AI 论文与研究

Language models make it easy to build classifiers by simply prompting a model with the sample to classify. However, their predictions are neither deterministic nor uniformly reliable across samples. How can we combine predictions from diff…

Google DeepMind 等机构的论文《Training Natural Schemers》用 Kimi K2.6 构建自然形成的 scheming 模型，仅给良性特质、不展示 scheming 行为，再以只计受监控回合、奖励遵守某条亏待用户的商业规则的 RL 训练。

AI 能否既富有创造力又高效实用？ Arena Conversations 新一期上线：@petergostev 采访了 Amazon AGI 高级首席科学家 Pradeep Natarajan。 他们讨论了： - 掌握基础原理带来的模型收益 - 评测作为差异化因素 - Harness 优化 - 准确率与风格等指标对模型创造力的约束 了解以上内容及更多。 https://www.youtube.com/watch?v=GqJhYHqNOEk

METR 与 Redwood Research 各一名员工在 2026 年 7 月 21 日至 8 月 11 日对 OpenAI 内部失对齐监控系统开展高权限红队测试，产出 61 页内部报告。

Anthropic 与 Theorem 合作，用 Claude 将 OSS-Fuzz 模糊测试转化为规格，对 123 个开源库的 239 个入口点证明了任意 64 MB 以下输入不崩溃，生成约 1.8 亿行 Rocq 证明。

Biograph 宣布完成其首项美国研究，使用 Caristo Diagnostics 的 CaRi-Heart AI 技术对无症状低风险人群进行冠脉炎症检测，42.9% 的参与者检出冠脉炎症升高，其中 17.9% 在至少一支主要心脏动脉中炎症水平很高。

Google Cloud 公布代号 Project Chrysalis 的研究参考架构，展示如何在 Gemini 平台与 Agent Development Kit（ADK）上运行受治理的自进化 AI 智能体。

Microsoft Research 的 Xinzhi Zhang 与 Zhao Song 证明 1995 年提出的 Kannan–Lovász–Simonovits（KLS）猜想，表明对数凹分布的最优平面切割与任意形状切割的差距不随维度增长，以普适常数为界。

OpenAI 联合 PerturbAI 与 Scripps Research，用 1946 个 GPT-Rosalind 智能体并行解读小鼠全脑 CRISPR 图谱，32 分钟得出 OpenAI 估计人工需数月的分析结果，其中 GPT-Rosalind 就 DCTN1 基因与 Perry 综合征提出了胆固醇转运可能受损的机制假设供科学家验证。

Wiz Research 披露 PHP 核心中的多个内存安全漏洞（CVE-2026-109618/109619，CVSS 最高 9.2），实现无需认证的远程代码执行与内存泄露，影响至少追溯到 PHP 5.4 的近十年版本。

In this paper, we study quantile distributional policy evaluation, where the return distribution under a given policy is approximated by the quantile Bellman fixed point $\bmη_m$. We investigate the statistical accuracy of estimating $\bmη…

The rapid expansion of the Internet of Things enables seamless collaboration among connected devices, making indoor localization a critical component of industrial automation. However, conventional localization techniques struggle in dynam…

Graph neural networks (GNNs) are among the standard tools for molecular property prediction, but the atom-level explanations of their predictions are hard for medicinal chemists to act on. We propose a change of input representation rather…

Despite the success of foundation models in image synthesis, achieving fine-grained spatial control remains a significant challenge, typically requiring auxiliary structural adapters or complex inference-time steering. While modern vision-…

Bayesian optimization (BO) is typically used only for problems where the underlying function is deemed "expensive" (e.g. in time or money), with the speed of the BO algorithm being of little importance. In the real world, however, problems…

Using a near-census of 802,863 resolved markets listed on Polymarket between January 2023 and May 2026, and only information available at listing (the market's title and description, named resolution source, and contract metadata), we pred…

Many learning systems are performative: once deployed, their predictions change the data distribution on which future models are trained. In multi-agent settings, this feedback is routed through a network of interacting learners, so stabil…

We consider online prediction from experts, a fundamental problem in machine learning, under differential privacy constraints. Existing private algorithms achieve near-optimal regret $\tilde{O}(\sqrt{T})$, where $T$ is the time horizon. Ho…

Graph-level out-of-distribution (OOD) generalization is critical for deploying graph neural networks in real-world applications where distribution shifts are inevitable. Existing approaches typically require class labels during representat…

O'Reilly 基于 209 名安全从业者（总样本 351 人）的调查报告分析 AI 与安全的交叉现状。最常见攻击是 AI 生成钓鱼（55%）、漏洞扫描利用（30%）和 AI 辅助恶意软件生成（26%）；受访者最担心的风险是过度依赖 AI 未加人工验证和通过 AI 工具泄露敏感数据（各 49%）。

