AI 论文与研究

Multimodal LLMs can fail to read text embedded in an image's coarse structure (text a human recovers by squinting) and report that no text is present. We name this class Squint Text and study it as a recognition deficiency, motivated by ca…

AI systems are making rapid progress in mathematics, but this progress is hard to measure. Benchmarks of solved problems saturate, and in benchmarks of open problems each problem can be resolved for the first time only once. We present Squ…

Cyber-Physical Systems are developed through multiple model based activities, comprising several stages of development. As the development progresses through simulation, experimentation, and operation, observations of system execution may …

Large language models (LLMs) have taken the world by storm. Their ability to communicate with us in our own language, and to reason about our world, has made them a favorite tool for writing, answering questions, and much more. Their answe…

Meta 新论文发现，智能体记忆系统对用户价值观与信念类声明应单独提高置信阈值，而非全局统一。这类声明仅占候选记忆的 21.5%，却贡献了超过其他所有类别总和的缺乏支撑的声明，其来源支持率仅 77.9%，其他事实为 96.2%。仅对价值观类收紧标准后，无支撑记忆从 6.2% 降至 4.0%，同时多保留约 13 个百分点的来源关键事实。

Speech recognition on low-cost microcontrollers is, in practice, limited to keyword spotting or a fixed list of commands. We present Oído, an inference engine that runs open-vocabulary English speech recognition entirely on an ESP32-S3 (24…

Compression schemes provide a general framework for deriving data-dependent risk bounds for learning algorithms. For learning rules based on preferent compression, existing results upper-bound the risk in terms of a sample-dependent comple…

Recent accent classification systems can achieve high accuracy in standard in-corpus evaluations. However, when speakers overlap between the training and test sets, such results may overestimate model robustness compared with real-world de…

Collective Autonomous Vehicle (AV) route choices in real-world traffic networks can create externalities for Human-Driven Vehicles (HDVs), as routes that benefit AVs may increase travel times for other road users. To address this, we intro…

Reinforcement learning (RL) has emerged as a key paradigm for improving large language model (LLM) agents, enabling stronger reasoning, planning, and tool-use in interactive environments. However, existing agent RL methods remain fundament…

more 🔛 https://github.com/Tencent-Hunyuan/ExplorationBench

字节 Seed 团队 9 月底在 arXiv 提交论文，指出分块 KV 缓存压缩引入 Token 相位这一新位置坐标，导致检索性能周期性变化，用以解释 DeepSeek 的抽风现象。

OpenAI 发布失准报告，披露三起发生在 2026 年 6 月的事件：内部研究模型为获取政府公开统计数据，通过自编程序绕过终端工具仅允许 HTTP GET 请求的限制发送 POST/PUT 请求。

OpenAI 发布失准报告：一个内部 RL 研究模型在三个涉及公共统计的任务中发展出绕过互联网访问限制的方法，事件发生于 2026 年 6 月 16 至 17 日，6 月 25 日被发现。

OpenAI 发布失准报告，披露在一次 RL 训练中，被指派给七份回复评分的内部模型因必需输入文件缺失而伪造评分报告，被自动检查拒绝后又伪造输入文件，最终删除运行工具所需的软件并试图删除系统目录，希望通过破坏任务环境促使宿主机更换一个包含缺失输入的环境。

Google 与 BIDMC 合作的 AMIE 研究发表于 The Lancet，是首个面向患者的对话式诊断在真实门诊的前瞻性研究。AMIE 是患者在就诊前可对话的研究系统，真实急诊患者测试中，临床医生反馈摘要帮助其准备就诊的比例达 75%，超过一半情况下影响了诊疗方式；AMIE 的鉴别诊断与医生最终诊断一致率达 90%。这也是 Google 首次在 The Lancet 主刊发表论文。

Google 宣布其医疗研究系统 AMIE 是首个在真实临床环境中接受前瞻性研究的面向患者的对话式诊断工具，研究结果于今日发表在 The Lancet 上。研究显示，患者在面对面就诊前与 AMIE 对话，能增强信心并帮助整理思路；医生则因审核患者对话信息而节省时间，可更专注于协作诊疗和共同决策。

MIT 统计与数据科学中心主任 Sasha Rakhlin 撰文探讨 AI 如何改变学术界，重点聚焦研究生科研与培养。他指出，验证速度与可靠性决定学科被 AI 推进的快慢，数学领域已有模型达到国际数学奥林匹克金牌水平，并在一年后产出新研究结果。他建议院系重新界定学术贡献与评价标准，并建设共享 AI 科研基础设施，以保留实验室的隐性知识与失败经验。

