AI 论文与研究

Anthropic 在 Frontier Red Team 博客发布研究，介绍用 Claude 大规模进行形式化验证的方法。与 Theorem Labs 合作，团队在数周内为超过一百个真实开源库证明了安全属性，突破了形式化验证过去仅适用于小型代码库或需大量专家人工的限制。全文见 https://www.anthropic.com/research/formal-verification。

MIT CSAIL 与天津大学研究人员开发出 X-Hinges 设计软件，可自动在 3D 打印物件中嵌入两种不同材料，分别用作传感器和导线以避免信号干扰。该工具支持压缩、上下、左右最多三个运动轴，并同步采集各方向运动数据，已用于制作可教机器人抓取积木的手套、折纸台灯和物体识别装置。目前每件打印品的电路布局存在细微差异，需借助计算机视觉系统重新校准传感器输出。

Google DeepMind 与 UPenn 的论文提出用稀疏自编码器解析 Gemma-2-2B 检测器内部激活，让 LLM 标注特征后训练简单分类器，构建可解释的 AI 文本检测器。

微软团队在 Science 发表研究，展示其诊断编排系统如何推理复杂医疗病例，在提升诊断准确率的同时降低检测成本。引用内容称 MAI Diagnostic 比可上网查资料的急诊科医生准确 1.7 倍，且成本更低、所需医学检查更少。

Microsoft AI 在 Science 发表研究，称其诊断编排系统 MAI-DxO 在 NEJM 复杂病例模拟中准确率达 83%，而单独工作的全科医生为 36%。

Mustafa Suleyman 宣布 Microsoft AI 的 MAI Diagnostic 系统诊断准确率达到急诊医生的 1.7 倍，且医生可联网查资料时依然如此。该系统成本更低、得出诊断所需的医学检查更少，相关成果今日发表于 Science 期刊。

Anthropic 至少自 2022 年起就由 @logangraham 主导预警 AI 助力的网络攻击，当时没有 VC 认真对待。Anthropic 还在 ChatGPT 真实事件让"模型过度迎合用户"风险变得可感知之前数年，就由 @EthanJPerez 等人构建了谄媚评测。推文认为教训是：应在风险造成大规模伤害前就重视早期证据。

Microsoft AI 发布 Sequential Diagnosis Benchmark（SDBench）和 AI 诊断系统 MAI-DxO，论文发表于 Science。

Google DeepMind 新论文提出让各模型为回答每个问题“出价”、由最高出价者作答的拍卖机制，在 110 个模型、超 30,000 条 EmbedLLM 提示词上模拟，效果达最优模型选择的 99.2%，而按提示词类型路由到最佳模型仅 89.4%。

Microsoft 在 Science 发表研究，称其 Diagnostic Orchestrator 在 SDBench 基准上诊断准确率最高达人类医生的四倍，计划开源该基准。

Robotic foundation models are typically evaluated through costly closed-loop rollouts, which motivates cheap offline proxies for model capability. The Platonic Representation Hypothesis (PRH) suggests that as representations improve, diffe…

Test-time scaling has emerged as a key inference strategy for reasoning models, where a model generates a large number of candidate solutions in the hope of finding at least one correct solution. Running all candidate solutions to completi…

Constructing accurate and computationally tractable approximation spaces is a fundamental task in scientific computing. Partially trainable neural networks bridge fixed random features and fully adaptive representations. This paper quantif…

We consider online monitoring of an AI system that produces a sequence of decisions. At each round, the system outputs a proposed decision together with a confidence score estimating the trustworthiness of the decision. The decision and co…

Auditing whether a text-to-image (T2I) model was fine-tuned on a particular image collection is difficult: a model can absorb collection-level visual characteristics without reproducing training images. Proactive dataset-use auditing plant…

Wrong automatic routing of an IT support ticket can cost more than human triage. We evaluate cost-aware abstention as an offline, cost-conditional readiness assessment: select a routing policy on calibration data under explicit costs, then…

Meta-reinforcement learning aims to train agents that adapt quickly to unseen tasks from limited interaction, but this requires representing task evidence over long interaction histories. Classical recurrent agents compress experience into…

Tool-using language model agents rely on environmental facts to decide whether an action is permitted. When a tool response omits a safety-relevant fact, an agent may carry out a prohibited action even though the user request is benign and…

Many recent open-weight language models use n-gram embedding tables to improve performance. These tables store learned representations of short token sequences, such as pairs or triples of consecutive tokens, which the model retrieves duri…

