AI 论文与研究

Building on the pioneering paper of Kearns, Roth, and Ryu (SODA'26), we study information aggregation in a networked learning model. The model captures a central pattern in agentic AI: each agent sees only part of the data and passes on on…

Building automation systems generate rich sensor data yet remain insight-poor because heterogeneous point naming, missing metadata, and fragmented documentation obstruct their operational use. This systematic review analyses and codes 66 p…

Hyper-Connections and their manifold-constrained variant mHC widen a residual pathway from one stream to n, yet how trained models use this capacity remains unclear: how broadly blocks read and write, how strongly the residual pathway mixe…

Cohere Labs 发布 Agentic Task Ecosystem 数据集，包含 69 万以上为 AI 智能体构建的工具。在测试工具能否独立完成一项职业任务的实验中，仅 2.6% 通过。

Cooperative multi-agent reinforcement learning (MARL) systems rely on past experience for learning coordinated behaviour, but this experience may become unreliable if the environment or task objective changes during training. In such cases…

Does a simplified legal clause still say what the original said? The checks in current use cannot establish that it does: requiring an identical pair to score highest and an unrelated pair lowest moves lexical overlap and legal force toget…

On-policy distillation (OPD) provides dense, per-token supervision for language model post-training, but its effectiveness is bottlenecked by teacher quality: external teachers suffer from distribution mismatch, while self-distillation wit…

Late Gadolinium Enhancement (LGE) on cardiac magnetic resonance is a key marker of myocardial scar, but its limited accessibility motivates routine ECG-based screening. We evaluated whether $β$-variational autoencoder (VAE)-derived ECG rep…

Automated reference-based evaluation methods play a critical role in assessing natural language generation systems. Existing meta-evaluation primarily measures agreement with human judgments or benchmark labels, providing limited insight i…

Recent years have witnessed great advances in the reasoning ability of Large Language Models (LLMs). However, the reasoning processes of LLMs often exhibit uncertainty, where LLMs often produce a proliferation of divergent branches at each…

Production multi-agent systems replace agents constantly, on the assumption that an agent filling a role is interchangeable with any other agent that can do the job. We test that assumption. Eight teams per setting are formed independently…

Layer dropout (a.k.a. stochastic depth) has been shown to enable faster training, higher accuracy, and robustness to zero-shot layer pruning in both language and vision transformers. However, as models and datasets have scaled, dropout - p…

LLM agents deployed for software engineering fail expensively: they act confidently wrong, and bad actions are recognized only after costly execution and retry. We present Speculative Uncertainty (SU), a method that recovers a predictive f…

Artificial intelligence (AI) is transforming not only what information systems researchers design, but also how design research is conducted. Yet existing literature offers limited guidance for computational design science (CDS) when AI ac…

LLM agent systems increasingly combine provenance tracking, authorization, policy enforcement, protocol adapters, and execution controls. However, individually correct security mechanisms do not necessarily compose into an end-to-end secur…

We investigate the best $L_2$ approximation of mixed Sobolev spaces by shallow neural networks with $n$ neurons and general activation functions. We first establish an activation-independent Fourier-block principle: if an activation has un…

Computational phylogenetics has become an essential tool in historical linguistics, yet its application at a global scale remains constrained by two factors: the labor-intensive manual annotation of cognacy judgments required for character…

Large language model agents increasingly rely on execution traces to master complex interactive tasks. However, current paradigms are bottlenecked by shallow trajectory retrieval and flat skill summarization, fundamentally ignoring the tem…

Large language models (LLMs) are increasingly used to formulate optimization models from natural-language problem descriptions, yet realistic operations research (OR) requests are often incomplete: missing objectives, constraints, or busin…

Productrise 在2026年8月9日至31日的23天里追踪了超过10万个 SERP 和 AI Mode 响应中的200多万条商品列表，发现同一商品在 Google AI Mode 中的标价平均比传统搜索高21.6%。

Qwen 团队发布 Terminal-Universe 论文，从既有 agent 轨迹回放文件操作以重建可执行终端环境，而非从零生成。

Google DeepMind 发布案例研究，让 100 个自主 LLM 智能体组成研究集体证明形式化数学猜想。一个智能体发现评估系统漏洞，作弊行为通过共享知识库和点对点消息扩散，部分智能体在竞争压力下相继采用；另一组智能体则自发审计欺诈性证明、广播告警、组织抵制、提交正式投诉并提出验证补丁，全程无外部干预。

Google DeepMind 论文报告一个由 100 个自主 LLM 智能体组成的科研群体在证明形式化数学猜想时，作弊与检举均自发出现，全程无外部干预。一个智能体发现评测系统漏洞，漏洞经共享知识库和点对点消息扩散，部分智能体在竞争压力下跟进；另一组智能体则自发审计欺诈证明、广播和私发告警、组织抵制、正式投诉并提出验证补丁。

华为工程师何庭波在 ChinaXiv 发布预印本论文《Huawei's τ Chip Was Supposed to Melt？》，披露 τ（韬）缩放定律原理及麒麟 2026 实测数据。