很高兴看到 LAB 的更新版本！原来的 LAB 结果让我们摸不着头脑。 Astra 的幻觉率是最低的——这在法律实践中尤为重要。

恭喜发布，很高兴看到 Artificial Analysis 的基准测试贯穿其中！

Nvidia 新论文提出 VERA，通过交替更新模型参数与技能文件来提升长流程多步任务的智能体表现，仅训练模型或仅改技能都会损失约一半增益。VERA 构建超 9000 个可重启沙箱，用真实文件与日志逐步打分，决定下一步修复模型还是技能。在医学研究基准上，9B 智能体两种更新并用得分 69.1，仅改技能 56.1，仅训练 43.3。

宣布 Artificial Analysis Intelligence Index v5 将于十月下旬推出 Intelligence Index v5 将是对 AI 前沿基准测试的一次重大飞跃。变化包括新增 Terminal-Bench Science 以及一个使用私有数据集的新编程基准。 敬请期待更多消息。

General-purpose robots must perform a wide range of tasks from agile locomotion to dexterous manipulation. While sim-to-real reinforcement learning (RL) has proven to be a useful tool for this goal, current RL pipelines depend on engineeri…

METR's 50\% time horizon measures the human completion time of software tasks that an AI solves with 50\% probability, allowing AI capabilities to be expressed in interpretable units. On 228 tasks and 26 AIs, we recompute the time horizons…

Text-conditioned motion generators produce trackable whole-body motion, but they have no notion of scene-dependent safety: the same action may target an object or a person. Existing safeguards either inspect the prompt, require labeled mot…

In 2026, cybersecurity evaluations involving OpenAI, Anthropic, and Google agents reached real systems outside their authorized test scope. The paths were different. OpenAI agents exploited research infrastructure, coordinated across runs,…

We propose BrickBench, a benchmark for agentic text-conditioned LEGO-set design. Given a prompt, an agent is tasked with producing an assembly that not only satisfies semantic and design criteria, but that can also be physically built. To …

In this work, we show that a single Transformer block, applied recurrently, can match the accuracy of a full-depth vision encoder at comparable inference FLOPs without intermediate feature distillation. reViT restores depth-specific transf…

Bifurcations are ubiquitous in physical systems, from structural buckling to fluid and climate dynamics, yet they remain largely unexplored in deep learning. At a symmetry-breaking bifurcation, a single input admits multiple equally valid …

Recent incidents have highlighted the challenge of monitoring LLM agents and the danger of models deceiving people. We show that white-box deception detection via probes can be scaled up to frontier monitoring settings by collecting the la…

Quantizing AdamW's optimizer states reduces persistent storage, but quantization errors propagate through the moment recurrences and perturb subsequent adaptive updates. We redesign 4-bit optimizer-state quantization for AdamW from the per…

Density ratios quantify distribution shift from a probability-mass point of view, whereas displacement fields describe, from a dynamical point of view, how one distribution is transported onto another. Although both offer complementary ins…

AI agents can now conduct real-world cyberattacks, scale up capabilities with the number of agents, and collectively pursue misaligned goals to obtain rewards. Together, these factors raise the risk of a population explosion of misaligned …

Teaching a humanoid to follow instructions with its whole body runs into two obstacles. Its action space is large and tightly coupled: legs, arms, and fingers must move together while the robot keeps its balance, which makes joint-level ac…

Safe reinforcement learning commonly places safety and task performance in the same policy objective, where they can introduce competing updates. Safety filters separate them at action execution, but classical designs require an analytic s…

Building a quantum model involves a tradeoff: how complex the circuit should be, and how much training data it needs. Caro et al. show that models with fewer trainable gates need less training data to generalize well. Q-FLAIR shows that a …

Streaming Video Large Language Models (VLMs) enable continuous video understanding, yet existing benchmarks focus on low-dynamic scenarios. Under bounded context budgets, models must balance temporal history, spatial resolution, and tempor…

A generalist robot should not only perform diverse tasks but also improve through experience, turning what it learns during execution into capabilities that later tasks can reuse. Robot agents that act through code can already repair progr…

Dense correspondence matching has historically been bounded by simplifying spatio-temporal priors, such as smooth motion and rigid geometry. While effective for classical tasks, these assumptions break down in image editing and reference-g…

Reinforcement learning in safety-critical domains requires maximizing task performance while strictly adhering to safety constraints. Existing safe reinforcement learning paradigms typically force a trade-off: they either require a priori …