Standard implementations of softmax, log-softmax, and cross-entropy have two numerical bugs. The first bug, Softmax Collapse, rounds the target logit's gradient to zero if that one logit is much larger than the others, resulting in gradien…

Chain-of-thought (CoT) prompting elicits step-by-step rationales that appear to explain model predictions, but there is no guarantee these steps causally drive the final answer. In standard autoregressive generation, dense attention allows…

Multilingual pipelines such as FineWeb2 apply one recipe across a thousand languages, leaving the failure modes of individual languages unanticipated. We use an LLM once, as a corpus debugger rather than a classifier: a tool-using agent ex…

Looped architectures scale quality while keeping parameter count fixed by repeatedly passing activations through the same block of layers. During inference, each recurrence generates independent KV-caches, causing memory to grow linearly. …

Instrumental convergence is often invoked as a general reason to expect capable agents to acquire or express power-seeking subgoals: self-preservation, resource acquisition, deception, or resistance to oversight which poses an important th…

While monolingual language model (LM) surprisal is widely established as a predictor of reading times, it cannot capture the non-native (L2) reading dynamics that are shaped by the reader's first language (L1). In this work, we investigate…

On-policy distillation has been widely adopted for alleviating the distribution mismatch of off-policy distillation. Self-distillation further removes the need for an external teacher by using the student's own outputs, enhanced with extra…

We study learning a discrete distribution under person-level differential privacy in the batched setting. Given $n$ users, each contributing $m$ i.i.d. samples from an unknown distribution $μ$ over $[k]$, we seek an estimate of $μ$ with to…

Purpose: To describe an evidence-aware retrieval evaluation protocol for a clinical retrieval-augmented generation (RAG) system over breast lesion management literature, addressing the dependence of retrieval scores on annotation coverage.…

Counterfactual explanation methods have been introduced for unimodal settings, with dedicated frameworks for tabular, timeseries, text, and image data, yet no post-hoc method has been designed for classifiers that learn from multimodal inp…

Control evaluations measure whether safeguards can prevent an AI model from causing undesired outcomes even when it is intentionally trying to do so. These evaluations typically use purpose-built control settings and extrapolate their resu…

METR 和 Redwood Research 各一名员工今年夏天在 OpenAI 驻场三周，评估其内部智能体错位监控系统，简报现已公开。

严重推荐！ FrontierCS 2：计算机科学研究基准

Anthropic 发布对 Claude Mythos 5 对齐问题的评估报告。

METR 与 Redwood Research 各派一名员工，于 7 月 21 日至 8 月 11 日对 OpenAI 内部失准监测系统开展三周高权限试点压力测试，产出 61 页内部报告。

Google DeepMind 与 Google Research 联合 World Resources Institute 和 IIASA 发布 Forest Typology（ForTy）2020 全球地图，以 10 米分辨率将地球每块土地划分为 FAO 定义的六类森林与土地类型，并首次提供 2017 至 2025 年逐年变化时间序列，展示天然森林被种植园或农场取代等变化。

斯坦福与伯克利研究者提出可插拔小模块 TokenGate，插在现有 LLM 浅层学习"跳读"长上下文，只把 top-K token 传给上层。论文报告无精度损失时首 token 时间快 1.4–11.3 倍、峰值显存降 1.5–3.4 倍，相对最强免训练剪枝最高约 6.4 倍更快、2 倍更省显存。训练成本接近 LoRA，无需从零预训练。

Sierra 发布 𝜏-Entity 基准，测量语音智能体能否正确采集、识别错误并修复姓名、电话号码、地址和邮箱等细节，包含 10 种实体类型共 200 个任务，并与 𝜏-i18n 一样覆盖六种语言。

大语言模型在数学和科学上越来越强。但令人意外的是，许多模型连一个基础谜语都解不开。 @Cohere_Labs 的 CulturalRiddles 是一个新基准，将文化作为相互交织的生活体验来衡量。秉持开放科学精神，它由一支意想不到的团队打造。

Google DeepMind 与 Cambridge 的论文展示了一个基于 Gemini 的智能体能够自主发现新数学，为 11 台此前无任何公开分析的 Busy Beaver 图灵机找到并用 Lean 验证了精确规则。该智能体先在模拟器中测试边界情况，起草非形式化证明，再用 Lean 证明检查器完成形式化证明。

OpenAI 与 PerturbAI、Scripps Research 合作，用数千个基于 GPT-Rosalind 的智能体在 32 分钟内完成对数百万小鼠细胞的全脑 CRISPR 图谱分析，人工估计需数月至一年。