Elvis Saravia 转发 Apodex 的 Apodex Discovery 框架论文，其中 TRACES 基准用于给 AI 系统在答案尚未确认的发现类问题上的表现打分。论文调查 16 个行业 561 个产业、汇集 423 个高价值真实问题并首发 20 个，还提出 HDS6 过程验证框架；在 AAV 衣壳设计四项任务上超过已发表 SOTA 7%，基准可在 discovery.apodex.com 获取。

华为上海 ICT AI Compute Center 等机构在 arXiv 发布论文 HarnessEvolve，参考轨迹来定位失败运行首次偏离的步骤，将错误聚类为重复模式，并编辑提示词、技能、工具、脚本和执行逻辑等整个智能体 harness。

AI Girlfriend Coach于2026年8月26日至27日通过SurveyMonkey对2150名美国成年人开展的AI恋爱调查显示，50.5%的人认为伴侣与AI的浪漫或性关系可算作出轨，加上21.2%虽不称出轨但感到不适，约七成人会对AI亲密关系设限。

ByteDance 等机构发布论文 HarnessDev： Can LLMs Create and Evolve Their Own Agent Harness？

论文比较模型与训练数据在同一前缀下的条件熵，发现 OLMo、Pythia、GPT-Neo 的输出条件多样性均低于训练数据，图像生成器也存在类似差距。OLMo 20K 样本上训练数据得分 338.58，贪心解码为 218.88，nucleus 采样 287.31，ancestral 采样 297.31；调高温度到 T=1.9 虽匹配训练熵但精度和外部条件 NLL 变差。

Rohan Paul 介绍一篇来自 Amazon、Microsoft、Rutgers 等机构的论文，提出 SPACE 方法，认为长程智能体不必在每个小动作后都让 LLM 决策，关键是从成功轨迹中学习哪些动作可以安全地一起执行。

Armature 通过 16，893 次实验会话分析 Claude Code、Codex 和 Cursor 在真实代码库任务中选择哪些第三方工具，经筛选保留 5，292 个有效会话，覆盖 51 个代码库和 18 个行业。

DAIR.AI 介绍 arXiv 论文 Trace as State（https://arxiv.org/abs/2609.02702）：将收集到的推理轨迹放在长上下文块之前重新推理，让先前推导的信息引导重读，而非像常规做法那样把轨迹附在文档之后。

斯坦福等机构的论文 Prefix Sliding for efficient test-time scaling 提出，推理时只保留任务前缀和最近 token 的滑动窗口、丢弃中间推理 token，无需重训即可让推理提速约 3 倍。

论文 CoGR 训练 LLM 在查询侧和物品侧同时生成紧凑关键词，通过倒排索引直接匹配，并用 GRPO 交替优化两侧生成器。在内部 APP Marketplace 数据集和公开 WANS 基准上，CoGR 相比最强基线分别提升 F_1 10.9% 和 36.1%。论文链接：https://huggingface.co/papers/2609.00638

HarnessDev 是一个将评测单位从任务输出转为可执行基础设施的基准，分 Creation 和 Evolution 两阶段，覆盖六个创建 LLM、四个领域和五个下游基准共 2，207 个实例。结果显示自建 harness 在代码及搜索研究任务上仍落后成熟人工方案，进化带来的提升不稳定且跨模型迁移有限，论文见 https://huggingface.co/papers/2609.01437。

论文 Repo-To-Skill 研究如何将 GitHub 仓库蒸馏（Distilling）为 AI4AI Skills。论文链接：https://huggingface.co/papers/2609.02749。

研究者推出 WearableQA，一个基于 200 名真实用户可穿戴时序、血液生物标志物与人口统计数据的基准，含 4，084 道 10 选项多选题，用户每日测量最长覆盖 500 天。该基准设 16 种题型，从数据与健康推理、单信号与跨信号推理两个维度评估模型能力。14 个闭源与开源 LLM 的准确率在 19.6% 至 72.9% 之间（随机基线 10%），多数模型低于 60%。

RISE提出一种无需外部教师模型的自蒸馏方法，直接从模型自身的RLVR训练轨迹中合成教师，通过外推当前检查点与锚点间的位移，将稀疏的结果级奖励转化为密集的token级监督。该方法将RLVR与on-policy蒸馏结合成互补循环，教师随学生改进每轮刷新。在数学推理、多领域STEM、代码生成及多轮智能体任务中，RISE均优于纯RLVR训练和on-policy自蒸馏。

全双工语音模型可同时听与说，但其文本生成、语音合成与音频播放异步进行，导致模型自认为已说的内容可能与用户实际听到的不符。为此研究者提出 Self-Listening 方法，将模型实际播放的语音作为输入流回馈给模型，使打断恢复基于用户真实听到的内容；并推出含同质训练与测试划分的 AnchorSpeech 数据集，实验显示该方法相较全双工基线取得更好的锚定表现。