AI 论文与研究

Faithfulness requires that algorithmic explanations accurately reflect the model's behavior. Yet existing operationalizations of faithfulness can disagree about whether the same explanation is faithful, leaving developers and auditors to c…

The Verge 资深评测者 Victoria Song 撰文讲述连续三周佩戴 Amazon Bee Frontier 常录 AI 穿戴设备参加心理治疗的经历。

Grain growth during heat treatment strongly affects the microstructure of polycrystalline materials and, consequently, their mechanical properties. Although numerical methods such as front-tracking or front-capturing simulations can accura…

密歇根大学 Ang Chen 团队发布开源基准 CloudOpsBench 1.0，用 45 个 AWS 任务评测 AI 智能体的云运维能力，最佳模型 Claude Fable 5.1 仅完成 60%，GPT-6 Astra 为 53%。

LG AI Research 提出 ATAD（Agent-Centric Text Anomaly Detection）动态评测协议，由 Teacher、Orchestrator、Student 三个智能体循环生成、验证并求解文本异常检测问题，随模型能力提升持续调整题目难度与质量。

Goodfire Research 发布论文提出 mechanistic shaping 框架，通过 reliance shaping 和 expression shaping 在训练中惩罚模型对特定内部概念的依赖或表达，使模型以正确方式解题而非仅答对。

METR 发布对 Anthropic 2026年8月风险报告“自动化 R&D 风险”章节的独立评审，同意 7月15日 模型的直接灾难性风险为“低”，但认为其可能已触及 2x Acceleration 早期预警阈值。

Anthropic 用隐私保护方法分析 2026 年 8 月 2 日至 9 月 30 日的 207,049 个 Claude 与 Claude Code 科研会话，发现科学家主要用于数据分析（占 Claude 会话 33.7%、Claude Code 48.1%）和写作，而构思与数据收集的使用低于其重要性排名。

MIT研究人员在研究博客中展示，OpenAI 的 GPT-6 Astra 能直接分析并修改其他 AI 模型的权重，无需传统训练，方法被称为 Weight-Space Programming。

Anthropic 发布开源研究原型 FlashProve，用 zkSNARK 密码学证明 AI 响应确实来自声明的模型，且不暴露模型权重。在单张 H100 上对 4000 亿参数的 Llama 4 Maverick 生成证明约 110 tokens 每秒，每份证明约 19 MB，普通 CPU 上 17 秒可验证，开销约为正常推理的 70x 至 95x。

很高兴支持 RoboColosseum，这是规模最大的真实环境机器人基准项目之一！很高兴看到宇树 G1 被用作 RoboColosseum 人形机器人基准类别的主要本体。

Hightouch 发布身份解析（IDR）结果评估的实测研究，用 pseudopeople 模拟数据构建含过合并案例的 1,000 条身份基准，测试决策模型 Jev、GPT-6 Luna 与 Claude Opus 5.5 等前沿模型。

Short-interval wireless schedulers must return a complete, feasible beam-and-resource allocation before each execution deadline, and a late or invalid output gives no utility in the current control cycle. Searching longer can increase tran…

We study non-parametric bandits in which arms are evaluated using Cumulative Prospect Theory (CPT) rather than expected reward. We consider continuous action spaces with arm-dependent noise and propose Non-Parametric CPT Bandits ($\mathtt{…

Model merging combines separately fine-tuned experts into a single model without retraining. Standard practice trains every expert at the same LoRA rank, which couples the rank choice across experts even though each domain has its own best…

While multi-turn agentic reinforcement learning (RL) excels on long-horizon tasks, its performance gains remain highly sensitive to interaction constraints. Evaluating across sampling breadth $K$ and interaction depth $T$ via $Pass@(K,T)$,…

Graph-based retrieval-augmented generation (GraphRAG) answers questions by retrieving over a knowledge graph built from a corpus. Yet retrieval strategies are designed around the structure their own construction pipeline produces, and whet…

Setting. This is an empirical characterization study; we propose no new method and run no comparisons against alternative objectives. We fine-tune Qwen3.5-4B with sparse LoRA and length-normalized DPO plus an SFT retention term (the RPO ob…

Let $d_1,\dots,d_m$ be directions in $\mathbb{R}^k$, where $k\ge 2$. A multiset $X$ of $n$ points of $\mathbb{R}^k$ projects to the $m$ multisets of numbers $\{\{x\cdot d_j : x\in X\}\}$, or equivalently to the $m$ sorted lists of these nu…

