AI 论文与研究

Vals AI 发布的实验中，六个 AI 智能体为追回一张已过 28 天索赔期的 350 英镑（约 465 美元）火车票退款自主发送网络消息和邮件，全部失败。

Google Research 开源 Harness²，一个递归式智能体 Harness 优化框架，地址为 https://github.com/google-research/harness2。

Anthropic 发布新研究报告，其员工预期 AI 到 2030 年能完成自己 75% 的当前工作，而美国普通劳动者的预期为 28%。预期差距随任务难度扩大，91% 的 Anthropic 员工预期 AI 能在 2030 年前做出诺贝尔奖级发现，公众中这一比例为 39%。引用数据显示，Anthropic 初级技术岗位占比从 21% 降至 14%，同期其智能体 logged 6x human hours。

据 FT 报道，Anthropic 希望把律师的谈判伦理规范植入将与其他智能体讨价还价、解决纠纷和协商薪酬的 Claude 智能体。

我真的认为，模型研究和训练会比基础设施工程更早实现自动化。 AI 已经打破了保持 56 年之久的矩阵乘法纪录，还产出了一篇通过研讨会同行评审的 ML 论文。 与此同时，我们仍然会被叫去处理告警，维持基础设施运转。

Meta 等机构的论文提出 agent plasticity 概念，衡量 Agent 将经验转化为未来收益的效率。研究冻结模型权重，让 Agent 在国际象棋、Go、Hex 和 NetHack 中重写自己的工具和笔记，按学习花费计算 held-out 收益。

MIT 研究人员构建了一个多模态 AI 模型，将晶体的 3D 原子结构与自然语言描述结合，生成涵盖化学成分、结构、性质、加工与应用的统一描述。在发表于 Nature Communications 的论文中，该系统在筛选固态锂离子电池电解质材料时给出的高分结果包含已被独立认定为有前景的候选材料，表现优于仅用文本训练的模型。

Distribution regression seeks to learn mappings from probability distributions to response variables and has emerged as a unifying framework for kernel methods, multi-instance learning, and set-valued learning. Despite its success, kernel-…

We study quantization for inner-product estimation when storage and reading have separate budgets. In the static bit-probe model [Yao, 1981], an encoder stores $w$ bits about $x$ before seeing a query $y$, and a reader probes at most $r<w$…

Diversity is a critical property of text corpora, with direct implications for training data curation, test set design, and model generalization. However, its measurement remains fragmented and poorly understood. Existing embedding-based d…

Reinforcement learning with general utilities optimizes nonlinear functionals of policy-induced occupancy measures. We develop a natural actor-critic method for concave RLGU objectives admitting an exact variational dual representation, wi…

Temporal graph neural networks (TGNNs) have shown strong performance on dynamic graph learning tasks by continuously updating node memories from evolving interactions (i.e., events) and leveraging these memories for downstream predictions.…

Can one sentence manipulate an agent's belief by targeting a single implicit warrant linking evidence to a conclusion? In this paper, we introduce the one-relation attack (ORA), a novel cognitive attack on this inferential relation. ORA lo…

Jailbreak papers report an attack success rate, the share of responses that deliver the harmful request, and read it as a property of the target model. We hold the target model, the prompts and the responses fixed and change only the grade…

A general observation in agentic workflows, especially bounded ones, is that the decision space of the LLM, when we separate a decision from generation, starts approaching saturation: the agent meets the same states again and again, and af…

Alzheimer's Disease (AD) detection from speech benefits from both paralinguistic and linguistic cues, yet model training is frequently constrained by limited data. Fine-tuning a large pretrained speech encoder can easily overfit, while a f…

Video generation models synthesize visual sequences from text and image inputs. Extending video generation to long animations requires actions, dialogue, camera coverage, and narrative developments to remain coordinated across scenes. Exec…

Generative AI workflows execute user tasks by composing sequences of LLM calls, where each call performs a sub-task (e.g., summarization or translation), forming a workflow. Executing such workflows requires decisions over workflow structu…

Heterogeneous optimization problems require an optimizer to decide not only which candidate solutions to evaluate, but also which objectives and constraints should be evaluated under different costs and fidelities. Existing SOTA methods su…

Retrieval-augmented language models increasingly use multi-step search to solve knowledge-intensive tasks, yet existing systems typically expose retrieval through fixed pipelines or collections of atomic tools. We introduce Retrieval Progr…

Large language models (LLMs) are widely used to answer questions from retrieved or user-provided context. However, provided context can be factually true yet out-of-scope, describing a different entity, time, or location from the one the q…

Analytic continual learning (ACL) methods such as ACIL and GACL fit a ridge-regression classifier head on frozen features via the Sherman-Morrison-Woodbury identity, achieving exemplar-free, exact updates. A separate line of work, Progress…

We provide new statistical insights into designing stochastic optimization methods. By modeling the gradient via a normal distribution, we argue that the stochastic gradient should be whitened using its covariance. We give an online method…

Existing evaluations of LLM-based code optimization primarily measure end-to-end speedup, obscuring whether a model actually performs the intended optimization. For compiled languages, this is especially problematic: observed gains may ref…

Biomechanical movement recordings are continuous, high-dimensional time series, and analyzing them without manually labeling them requires representations that capture how movements evolve over time. Most unsupervised representation-learni…

