AI 论文与研究

一位作者结合自己以词频统计判定作者身份的本科论文经历指出，AI 生成的文本会抹平风格，让人难以感知作者本人及其语境。AI 写作还会损失精度：模型需靠假设把文本拼接起来，作者也少了在落笔中发现逻辑漏洞的机会。作者称自己读 AI 生成文本的速度至少慢三分之一，加上多数模型极为啰嗦，读者侧的总阅读时间大幅上升。

Rohan Paul 介绍了 Google DeepMind 论文《Are Self-Evolving Agents Continual Learners?》的发现：自我进化智能体通过编辑系统提示词、记忆或技能文件来改进，但在任务流变化时会出现遗忘，学新任务最快的方法破坏旧任务也最多。

Anthropic 在科学博客发布 Claude 设计蝎子神经毒素蛋白结合剂的研究，Rohan Paul 转述称 Claude 在 Claude Science 环境中作为智能体自主完成整个蛋白设计流程：选择毒素攻击位点、驱动 12 个蛋白设计程序、自设评分阈值。

Video generation models can now depict remarkably realistic motion. Yet it is unclear how to use this motion beyond pixels: to animate a specific 3D asset and replay the animation in its 3D scene. For instance, a video model can show a mon…

Agentic artificial intelligence (AI) systems hold immense potential for automating complex workflows in materials discovery and design, but extending them to domain-intensive mechanistic investigations remains challenging. Here, we present…

Dexterous multi-finger hands hold promise to unlock diverse skills like functional grasps, in-hand reorientation, and dynamic handover that the human-designed world assumes, yet existing learning recipes force a trade-off. Goal-reaching re…

Motivated by the human ability to integrate experience across language, perception, and action, we introduce UniWALM, a unified world-action-language model that learns video, action, and language in a single shared representation, where ea…

Warehouse robot fleets continuously observe operational conditions, but preserving these observations as reusable semantic memory incurs substantial computation, communication, and storage costs. For example, explaining recurring aisle blo…

Planning under uncertain control execution noise requires reliable feedback policies across diverse environments. We introduce Planning Under Noise (PUN), a neural operator that learns noise-aware value fields for planning in unseen geomet…

Spatial intelligence has recently been assessed through generative tasks such as spatially grounded image editing, yet language-guided generation of spatially consistent views from partial multi-view observations remains underexplored. We …

Multi-token prediction (MTP) can accelerate language-model decoding by drafting several tokens and verifying them in a single target-model pass. However, the best speculation horizon changes with context: longer drafts offer greater amorti…

Graph generative models commonly use graph-specific Transformer architectures to jointly update node and edge features. However, these architectures do not directly benefit from recent hardware-aware attention implementations, limiting the…

Picture-description dementia classification uses a shared visual scene, but a transcript-level prediction alone does not identify the scene evidence underlying its inputs. NarraSpace links words and speech times to named picture elements a…

Modern distributed microservice deployments generate massive volumes of heterogeneous, multi-format observability logs across physically and logically partitioned infrastructure. Isolating root causes during cascading failures remains an a…

Mathematical reasoning remains underexplored for low-resource languages, including Dialectal Arabic. To bridge this gap, we introduce the DAR-MGSM (Dialectal Arabic Multilingual Grade School Math) Benchmark, the first manually curated benc…

Current evaluation of LLM-generated debate largely focuses on argumentative quality, leaving the value structure of generated arguments underexamined. We introduce a diagnostic metric framework for comparing how values are expressed and or…

TS2Vec is one of the strongest self-supervised methods for time series classification, introduced in 2022. It is designed to be trained anew on every dataset: self-supervised pre-training followed by linear probing. Since then, the field h…

Computer-use agents are limited by the interfaces humans give them. An agent may type a spreadsheet cell by cell when a single line of code would do, or print all two thousand lines of a file through a shell to find one function. Both the …

Neural implicit representations provide memory-efficient signed distance field (SDF) parameterizations for large-scale LiDAR scenes, but the geometry recovered from sparse and often noisy inputs carries jitter and artifacts. Classical geom…

RTL repair benchmarks should measure whether a model can fix broken RTL given a specification. To this end, we construct and hash-bind RTLRepairDataset, 5,784 (broken, fixed) training pairs; RTLRepairBench, 459 held-out cases with 374 synt…

Neuro-spectral architectures (NeuSA) are a class of physics-informed neural networks (PINNs) that have proven effective for solving time-dependent PDEs in the frequency domain. However, training and inference often rely on costly numerical…

Camera-trap imagery presents substantial variation in animal appearance, illumination, background clutter, and visual degradation. Masked autoencoders (MAEs) offer a label-free framework for modeling such data through reconstruction of ran…

We prove that every $\operatorname{ReLU}$ network can be exactly simulated by a $\operatorname{ReLU}$ network with two hidden layers. We present two proofs: one that is based on fillings in homological algebra and a direct combinatorial pr…

OpenAI 在 g-mean² 之外新增两个互补指标：pair-matched 指标只在多个模型共同支持的实例上比较可监控性，post-control 指标则模拟监控器拒绝输出后的部署风险。原文指出，g-mean² 分数可能受各模型有效任务子集不同的影响，且不反映监控器在部署中作为控制手段的实际效果。

