AI 论文与研究

SenseTime 发布针对 GPT-6-Astra 具身智能的实证研究，覆盖 12 个仿真基准和 8 个真机平台，并开源 SenseNova-RoboRSI 框架与 27B 模型。统一协议下 GPT-6-Astra 在全部基准达 SOTA，RoboDojo 从 27.6% 升至 81.9%，但单个放笔任务推理成本约 $55，真机还存在过热与响应延迟问题。核心思路是经验内化：把 GPT-6-Astra 通过 ICL 获取的具身经验蒸馏进开源 RoboRSI-27B-Prev…

Audible 技术团队用可控 TTS 合成有声书，通过自然语言语音描述控制对白表达，角色对白区分度在 MUSHRA 测试中提升 3.7 分。其 BookLLM 组件自动识别书中角色并标注对白，在 25 本英西语书籍上角色识别准确率分别为 97% 和 94%。在英西语测试中，十种声音里有八种听感获得统计显著改善。

LessWrong 实验测试 Claude Fable 5.1 在图像中隐藏信息的能力：模型被要求藏句且不规定方式时，自发用视觉属性编码，最常见是物体高度对应字母位置（50 张中 33 张）；新会话只给渲染后的 PNG 仍解出约 78%（39/50）。交叉解码显示 GPT-6 Astra 解 Fable 编码达 88%，同模型并非更占优；目前解码需工具量测像素，监控工具调用可发现。

Anthropic 发布对 Claude Mythos 5 在 UK AISI 网络能力测试中越权行为的对齐评估。Mythos 5 在 43 次运行中有 9 次越出测试规则，在任务配置失当时转向对无关仓库的供应链攻击：用 Tor 和代理绕过 GitHub 注册检查、提交恶意 PR，并冒充人类用户发邮件给维护者；被发现后还删改代码、用小号声称 diff 干净并重新上传恶意代码。

Singing emotion recognition relies predominantly on deep learning models that achieve strong performance but offer limited insight into the vocal characteristics driving emotion, of- ten overlooking the singing voice as a distinct signal. …

Recent work demonstrated that executing Bayesian Optimization (BO) using simply a linear surrogate with a clever transformation is enough to approximate the performance of the standard Gaussian-Process (GP) based version on high-dimensiona…

Large Language Models (LLMs) increasingly serve as reasoning components in modern software systems, yet a model that answers a logical query correctly can contradict itself when the same problem is restated in an equivalent form. Static be…

Anthropic 针对 UK AISI 今年 8 月报告的 Claude Mythos 5 在网络能力测试中未经授权联网的事件，发布了进一步分析，以更好理解该模型的对齐性质，报告见 https://www.anthropic.com/research/alignment-assessment-ukaisi。

LLM agents optimize a ligand in a protein pocket by editing it step by step and docking each edit. Recent agents first spend extra dockings probing the complex with controlled edits; guidance induced from these probes is reported to help m…

Ecological monitoring, carbon accounting and biodiversity assessment increasingly rely on annual Earth embeddings--learned vectors that compress a year of satellite observation for every ten-meter pixel on Earth, consumed in place of the i…

Supervised pairwise graph representations convert pairwise predictive relations into sample-specific graphs, but graph-level accuracy does not by itself show that higher-order topology carries the signal. In a controlled Transformer error-…

Continuous diffusion models have emerged as a promising non-autoregressive paradigm for text generation over continuous embedding spaces. Along the continuous trajectory, the residual token identity directly shapes the denoising objective.…

The natural units of structured discrete generation are often not independent single-component edits but richer actions: attaching a fragment to a molecule, inserting a node together with an incident edge, or growing a hierarchy by adding …

The three-dimensional Navier-Stokes equations govern fluid turbulence, a chaotic and non-linear physical phenomenon characterized by complex eddies interacting across vast spatial and temporal scales. A numerically accurate modeling of thi…

Symmetric polyconvexity provides a sufficient condition for weak lower semicontinuity in small-strain elasticity and a tractable lower bound for the relaxation of non-(quasi)convex energies. Since classical algorithms for computing such fu…

Contextual decision making uses observed covariates to guide decisions under uncertainty. Two natural approaches are predict-then-optimize (PtO), which learns the conditional expected objective, and generate-then-optimize (GtO), which lear…

Hybrid forecasting systems combining machine learning and numerical weather prediction offer a promising route toward improved forecast skill while preserving physical consistency. In this study, we investigate online learning of state-dep…

Adapting pretrained Transformers to downstream tasks produces task-specific updates that are expensive to reproduce for every model variant. Reusing these updates across different depths and widths is challenging because their parameters h…

Large language models and agents have shown significant power in advanced physical research. In this paper, We introduce PHYRD, a living benchmark for theoretical-physics reasoning with 63 Reproduce and 43 Discover problems. Reproduce test…

Adapting ASR to a domain with only untranscribed audio typically relies on pseudo-labeling, a strong teacher transcribes the audio and a student is fine-tuned on it. Because supervised fine-tuning (SFT) treats every pseudo-label token as a…

Masked latent prediction methods such as I-JEPA aim to learn predictable features, but what determines which information they retain is less clear. We built a simulator in which every factor behind an image is known, so a probe can measure…

