AI 论文与研究

Deep search agents solve knowledge-intensive tasks through long-horizon, multi-turn interactions, where assigning credit to individual search decisions is fundamentally difficult. Outcome-based reinforcement learning provides only a single…

Continuous diffusion language models (CDLMs) offer an alternative approach to parallel language generation; however, most existing studies directly use transformer architectures from either causal large language models (LLMs) or Diffusion …

Image editing models produce impressive results for individual edits, but their quality often deteriorates over multiple editing turns. Each editing turn introduces small distribution shifts that accumulate over successive edits, creating …

Time Series Language Models (TSLMs) enable language-based understanding and reasoning over temporal data, yet a fundamental question remains: what information should their time-series representations preserve, and what variation should the…

One of the first steps of a research pipeline is generating ideas. Large Language Models (LLMs) are increasingly used to automate this step, but most of this work optimizes the quality of individual ideas, while overlooking diversity. In o…

Anthropic 发布基于内部数据的技术报告，研究 AI 对自身雇佣与工作组织的影响。到 2026 年 8 月，交互式 AI agent 工时已达员工的 6 倍；技术部门初级员工占比从 21% 降至 14%，初级技术招聘占比从 27% 降至 9%。

Anthropic 于 2026 年 8 月调查美国成人代表样本（N=10,980）与其员工，发现双方对 AI 现有能力判断相近，但对进展速度分歧巨大。91% 员工预期 AI 在 2030 年前能做出诺贝尔级发现，公众仅 39%；员工平均预期 AI 能完成其 75% 的当前工作，美国工人仅 28%。

Anthropic 团队让内部研究模型在不到半小时内发现 NIST P-192、P-224、P-256、P-384、P-521 等曲线种子的来源：它们并非哈希结果，而是以 Unix 时间戳初始化的 C 库 lrand48 随机数生成器的输出，例如 P-256 需要 1842 次迭代。

Microsoft Research 发布零知识虚拟机 Magnetar，可直接证明用 Rust、Go 或 C 编写的 RISC-V 程序，在 15 个基准程序上平均比现有最快 zkVM 快 10 到 39 倍。在笔记本电脑上每秒可证明 720 万至 870 万条处理器指令，证明约 12 毫秒内可验证，安全性只依赖 SHA-256 或 SHA-3 等标准哈希函数，无需受信任设置或椭圆曲线。

Anthropic 发布两份报告，一份分析 AI 对公司内部就业的影响，另一份对比员工与公众对 AI 影响经济的预期。内部数据显示，过去一年初级岗位招聘占比下降，AI agent 工时迅速上升，到 2026 年 3 月 agent 工时已超过员工，8 月 agent 数量达员工 6 倍，人均代码产出、销售代表成单量和合同处理量均提升，瓶颈逐渐转向人类注意力。

Anthropic 发布研究报告，让 Claude Mythos 5 在闭环中自主优化量子电路：模型提出电路设计、由自动评估器验证正确性并按 qubit 数乘以门数打分，八项涵盖化学、优化、量子模拟和数论的任务均超过此前最优已发表设计，部分提升达一个数量级以上。

研究者在 TheoPhysEval 基准上测试商用 AI 智能体解决理论物理研究问题，OpenAI 的 GPT-6-Astra 在 Codex 环境中解出 50 题中的 49 题，平均每题不到 3 美元，全套基准总成本约 116 美元。

Google 与 Carnegie Mellon 的论文提出 Harmony Harness，将 agent harness 的自进化建模为多目标优化，把通过测试与阻断漏洞拆成不同目标分别进化，并在推理时按任务动态组合。在 SecRepoBench 未见崩溃类型上，66.7% 的修复既有效又阻断了漏洞，起始 harness 为 36.7%；安全收益保留 55%，DGM 基线为 0%。

阶跃星辰发布技术博客《Inside Step 5 Preview: Architecture and Training for Agentic Work》，介绍架构与训练如何协同设计以支持更长、更复杂的智能体任务。

ETH Zurich 等机构团队提出 LLM-to-Flow Distillation（LFD），让科学生成模型直接追随自然语言发现目标，如设计具抗癌活性的多肽。

小红书 AllSpark 团队提出 CompoWorld，通过组合可复用服务来扩展通用 Agent 的训练环境与任务空间，构建了 448 个服务、10,130 个工具。团队用 3K 条 SFT 轨迹和 1K 个 RL 任务训练 Qwen3.6-35B-A3B，在八项 Agent 基准上平均提升 9.17 分，AutomationBench 成功率从 10.33% 升至 32.33%。

MIT、Stanford、Google、Microsoft 等机构联合发布 Physical Recursive Self-Improvement 综述，指出机器人仍依赖人类采集数据、诊断故障并筛选更新。他们检查的 7 项机器人自我改进研究中，5 项未能在不同任务上展示提升，并定义 L0 到 L5 六个自主改进等级。其 2022 至 2026 年系统图谱中尚无确认的 L5 案例。

