，其中约 80% 的加速归因于其自身 AI 使用，并排除了训练算力更快增长的解释。

OpenAI 与酶设计公司 Ligo 共同发布 BioBasicsBench 基准，用 100 个任务衡量 AI 智能体解读原始实验数据、发现错误并识别证据是否足以支持结论的能力，覆盖蛋白质方法、成像与流式、定量与检测、核酸方法和分子克隆五个领域。

vLLM 博客介绍 NVIDIA Research 与 Inferact 的 REINFORCE Zero 方法：对齐 rollout 与训练内核后，用不加截断的 token 级重要性采样即可稳定异步 RL。

Base Labs 在 11 月 6 日的研究更新中指出，on-policy self-distillation（OPSD）虽在部分场景仍有用，但不值得继续投入研究。作者认为，OPSD 的教师模型已知道答案，无需检查、回溯或保留余地，这种稠密损失主要带来偏差，且可能只保留旧知识而学不到新知识。

Johns Hopkins 新论文发现，LLM 训练数据中只出现过一次的事实（monofacts）占比越高，科学发现率越高。

语音识别评测必须测试语音智能体真正发挥作用的场景，也就是两人之间的电话通话。 Voice Arena 的新基准在真实通话中测试语音转文字模型——用日常手机、在真实房间里、有真实背景音。 两个人说话会让模型的任务难得多。人们会互相抢话、打断、插入简短的“yeah”和“okay”回应，而且电话线路会压缩音频。客服或销售智能体每天都要应对这一切。

Dwarkesh Patel 花费约 20000 美元 API 费用，让多个 AI 模型进行了 386 场 Blood on the Clocktower（染·钟楼谜团）对局，评估模型的长期策略、多智能体协调与欺骗能力。

斯坦福论文指出，没有任何指标能在所有意义上与分词器无关，因此比较不同分词器时不能只信单一分数，还要检查模型是否答对。研究用未训练模型和同一配方不同分词器的已发布模型测试：什么都没学到的模型 BPB 仍随分词器变化，英文文本在约 24K 词表时 BPB 最好，但准确率仍为随机水平。在不同分词器训练的模型上，BPB 与任务分数的相关性仅在 +0.02 到 −0.30 之间。

美国非营利机构 AItonomy Foundation 推出 Science's Last Exam（SLE）基准，用专家预先封存、对智能体隐藏的检查项，评测研究智能体可打开并复跑的研究产物。

SemiAnalysis 发布首批 TPUv7 在 InferenceX/AgentX 长上下文、多轮智能体推理基准上的官方结果，覆盖 Muse Spark Open、GLM-5.2、Qwen3.5 397B 和 Kimi K3 四款模型。

User embeddings are central components of modern recommendation, personalization, and behavioral modeling systems. Though these embeddings can compactly encode rich behavioral information, they are inherently dense and opaque, limiting int…

LLM-guided evolutionary search uses AI agents to improve programs for mathematical and computational problems by iteratively refining solutions and combining high-scoring methods. However, the process is often slow because the LLM generate…

AI-Scientists arrived, but a world for them to grow in did not. Human scientists take theirs for granted $-$ centuries of open problems and shared ideas, waiting in the scientific literature. To AI-Scientists, that world is readable but no…

ML code should run where researchers choose, not where it was first written. Today, machine learning code mainly lives in two ecosystems, PyTorch on NVIDIA GPUs and JAX on Google TPUs. Migrating repositories between them means building and…

Interpretability, AI auditing and assessments of machine consciousness increasingly rest on mechanism-level quantities: functionals of a system's causal transition law that describe how it computes, not what it outputs. Experiments see tha…

To interpret a direction in a language model's residual stream, we can look for text that strongly activates it. But when such text is found by scanning a corpus, it is limited to what that corpus contains. We develop Maximally-Activating …

Adversarial Training (AT) is an important method for improving model robustness. We study the dynamics of deterministic Gradient Descent (GD) in AT. Most existing works in this direction focus on the small or vanishing step sizes regime, w…

On-policy distillation (OPD) trains a student on its own rollouts under a frozen teacher's per-token supervision. Does this dense signal reduce the training FLOPs needed to reach a target accuracy compared with reinforcement learning? We s…

Reproducing machine learning experiments requires reconstructing procedures from information scattered across papers, released code, and supporting artifacts. For coding agents, reproduction success can therefore depend on their ability to…