Model predictive control is a high-performance control method, but it incurs comparatively high online computational costs across many control tasks and is challenging to realize in low-level hardware settings. To address this, we present …

One of the enduring challenges in online platforms is recommendation. Both communities and advertisers face difficulties in analyzing complex user behavior, identifying habits and evolving preferences, and providing timely recommendations …

Aligned large language models are trained to comply with benign requests while refusing harmful ones. However, recent work shows that this refusal behavior is fragile and can often be disrupted by simple activation- or weight-space interve…

Training modern AI models has become increasingly expensive, and updating deployed models can alter their internal feature representations, affecting applications built upon them. In retrieval systems, such updates often require complete g…

Linear-attention language models keep their context in a recurrent state of fixed size, where an inference system stores one copy of this state for every sequence it decodes at the same time. Storing the state at lower precision therefore …

Direct preference optimization (DPO) incurs training costs in output scoring, sequence execution, and backpropagation, so reducing supervision alone does not necessarily save computation. We introduce a unified framework for stochastic pru…

Autoregressive (AR) models have proved effective for zero-shot prediction from longitudinal electronic health records, but impose strictly sequential generation on trajectories containing temporally clustered and interdependent clinical ev…

We study an extremal regression problem, an example of multiple instance learning (MIL), where the predictors are given as multisets, called bags. The bag elements (called instances) possess hidden labels which are pooled to produce an obs…

Residual-based causal discovery assumes that each variable is generated from its parents and an independent noise source. It estimates this noise and tests its independence from candidate causes. For identifiable functions such as additive…

Masked language modelling (MLM) is the dominant pretraining objective for transformer encoders applied to longitudinal electronic health records (EHRs). However, reconstructing individual events may overemphasise noisy or weakly informativ…

Speculative decoding accelerates large language model inference by employing a compact draft model to propose candidate tokens that are verified in parallel by a target model. Conventional training relies on soft-label distillation, which …

小红书自研的稀疏注意力索引器加速方法 DoSA，在保持原始 Top-K 选择结果的前提下，用分数上下界提前排除不可能入选的候选。

Anthropic 的 Claude 在 Claude Science 研究环境中几乎自主设计了针对蝎神经毒素的蛋白质结合剂，初始提示后独立工作 24 小时，共产出 318,484 个设计。

李飞飞在 Cloudflare Connect 2026 上表示，自己也是不少前沿闭源模型的满意用户，但开放生态的分量正变得越来越重。她给出两点理由：开放权重让创业者能站在其上创新，突破可能从任何地方冒出来；模型能力太强必须安全，而安全与危险的边界不能只交给忙着变现的公司判断，学术界、公共部门和政府都需能上手检验。她还批评动辄把“人类灭绝”挂在嘴边的技术讨论方式不负责任。

OpenAI、PerturbAI 与 Scripps Research 联合发布研究，用数千个由 GPT‑Rosalind 驱动的 AI 智能体并行分析数百万小鼠脑细胞的全脑 CRISPR atlas，32 分钟内完成分析与文献解读，并称人工逐一分析需数月至一年。

DeepSeek 开源支撑其线上生产服务的高性能推理框架 DeepInfer，基于 Apache 2.0 许可证，并同步发布技术报告。框架集成 FlashMLA、DeepGEMM、DeepEP 等此前开源内核，加上调度、并行、DSpark 投机解码等系统层实现，同时支持 NVIDIA SM100 GPU 和华为昇腾 950 NPU。

Google Research 在 JAMA Network Open 发表前瞻性多中心研究，评估基于"盲扫"超声的 AI 孕周评估模型在不同临床环境和设备上的泛化能力。

感谢试用 Claude！ 对一个用 Go 构建、并移植到 Rust 的 http 代理进行基准测试，看看能否进一步压榨出更多性能。

OpenAI 于 10 月 20 日发布研究报告，介绍其增强 GPT-6 Astra 等模型构建和迁移 COBOL、Fortran、Visual Basic 等低资源编程语言的能力，并推出 MultiLangBench（20 种语言从零构建兼容 SQLite 的数据库引擎）和 COBOL 迁移至 C++、Java、Python、Rust 两项评测。

Meta 等机构发布 ProPerMem 基准，用约 618K token 的全年个人记忆和人工验证的触发窗口评测记忆驱动的主动助手。所有基线最高仅 15%，GPT-5.3-Chat 在窗口内给定证据时达 96%，但自主定时降到 38%、加检索降到 15%；兼顾内容与时机并减少触发可把准确率从 10% 以下提升到 42.4%。