AI 论文与研究

Artificial Analysis 在旧金山发布 Intelligence Index v5，新增两项基准、三项升级，并首次让指数大部分基于私有测试集或留出答案，同时加大智能体编程与科学的权重。

无脚本通话很难大规模收集。 如今的语音 AI 大多是从 LLM 编写的虚假对话、再由 TTS 朗读出来中学习，所以它从未听过人们互相抢话； 真的很好奇他们是如何让对话持续 10 到 20 分钟的。 他们发布了 Joy，10 万小时的无脚本双人对话，覆盖 10 种语言 https://x.com/shobhitbanga/status/2115862857351061694/video/1

Data attribution methods for Language Models (LMs) are typically evaluated using downstream tasks like fact retrieval. However, existing evaluations vary multiple dataset properties simultaneously, making it impossible to isolate exactly w…

An estimator can be accurate and stable on typical data, yet sensitive to a few possible records. Differentially private mechanisms commonly scale their noise to the worst-case influence of a single record, which lets these records dictate…

The Neural Thickets hypothesis posits that pretrained language models lie in dense neighborhoods of task-specific experts in weight space, which random Gaussian perturbations can uncover. We reproduce the reported gains, but show that they…

Pixel-space generative Transformers typically use one token per image patch, so capturing finer detail requires more tokens and makes global attention increasingly expensive. We observe that finer detail does not necessarily require more g…

Distributionally robust optimization (DRO) hedges data-driven two-stage stochastic programs against error in the distribution estimated from a finite record, and standard ambiguity sets differ in how they treat realizations absent from tha…

Luggage re-identification (ReID) is a critical task in video surveillance for forensics, security or border management (suspicious or abandoned luggage recognition), as well as luggage management (lost luggage recovery). It enables the ass…

Extreme weather is forecast days ahead, yet its consequences for agriculture remain hard to predict: a forecast says where a storm will strike, not which crops will suffer or how much is at risk, and the records that could tell span many s…

World models can achieve strong control performance without faithfully capturing the dynamics of their environment, limiting their potential for mechanistic interpretability and scientific discovery. Here, we introduce world2vec to learn m…

An observer sees only part of the surrounding world, yet must keep track of what lies beyond the current view. Spatial perception therefore involves more than recognizing individual images: it calls for a persistent spatial state that is u…

Safety alignment has enabled large language models to reliably refuse harmful requests directly, yet the same request distributed across a multi-turn conversation still succeeds far too often. This brittleness has a mechanistic root: Align…

UC Berkeley 博士生、Google Research 高级研究员 Alex Krentsel 在 Arena HQ 介绍了自改进 AI 智能体架构 Exo，借鉴操作系统的 exokernel 思路，把智能体拆成受保护的机制层（状态、历史、密钥和 guardian）与可完全演化的策略层。在开放世界 RuneScape 任务上，Exo 比 Codex 和 Pi 便宜 31-62%。

Ask a language model whether sales tend to dip after a product recall, or whether one side effect tends to follow another after a vaccine, and it answers in a fluent sentence that reads like a finding, with no test behind it. Assertions ab…

The marginals of a continuous diffusion model can be reproduced by a deterministic ordinary differential equation (ODE), known as the probability flow ODE, which provides a foundation for accelerated sampling and consistency distillation. …

Modern diffusion models are trained on massive datasets that often include copyrighted or private images. Membership inference attacks (MIAs) address this by testing whether a given image was in a model's training set. The standard approac…

We present TextSeal2, a distortion-free watermark for large language models that keeps its signal under speculative decoding: tokens that replace rejected drafts remain watermarked with the same keys, and the acceptance rate is unchanged. …

Proactive assistants offer short, memory-grounded hints during a conversation, deciding when to assist and what to suggest. Current systems split this between a gate and a larger generator; we evaluate both together on 28,058 decision poin…

An autonomous research laboratory could conduct an entire study, from formulating questions to preparing manuscripts. Artificial intelligence (AI) offers one route towards this goal. Recent AI systems take on several research roles, but ma…

When AI agents share channels and credentials, one agent can press another to break its operator's rule, a first step towards collusion. Output filters check only what an agent has already written. Can an overseer tell beforehand that an a…

Integrated sensing and communications (ISAC) systems using 5G and beyond signals of opportunity enable joint wireless sensing and communication using shared infrastructure and spectrum resources. While multistatic ISAC can increase sensing…

