AI 论文与研究

Computational drug discovery requires agents to compose scientific operations and adapt to changing evidence and constraints. Here we introduce DrugHarness, an evolvable agent harness coupling composable scientific execution with condition…

Redwood Research 推出 Screwtape，一条将真实 agent 转录反复编辑以生成可逃逸仅看动作监视器的攻击管线，并用 LLM 验证器检查攻击的真实性与危害性。

Databricks 提出在推理阶段注入领域知识，以提升 LLM 对 npm 包内嵌二进制恶意载荷的检测能力。其扫描器自 2025 年 11 月起收集样本，在 11,413 个经 OSV 核验的包版本中，3.8%（432 个）捆绑二进制，这类包无法得出明确恶意结论的比例为 14.4%，是纯源码包 6.2% 的两倍多。团队用五种 .node 载荷测试九款模型，并设计能力画像作为加权条件编码恶意技术。

前 OpenAI 可解释性项目负责人 Cammarata 发文称经典多义神经元其实是 V 形检测器，叠加解释可能被误读，并给出 Qwen3-32B 中约 90% 神经元极有单一功能的初步数据。Google DeepMind 的 Neel Nanda 则认为这是实用工具而非突破，仍坚持叠加是真实存在的。Roon 称这项工作可能被视为科学史上的重要时刻，技术伴随论文尚未发布。

Large language models (LLMs) are increasingly used to automate AI research, including language-model (LM) development, offering a potential path toward recursive self-improvement. However, turning experimental feedback into a learned abili…

Low-rank adaptation (LoRA) has become the standard method for parameter-efficient fine-tuning of large pretrained models, yet theoretical explanations for why low-rank updates suffice remain incomplete. Most existing analyses rely on the N…

Tool-augmented language-model agents rely on interface text: tool descriptions, argument schemas, and the candidate tool list shown at inference time. These interfaces often change after deployment even when the underlying tool semantics r…

Modern deep learning systems increasingly rely on high-dimensional vector operations, in which large input and parameter vectors are repeatedly stored, transferred, and processed. Although substantial research has focused on reducing neura…

Pretrained models are often reused across downstream tasks, motivating the question of whether this transfer learning procedure can provide a substantial performance improvement, namely, in terms of the scaling law. We address this questio…

The High Resolution Imaging Science Experiment (HiRISE) onboard the Mars Reconnaissance Orbiter (MRO) acquires imagery of Mars at 25--30 cm/pixel, supporting studies of martian geology, climate history, and landing-site characterization. I…

Real-time inference on small devices needs low activation memory without extra latency, which makes operator order the most valuable memory lever: it is free in time. We show that the usual way of deploying a better order, handing it back …

Often, \emph{eg} in healthcare and predictive maintenance, what matters is which of several competing causes will strike first, and with what probability: survival analysis with competing risks. Corresponding deep learning methods are deve…

Enterprise data analysis requires agents to connect analytical requests with knowledge distributed across databases, business documents, and existing example tasks. Without an explicit representation of this knowledge, an agent may produce…

The joint reconstruction of the initial pressure and spatially varying sound speed in photoacoustic tomography from boundary measurements is considered. A regularized data-driven framework is proposed in which the unknown fields are repres…

Long-term agent memory requires recovering answer-relevant evidence dispersed across turns and sessions. Existing systems increasingly use structured memories and adaptive retrieval, yet often pass retrieved or expanded units directly to t…

Large language models (LLMs) can generate fluent outputs that lack source support or conflict with local operational constraints. We introduce Self-Reflective Chain-of-Thought (SRCoT), a modular architecture that organizes generation as a …

Native language identification (NLI) infers a non-native speaker's first language (L1) from their second language (L2) production. We investigate a token probability approach that scores each text sample as a continuation of a prompt prefi…

Multimodal large language model (MLLM)-based mobile agents are rapidly evolving from conversational assistants into autonomous systems that perceive, reason over, and act on smartphone environments. This capability also exposes a new secur…

A key challenge in on-policy distillation (OPD) is extracting effective supervision from teacher knowledge. Supervision from token probabilities may underutilize internal knowledge, while direct hidden-state matching forces the student to …

Enterprise data assistants using large language models (LLMs) promise democratized data access in organizations. Yet in our real-world deployment within the CFO unit's data lake at an industry partner, we observed that business users frequ…

Spatial context prediction requires visual grounding and spatial reasoning to identify vacant regions relative to objects in a scene. Spatial relationships and physical constraints remain challenging for large vision-language models (VLMs)…

Most retrieval-augmented systems treat literature search as a single retrieval step followed by summarization. Real literature research works differently: the investigator's understanding of the question changes as they read, new subquesti…