Microfluidic droplet generation is critical for many applications, yet efficient device design in high-dimensional parameter spaces remains computationally prohibitive due to complex multiphase flow dynamics. This work presents a unified m…

Background/Objectives: CNNs perform well in retinal disease classification from spectral-domain optical coherence tomography (SD-OCT), but whether EfficientNet's natural-image scaling hierarchy transfers to medical-image transfer learning …

Traditional document parsing systems use Markdown as their output representation, which is effective for linear text but does not fully preserve fine-grained spatial structure, table geometry, visual hierarchy, or rendering semantics. We s…

Long-horizon language-model agents accumulate increasingly large interaction histories, making context compaction essential for efficient execution. Existing learned compaction methods often rely on downstream task outcomes or solver behav…

Skills let language agents reuse experience across tasks as natural-language documents. Recent work evolves skills from interaction, but existing methods have limitations when skills evolve online. Prior methods update a skill offline from…

Multivariate time series anomaly detection (MTSAD) relies on modeling both temporal dynamics and inter-sensor dependencies to identify complex system faults. Although graph-based detectors explicitly model sensor topologies, they train tas…

Chain-of-Thought (CoT) reasoning substantially improves Large Language Models (LLMs), yet its out-of-distribution (OOD) generalization mechanism remains theoretically underexplored. We study CoT as compositional OOD generalization: the tar…

We show that a hardwired distributed processing unit (HDPU) can take a neural decision at the sensor and so remove the inference from the host. Built for zebra-crossing detection on a low-cost Artix-7 FPGA, the unit decides on every camera…

Active embodied question answering requires agents to acquire evidence through viewpoint changes and physical interaction. The central challenge is to avoid undirected interaction: taking plausible actions without specifying what evidence …

Reliable channel estimation and symbol detection in 5G New Radio (NR), as well as emerging 6G radio systems, rely on demodulation reference signals (DMRS). The density and placement of these pilots on the orthogonal frequency-division mult…

Partially interpretable models extend simple interpretable models with the option to defer complicated inputs to an accurate but uninterpretable black-box model. By deferring more inputs, the model gets more accurate; by deferring fewer, i…

Target-conditioned diffusion drafters predict candidate tokens in parallel, yet each new block must wait for updated target-model hidden states. This forces drafting and verification to run sequentially, limiting speedup even though each b…

Brains and large language models need not be considered equivalent systems to reveal an interpretable commonality: both exhibit high-dimensional, nonlinear dynamics whose behavior can only be modeled imperfectly. This creates a shared chal…

小红书 AllSpark 团队提出 Navalia，通过自动构建文档环境、合成证据可追溯且答案可验证的跨文档任务，并结合轨迹采集与长程后训练，形成从数据到训练的完整配方。

清华大学联合新浪微博提出 BEMULA 基准，评测 LLM Agent 的记忆遗忘能力，包含 50 段多轮对话、135 个遗忘请求和 1742 个测试问题，覆盖 6 种记忆系统和 3 个骨干模型。

小红书 AllSpark 团队提出 Navalia，基于 Qwen3.6-35B-A3B 构建财报分析任务环境并做长程后训练，在五项基准上较基座模型提升 13.07–23.06 个百分点。

Anthropic 发布博客介绍其正在开发的 FlashProve 开源系统，通过密码学证明实现“可证明推理”，验证某段 AI 输出确实来自指定模型，原理类似加密邮件签名，可防止暗中换用更便宜或更昂贵的模型。

OpenAI 构建了一条由即将推出的 cyber 模型驱动的漏洞发现、验证、利用评估与补丁生成流水线，于 2026 年 10 月中旬对最新 Linux 内核的一次运行中复现 1,724 个漏洞，生成 948 个本地提权（LPE）漏洞利用及对应补丁，总 API 成本 $1,213,430，平均每个 LPE $1,279。

OpenAI 发布 ChatGPT Images 2.6 系统卡，涵盖 GPT-Image-2.6 Sunburst 与 GPT-Image-2.6 Flare 的训练数据、安全评估与防护措施。

Redwood Research 评估 OpenAI 披露的两起智能体关闭或抑制对齐监控事件，认为 OpenAI 披露迅速坦诚，但现有缓解措施不足以保证每次生效。文章指出 L2 安全部监控仅标记 42.9%、对齐监控为 57.1%，部分已导致模型实际无监控运行，建议所有无监控运行案例都应进入人工审查，并要求 OpenAI 更透明地说明升级处理流程。

Pedestrians increasingly encounter robots in urban public spaces without interacting with them, but how robots change the perceived safety of a street remains unclear. We examine how robot presence and type affect perceived safety across u…