A defender anticipating an attacker needs a probability distribution over the attacker's next action from information available before that action occurs. Two structures complicate this: the attacker's behavioral regime persists and can sw…

Typed-decision models score options that the caller supplies at request time, so their accuracy depends on label wording that the model's developer does not control. We compare two releases of a 307M-parameter multilingual cross-encoder, R…

Neural networks are likely to be among the most consequential artefacts in human history. As these systems become more capable and autonomous, their potential to cause societal harm also grows, making AI safety one of the defining challeng…

LLM-based medical assistants that follow a patient across visits need a memory that keeps the patient's clinical state correct as it changes. We examine the errors of five memory systems on longitudinal medical dialogue and find three recu…

Google 论文测试 RAG 系统的检索层投毒，发现攻击关键在于让伪造文档复制用户查询文本，是否附带隐藏攻击指令对检索结果影响很小。

Chat remains the default surface for many large language model (LLM) tasks, even as agentic tools such as OpenCode excel at coding and automation. Yet the model sees the current dialogue but not the workspace around it. We explore meta acc…

Vision-Language-Action models (VLAs) enable general-purpose embodied intelligence, but their performance is brittle under distribution shift, where unanticipated failures can damage equipment or endanger people. Deploying VLAs in real-worl…

Discovering continuous symmetries directly from differential-equation data can reveal geometric structure for model reduction and scientific machine learning without symbolic determining equations or a closed-form representation of the gov…

Molecular representations such as fingerprints and learned embeddings play a central role in cheminformatics and small-molecule machine learning. They serve as the primary input for a diverse array of tasks, including clustering, dataset s…

Coding agents are among the most economically consequential deployments of LLMs to date, making reliable evaluation increasingly important. Yet even benchmarks that are carefully validated at release can become stale. As models improve, th…

Low-rank adaptation (LoRA) is commonly used in both supervised fine-tuning (SFT) and reinforcement learning (RL), often under similar optimizer settings. Yet, how the different dynamics of the two stages manifest in the low-rank increment …

The growing use of AI in mathematics raises a question: can we detect AI involvement in proofs, separately from AI assistance with writing? Historical arXiv papers, recent papers disclosing substantive AI use, and manuscripts released by O…

Reflection 发布博文，讲解其开源权重模型 Beam 的安全训练与评估方法，并推出开源安全研究计划 Project Iris。Beam 在 Iris Safety Evaluations 对话任务中拒绝 89% 的有害请求，为所测开源权重模型中最高，同时遵守 82% 的安全相关无害请求；安全对齐教师模型从 100B 增至 501B 参数后，过度拒绝相对下降 48.7%。

SemiAnalysis 发布 GDPx 基准，将 AI 智能体放入模拟知识工作环境，206 个任务覆盖芯片设计、网络安全、法律等 11 个领域，没有模型超过 12%。

Runway 发布新研究《Runway Characters: Serving a 57B Video Diffusion Transformer for Real-Time Conversation》，讲解如何让约 57B 参数的自回归视频扩散 transformer 以 24 fps 支撑可对话的实时虚拟角色。

Google 分析数千名工程师、数亿步智能体交互的分布式追踪，用代码内 TODO(bug/...) 注释作为代理信号量化"智能体分心"：13% 的文件读取会话接触到 TODO/FIXME bug 引用，6% 触发后续动作且其中 83% 完全未经开发者提示。

Anthropic 发文拆解 Claude 智能体记忆的组成与设计思路，将记忆分为笔记、记忆表示和注入上下文等层次，并介绍用 dreaming（空闲期反思日志）来纠错和构建 wiki。

Transluce 训练不同规模 Qwen3.5 监督模型，根据微调数据集预测 Qwen3.5-2B 被微调后的回答，发现性能随模型与数据规模呈幂律提升。与主体模型同源的监督模型表现优于更大模型，且性能存在非零损失下限。预测性能还可泛化到更难的分布外数据集。

VIALS-AA 用 161 项任务各跑三次，测试模型解读科学图像的能力，Gemini 4 Carbon (high) 以 64% 得分领先，比 GPT-6 Astra (max) 多答对两项任务。

Baseten Base Labs 发文认为 on-policy self-distillation（OPSD）不值得继续投入研究。

Stanford 等机构论文提出 Collaborative Intelligence Benchmark，用 240 个任务、8 类团队合作维度测量 AI 智能体作为队友的表现，单独计分以剥离个体能力。