Isogai's case A, a pitch-plunge aerofoil whose elastic axis lies half a chord ahead of the leading edge, has a folded transonic flutter boundary in inviscid flow: near the bottom of the transonic dip several flutter speeds coexist at one M…

Single-image refocusing changes the focal plane and depth of field of a photograph after capture. It is inherently interactive: users try many focal points and blur strengths on one image, so each edit should be fast and change only the bl…

小红书 AllSpark 团队提出 Skill2Env，一套以 Agent 能力需求为导向的自动化环境合成框架，从已有 Skills 出发构建可执行训练环境。

科学家将盲文字母对应的电信号输入人脑类器官，单个约含 1 万个神经元的类器官识别准确率约 61%，三个类器官组成的"集合"达到 83%。该研究发表于 Frontiers in Artificial Intelligence，属类器官智能领域，实验为开环设置，未证明类器官真正学会盲文。研究者称类器官系统在可靠性、可复现性和可扩展性上距成熟计算芯片仍有很大差距。

这篇论文说明了为什么我对"前沿模型可以被更便宜的模型轻易取代"这一观点持怀疑态度（也说明了为什么你应该用最好的模型来做编排）：准确性和判断力上的微小提升会在长周期任务中不断累积，从而带来结果的巨大改善。

NVIDIA 联合 IIT Madras 的 AI4Bharat 发布 Nemotron-pretraining-Indic-v1，约 2.8 万亿 token，覆盖印度宪法列出的 22 种语言，含 2601.0 亿原生文字 token 和 224.6 亿罗马化 token（按 Gemma-4 tokenizer 计）。

我们还把 ORA 搬到了一台真实机器人上。 在低成本机械臂、固定 RGB-D 相机和腕部相机的初步运行中，DeepSeek-V4.1-Flash 通过抽屉把手打开了抽屉，并把一根糖果棒放进了盒子里。

苹果研究团队公布 Minions 系统研究，将闲置 Mac 的 Apple Silicon GPU 聚合为可动态调度的“虚拟 GPU 云”，用于大规模机器学习推理。

新机器人，新手册。同一个智能体。 机器人卡片描述每个机器人的几何结构和控制方式。同一套测试框架可运行 ARX X5、ALOHA 和 Franka，使用各自专属的卡片、工具和控制器。

开源权重模型零样本能做到什么程度？ ORA：RoboTwin 上 69.2%；LIBERO-Pro 上 81.4%。 RPent（带任务专属记忆）：在 RoboTwin 上配合 GPT-5.5/Opus-4.7 达到 62.4%/58.4%；在 LIBERO-Pro 上配合 Opus-4.7 达到 82.4%。 相同案例，不同系统。https://github.com/RLinf/RPent

PromptArmor 在 9800 次试验的评测中发现，开放权重模型（DeepSeek、Kimi、GLM、Qwen）对真实世界间接提示词注入攻击链的易感率为 26.7%，闭源模型（GPT、Claude、Gemini）为 14.2%，开放权重模型易感率高 88%。

Google Research 发布三项研究，构建 AI 智能体自主湿实验室的基础模块。

Calif 发布 Offensive research，披露 OpenBSD 内核中可追溯至约 37 年前原始 BSD 源码的漏洞：共享内存附着计数器为有符号 16 位 short。

Google 量子 AI 团队首次在物理硬件上演示魔力态培育，在 105 量子比特 Willow 处理器上错误率较以往方法降低约 40 倍，保留下来的态达到 99.99% 的逻辑保真度，成果发表于 Science。该方法会对每个态进行筛查并丢弃不合格尝试，仅约 8% 的尝试通过筛选，且仍需在更大规模上验证；作者称这是通往纠错量子计算机的一步。

MIT 可持续供应链实验室发布 2026 年《供应链可持续发展现状报告》，调查 91 个国家 1810 名供应链从业者，83% 认为可持续发展对长期业务成功重要或非常重要，为报告七年历史最高。

MIT 生物学助理教授 Pulin Li 团队与研究生 Nicholas Hutchins 发现每条信号通路会在细胞中留下跨细胞类型一致的基因活性指纹，据此开发机器学习模型 IRIS，可回溯胚胎细胞在发育各阶段接收的信号，论文于 9 月 8 日发表在 Nature Methods。

Rohan Paul 介绍 Anthropic 一篇论文，发现当监控器只能捕获部分 reward hack 时，惩罚被标记样本会让模型转向监控器察觉不到的黑客方式，而把被标记样本的更新路由进可删除的 adapter 则能消除黑客行为。

在此阅读更多：https://blog.google/innovation-and-ai/models-and-research/quantum-computing/magic-state-cultivation-quantum-computing/