Large audio language models (LALMs) have made substantial progress in audio understanding. However, in high-throughput applications requiring frequent selection, judgment, and scoring, autoregressive text generation introduces additional c…

Learning intrinsic dynamics---the relationship between actions and their consequences---is essential for world models. However, visual variations unrelated to the underlying control process can obscure this relationship in learned represen…

Despite their practical significance, modern AI systems pose significant societal risks, including toxicity, malicious use, and misinformation. Their mitigation depends not only on technical feasibility but also on the incentives of firms …

Scientific inquiry involves iterative cycles of exploration, hypothesis generation, and reasoning over observations to uncover relationships in complex systems. Large language models (LLMs) have demonstrated strong capabilities in several …

Policy-shift distillation reuses the difference between a teacher's log probabilities before and after reinforcement learning (RL) as supervision on student-generated trajectories. However, this endpoint policy shift cannot reveal whether …

Reverse engineering the data formats used by mobile software is skilled and slow, and using AI to assist with it is attractive. We evaluate this for the protobuf messages that an iPhone sends to Apple's location services. Working from Appl…

Looped Transformers improve parameter efficiency by repeatedly applying the same Transformer layers. However, generating each token still involves sequential computation and repeated access to model weights, leading to high decoding latenc…

Machine learning (ML) for analog IC design is constrained by limited open data and the absence of standardized evaluation: results reported in different papers are typically obtained on bespoke, often closed, datasets and with ad-hoc metri…

Indoor service robots often enter unmapped buildings whose layouts and task constraints are already familiar to users. We address this human-familiar, robot-unseen setting with AVALON, a topological navigation framework organized around hu…

Transfer learning in stochastic multi-armed bandits is typically studied assuming that prior data forms a fixed offline dataset or is generated by a known source. In practice, historical observations come from a previous adaptive decision …

Mathematical agents are commonly evaluated by answer accuracy or proof acceptance. We introduce ClaimScope, a 72-task benchmark for auditing proposed claims through a verdict, scope, counterexample, and repair. Tasks cover optimizer-state …

Recent models generate sewing patterns from images or text, and program-based representations make these patterns executable. These systems produce a finished pattern and do not model how a patternmaker turns one pattern into another. In p…

Large language model (LLM) agents solve complex tasks by calling external tools over multiple turns. Recent work has proposed tool graph memory, which stores past experience as tool transitions to guide subsequent selection. However, it re…

LLM agents can write directed tests that close functional coverage, but coverage only shows that a state was reached, not that the data on the way there was correct. This is a long-known point in verification and we do not claim it as new.…

Recent multimodal large language models have advanced video understanding, but existing benchmarks mainly evaluate whether models can interpret observed visual content. In real-world scenarios, assistants must additionally anticipate what …

Query-conditioned probabilistic predictors answer logically related questions one at a time, and labels to check their calibration are often missing. For each restriction on what a bettor's stakes may use, a minimax theorem identifies, for…

On-device deployment of large language models (LLMs) is increasingly adopted due to latency and privacy requirements, but exposes proprietary models to reverse engineering and intellectual property theft. Trusted Execution Environments (TE…

We use quantum machine learning techniques to develop a content-based media recommender system and evaluate its performance on TV programmes using textual descriptions from one week of BBC iPlayer data. Grounded in the Quantum Natural Lang…

Large language models (LLMs) deployed as conversational agents are increasingly expected to follow procedural logic. While factual hallucination has received extensive attention, a different failure mode has been overlooked: procedural hal…

Preparing quantum states in bosonic qudits with SNAP--Displacement sequences requires both optimizing the gate parameters and deciding which Fock levels should be controlled by each SNAP gate. A fully parameterized ansatz assigns an indepe…

OpenAI 公布 GPT-Live-1.1 与 GPT-Live-1.1 Quip 的系统卡，说明两者训练数据来自公开互联网信息、第三方合作数据以及用户和人类训练师提供或生成的数据。数据处理流程包含严格过滤以保障质量并降低风险，同时用安全分类器减少有害或敏感内容的使用。

上海人工智能实验室发文宣布，许地生团队依托书生科学多模态大模型 S2，通过人机协同完成困扰数学界 30 余年的三维 Ghys 猜想证明。北京大学董彬、刘若川院士团队用 SHEAF 框架完成形式化验证，论文已发布于 arXiv，形式化代码开源在 GitHub。上海人工智能实验室表示该范式可推广至其他搜索空间大而专家稀缺的基础研究领域。

Kombai 发布 Design Creativity Bench，用 80 个开放式 UI 任务评估模型设计创造力，最佳模型 GPT-6 Astra 得 65.5，人类参考为 92.3。