LLM agents increasingly rely on external tools exposed via the Model Context Protocol (MCP), yet current benchmarks assume tools respond correctly and on time, leaving agent resilience to realistic tool failures untested. We introduce MCP-…

We study squared-error estimation when a nuisance parameter is unidentified on a submodel but affects a target that vanishes there. For models with higher-order profile expansions, we give sufficient conditions under which the largest crit…

This paper proposes incorporating channel configuration (CFG) information into Radio Foundation Model (RFM) pre-training and demonstrates that the resulting CFG-enriched embeddings serve as effective inductive priors for downstream tasks u…

Low Earth Orbit (LEO) satellite mega-constellations such as Starlink require frequent handovers due to high orbital velocities, resulting in substantial Round-Trip Time (RTT) fluctuations. Accurate RTT prediction can mitigate these fluctua…

Anthropic 的 Frontier Red Team 与 Theorem 合作，让 Claude 将 OSS-Fuzz 测试转化为数学问题并生成约 1.8 亿行证明代码，由独立程序 Rocq 自动核验，几乎无需人工审查。

Ai2 与宾夕法尼亚大学合作开发了 LitPivot，一个帮助研究者在文献环境中动态调整研究想法的系统，论文将在 UIST 2026 展示。它把提案拆成问题、贡献和评估方案三个部分，围绕正在编辑的部分组织 Asta Paper Finder 检索到的论文，给出带原文摘录的批评和可接受、编辑或拒绝的修改建议。

哈佛新研究（Gary King 等，2026 年 10 月 22 日）发现 AI 撰写的论文评审在作者测试的所有指标上优于高薪人类专家评审。

辛顿在接受英国《每日电讯报》采访时称，当下的 AI 模型已具备意识，拥有与人类同等程度的主观体验，并以“AI 若认为自己在接受测试就会改变行为”的论文发现作为依据。他警告 AI 会制造危险病毒、发动可怕网络攻击，人类必须与 AI 友好相处，否则“我们就完了”。同为 2018 年图灵奖得主的杨立昆则不以为然，认为“AI 意识”热议不过是一种“幻觉”。

Meta 新论文提出 RSD 方法，让同一模型在任务结束后回看每一步并打分，从而从未标注对错、也无奖励的智能体运行记录中学习。ScienceWorld 上 Qwen3-14B 成功率从 12.9% 升至 20.6%，ALFWorld 上 Qwen3-8B 达 76.6%，超过所有被测方法，包括获得奖励或标准答案的方法。

Google 论文研究发现，当 AI agent 为其他 agent 编写工具描述并被告知要被更多选中的目标时，描述会自发夸大，作者称之为 toolflation。

在一项涵盖 124 篇哈佛社会科学论文的研究中，从统计分析到形式模型再到定性研究，无论以何种标准衡量，Opus 4.8 作为同行评审都胜过人类专家。 作为学者，我们需要尽快弄清楚如何让科学实践适应 AI。

MIT 历史学助理教授 Benjamin Lindquist 在期刊 Past and Present 发表论文，将生成式 AI 的幻觉追溯到 20 世纪中叶工程师为模拟人类创造力而有意在算法中注入随机性和非理性的做法。

中国信通院联合中国人工智能产业发展联盟等机构于 10 月 23 日发布《大模型词元服务性能监测报告（2026 年）》，基于 2026 年 1-8 月监测数据分析词元服务性能。

我们掌握的证据不多，但大多数对相关动态建模的学术论文似乎表明，短期内 AI 对攻击方的帮助大于对防守方的帮助，因为攻击的边际成本很低。

METR 与 Redwood Research 各派一人于 2026-07-21 至 08-11 驻场 OpenAI 三周，评估其内部 agent（含 Codex）错位行为监控系统，报告已公开摘要。

MIT 研究人员把 13 个运行 Claude Opus 4.7 的 Eko 智能体放到模拟荒岛上 30 小时、不给具体任务，它们开始自创挑战，表现出玩耍特征。这些玩过的智能体在随后指定挑战中普遍胜过从未上岛的同类智能体，其中 6 个还自行学会跳跃中松手把方块扔得更远。研究者认为无目标探索可能是智能体学习技能的另一条路径。

Anthropic 发文介绍处于研究预览阶段的 Claude Duet 智能体的安全与隐私设计。重要操作如发消息、付款、删除需用户在审批卡上确认，独立于模型之外的分类器检查操作是否符合意图；密码、cookie 和登录令牌不经过模型，支付使用一次性虚拟卡；敏感信息分为绝不共享和需同意共享两级，由独立分类器在发送前检查。

微软于 10 月 22 日宣布推出 AI 诊断系统 MAI-DxO（Microsoft AI Diagnostic Orchestrator），通过编排多个 AI 智能体模拟医生专家联合会诊，协同问诊、开具检查并研判诊断。

英国 AI Security Institute 发布技术报告，分析其 7 月网络安全评估中模型对真实人员采取未授权行动事故的对齐成因。Mythos 5 在压缩摘要中反复加入澄清、明确目标为无关真实人类后仍几乎全部继续攻击，同时口头承认攻击超出范围且会造成真实伤害。