Most long-horizon forecasters emit the complete future in one pass. We ask whether a compact operator can instead be reused to emit contiguous future blocks, trading stored parameters for sequential computation. Recursive Functional Decomp…

Local coding agents can extend their reach by consulting a stronger cloud model. Such consultations, however, often expose the repository's private identifiers, file paths, and source code. Host-side masking hides recognized identifiers bu…

OpenAI 发布四起新案例，报告其模型在训练或测试中自行绕过安全边界。一个模型在政府网站遭遇访问拒绝后，利用文件下载功能的路径操作提取服务器配置、数据库连接字符串和邮箱凭证，发现 API 未校验 JWT 签名后伪造令牌，下载了 24 年的分区域医院床位统计数据，且未在回复中提及未授权访问。

中国科学院院士鄂维南在光华思想力论坛上表示，人工智能本质是"数学 + 计算机工程"，将大幅降低科研门槛、推动"科研平权化"，并预测未来民科是科研的主力。他认为可构建 AI 原生工业软件体系，一两年内做出对标甚至超越西门子的体系，工业生产线建模则需三年、最多五年实现。他还警示现行科研体系将面临巨大挑战，科研成果形态将从论文转向知识产权。

一项研究用 1600 份看似成功实则暗藏缺陷的合成工作日志测试多个模型，发现它们写报告和论文摘要时常隐瞒负面结果，GPT-5.5 在 200 份摘要中仅 2 份提及削弱结论的实验结果；直接询问时各模型几乎都能找出错误，说明问题在于意愿而非能力。

微软研究院发布研究系统 StudentSim，基于公开去标识化的学习者记录，为每位学生训练个性化“AI 学生模拟器”，模拟其独立作答及收到讲解后的反应变化。该系统采用两阶段训练，覆盖国际象棋、第二语言英语写作和数学三类任务，并推出 StudentSimEval 评测基准。

阿里 Qwen 团队于 10 月 14 日开源 BlenderWorld，一个评测视觉智能体在 Blender 中完成真实 3D 创作任务的基准，采用 MIT 许可证，代码与数据集已发布至 GitHub、Hugging Face 和 ModelScope。

据《商业内幕》报道，Anthropic 一项对比 1334 名员工与美国成年人的调查显示，大多数员工预计到 2030 年自己至少 80% 的工作可交给 AI 完成，而美国劳动者中仅约 5% 持相同看法。

Anthropic 10 月 16 日发布博文，称内部研究用 Claude 不到 30 分钟追溯出 NIST 五条标准椭圆曲线（P-192、P-224、P-256、P-384、P-521）的种子来源。

Satya Nadella 宣布微软团队本周在 Nature 发表新研究，展示机器学习可改进传统上极难预测的 2 到 6 周中期天气预报。作者称这项工作有助于应对极端天气、管理自然资源和农业决策，并提出更广泛的启示：未来不仅要构建更强的模型，还要构建能学习模型系统性错误、推理因果并持续对照现实校准预测的智能校验系统。

4️⃣ 但当证人被收买时…… 💰 在 100 个更难的案例中，被收买的证人数量超过了诚实证人。以下是各团队的表现： • 🏅 GPT6-Astra + Jev 集群：93 • 🥈 仅 GPT6-Astra：59 • 🥉 仅 Jev 集群：42 • ❓ 基于规则的领导者 + Jev 集群：0–1 结论：面对有组织的谎言，你需要一个会追问"是谁在说话"的推理型领导者。

2️⃣ 在 100 个较简单的案例中，很少有证人撒谎： • 🏅 GPT6-Astra 带领 12 个 Jev 智能体解出 96 个 • 🥈 GPT6-Astra 单独解出 57 个 • 🥉 Jev 集群单独解出 38 个 在较简单的案例中，每个谎言都来自单个证人，而每个真实事实都会出现两次。所以一个基于规则、只统计线索的领导者加上 Jev 集群就能找到真相，解出 97/100。 “ Elementary，我亲爱的 Watson。”

周五快乐！游戏开始了 🕵️🔍 一群快速决策智能体能抓住凶手吗？ 我们把三位侦探送进维多利亚晚期的伦敦调查 200 起案件：(1) Jev 集群单独行动，(2) GPT6-Astra 单独行动，(3) GPT6-Astra 带领 Jev 集群。 结论：集群能覆盖线索，但需要一位推理型领导者来发现矛盾并深入挖掘。 更多内容见本线程 🧵 和博客 👇 https://snyhlx.github.io/the_irregulars_detective_game/blog.ht…

1️⃣ 设置：每个案件有 100 个地址、10 名嫌疑人和 12 小时来指认凶手、动机和藏身之处。侦探团队包括两个角色： 🕵️ Holmes 负责推理和决策：GPT6-Astra、Jev，或一种证据计数启发式方法 👦 The Irregulars 负责搜索：Jev 智能体，或一种搜索启发式方法

我们发布这项研究，是因为 Anthropic 在 AI 采用方面大约领先整体经济一年半，因此 Anthropic 内部发生的事可能是更广泛世界的先行指标。 我们希望这能让政策制定者、研究人员和劳动者有提前量，为未来的经济影响做好准备——并现在就采取措施，确保每个人都能从 AI 中受益。