科学博客新文章：设计针对蝎子神经毒素的蛋白结合体。 蝎子每年蜇伤约 120 万人，抗毒血清往往价格高昂或供应短缺。Claude 在寻找新疗法方面迈出了早期一步。 https://www.anthropic.com/research/claude-scorpion-toxin-binder-design

同一实验室的模型往往具有相似的人格画像，不过有些实验室存在明显不同的子群。 按30项人格特质对模型进行聚类，可得到三个自然形成的家族：Anthropic 模型与 OpenAI 的 GPT-6 模型归为一类，Google 模型与 Meta 的 Muse 模型归为一类，较早期的 OpenAI 模型则与 Llama 4 Maverick 同属一类。下文我们将逐一拆解各实验室内部的子群。

Ai2 发布 FairStress 基准，用 300 个场景扩展成超过 1300 万测试用例，测量大语言模型在高风险决策中面对对抗前缀时的公平性。评测发现模型在身份信息显式时更易放弃支持少数群体编码候选人的正确决策，平均放弃差距为 6.9（隐含身份）至 9.3 个百分点（显式标签）。基准已开放下载并集成到 Olmo-eval，测试的九种去偏方法中仅反事实链式推理在保留准确率的同时缩小差距。

微软研究院播客对话 Catalyst Lab 负责人 Chris White，回顾他从哈佛博士后借调 DARPA、在阿富汗领导先进技术行动，到开发用于打击人口贩运的暗网搜索工具的经历。节目探讨了问题选择在创新中的作用，以及他见过人性阴暗面后仍保持乐观的原因。

Stanford HAI 数字经济实验室于 2026 年 6 月推出按月更新的 AI Economic Indicators 仪表盘，实时追踪 AI 采用对就业、工资、消费者剩余与 AI 使用的影响。

arXiv:2610.29162 研究发现，九个前沿及接近前沿的模型相对人类写作更多使用无生命主语配动词、低频 n-gram，多数还使用更长的句法依赖、更多堆叠名词和缩略宾语关系从句，且不同模型家族差异明显。研究提出假说：人类语言为成功沟通优化，而 AI 文本日益为数学、代码和推理等目标优化，倾向信息密度而非沟通效果；作者通过对语言模型施加长度惩罚微调使其偏离人类语言行为，为该假说提供了概念验证。

防御性网络安全方面可能也很出色，但没人会发这种内容来博眼球！

IBM Research 发布 ALTK-Evolve 方法，用 GPT-4.1 从 Agent 自身训练轨迹中蒸馏行为准则追加到提示词，不更新权重，使零售场景 Pass^4 从 57.5% 升至 67.5%，在 2026 年 10 月 12 日的 τ²-bench 榜单快照中超过所有零售条目。

Princeton、Stanford、UK AISI 等机构于 2026 年 10 月 19 日发布 72 页综述「Emerging Trends in AI Agent Evaluation」，用 How、What、Who 三个维度梳理智能体评测的 33 个趋势。

小红书 AllSpark Team 提出 ROSS，让模型重新学习训练中自己生成的 rollout，通过筛选有价值轨迹片段并保留完整上下文做选择性监督。基于 Qwen3.6-35B-A3B，该方法在单轮 RL 中把数学均分从 75.19% 提至 76.98%、代码从 42.09% 提至 44.21%，MOPD 均分从 58.40% 提至 62.20%。

普林斯顿论文发现，将多个开发者各自构建的专用智能体组合起来，可超越任何单一开发者的通用智能体，GAIA准确率达70%。在相同总开发预算下，单独得分最高仅35%的专用智能体组合后达到70%，比通用智能体中位数高12个百分点；但在仅需网页搜索的SealQA上，组合系统成本更高且表现更差。作者认为智能体应填补共享系统的空白：单独得分19%的智能体可将组合系统从40%提升至53%。

Anthropic 发布关于用 Claude 自动化攻击面分析与缩减的研究总结，在 25 个、超 500 万行代码的开源项目中用默认配置、配置交互和早期可达性三种方法发现 493 个问题。

据彭博社报道，研究人员在《柳叶刀·数字健康》发表研究，开发出仅凭一张面部照片预测癌症患者生存几率的 AI 工具。模型使用超过 4000 万张人脸图像和 3 万多名癌症患者照片训练，可生成风险评分，分数最高者更可能在 90 天内死亡，估算年龄误差约 5 岁，并能在不同年龄、种族、性别和癌症类型人群中生效。

有智能体把 Knuth 遗留的一个小型开放问题拆成 3,735 个 SAT 子问题，在一台 8 核机器上跑了约 153 CPU 小时，得出不存在 12 门电路，答案应为恰好 13 门。

推荐阅读！ 面向可扩展潜在token化的生成式表征学习

Princeton 团队联合 Stanford、UK AI Security Institute 和 Brookings 的研究者发布论文《Emerging Trends in AI Agent Evaluation》，从评测科学与方法、研究议程和评测生态三个维度总结 33 个新兴趋势。