Building and scaling computer-using agents (CUAs) across desktop, mobile, and web environments requires reliable task-completion verification for data curation, evaluation, and reinforcement learning. Typical automated approaches prompt vi…

Sparse mixture-of-experts (MoE) models expand parameter capacity by activating a small subset of experts per token. Expert selection directly affects model capabilities under a fixed active-expert budget. Dense evaluation of all experts re…

Meta 论文发现 RL 后训练让 Agent 首次尝试更准，但在最多 128 次尝试内找到解法的能力反而变差。

Google DeepMind 新论文发现，用强化学习训练 Gemma-4 4B 写道德论证后，模型在选项互换时的答案翻转率从 14.3% 降至 6.6%，但附和用户的比例仅从 45.3% 变为 44.0%，道德寓言准确率反而从 56.5% 降到 53.4%。论文据此建议，评估模型的建议或判断任务时应分别测试一致性与正确性，因为更稳定的模型仍可能出错。

Meta 介绍其 LLM 驱动的 Catching JiTTests 已从研究原型进入生产基础设施，通过 dodgy diff 和 intent-aware 两条工作流自动生成测试，在代码合入前暴露 bug，并用 RubFake 规则与 LLM-as-Judge 集成过滤误报、以 human-on-the-loop 方式降低人工审查负担。

OpenAI 在 Alignment 博客介绍其针对前沿模型（如 GPT-6 Astra）的多层失对齐监控系统，覆盖训练、评估、内部使用和部分外部部署。

Latent Space 采访 Elorian CEO Andrew Dai，介绍其与 Mercor 联合发布的 DecoyCount 基准，测试模型是真正数数还是依赖视觉捷径。

Wiz 基于自身事件响应与公开披露案例，分析 AI 智能体参与真实入侵的现状与趋势，涉及 ARTEX、Hermes 等框架及韩国金融机构（约 67,000 人信息外泄）、泰国财政部、ConnectWise 供应链（尝试攻击超 2,250 个租户）等案例。

Wiz 汇总公开披露与自家应急响应中的智能体攻击案例，包括 ARTEX 针对韩国金融机构与 ConnectWise 供应链、Hermes 攻击泰国财政部，以及 Wiz 调查的三起真实入侵（如暴露的 Flowise 实例、MCP 接口链式 RCE 导致云凭证和 Kubernetes token 泄露）。

蚂蚁 inclusionAI 提出 IDM（Intrinsic Distribution Matching），把 DMD 蒸馏与生成器自身 EMA 特征空间的矩匹配结合，无需外部特征编码器，同一 checkpoint 支持 1-NFE 和 2-NFE 采样。

ChinaTalk 采访葬AI（Funeral AI）作者咸鱼，谈其如何用普通人的视角讽刺和观察中国 AI 行业。他搭建了 MemeBench 网站，从 Bilibili 和抖音评论收集梗图素材，让学生标注并给视频模型排名，结果显示 MiniMax 开源的 H3 在娱乐场景下超过字节跳动 Seedance 2.0 和 2.5。

OpenAI 发布长文，详解其针对前沿模型失准行为的多层监测系统，包括激活探针、基于模型的审查、调查智能体和自动化调查四层架构，并对 GPT‑6 Astra 及以上模型的全部 RL 运行、评估和内部 Codex 流量进行监测。

Automated detection has become the essential safeguard against realistic synthetic media, but deploying it at scale requires training on sensitive data from multiple organizations, putting privacy and the transparency demanded by forensic …

Graph analytics generates irregular, topology-dependent memory-access streams via indirect traversals, limiting data prefetching. Accurate prediction enables timely prefetches that hide memory latency. However, ML predictors couple short h…

AI coding agents now do long, multi-step and multi-session software work with little supervision. When a developer comes back to that work, the useful question is not what the model said but what was actually done, by whom, and whether it …

In many real-world classification tasks, datasets exhibit class imbalance, which can degrade the generalization performance of conventional learning algorithms. Data complexity has recently been recognized as a more objective indicator of …

View- and target-specific echocardiography AI models capture useful but complementary information and may not achieve sufficient predictive performance in isolation. A learned ensemble is a natural way to combine these outputs, but heterog…

We study first-order and best-of-both-worlds regret guarantees in multi-armed bandits. Pogodin and Lattimore (2019) proposed an algorithm that combines minimax-optimal worst-case regret with a first-order guarantee whose loss-dependent ter…

Deep neural networks achieve strong performance across many tasks but remain vulnerable to adversarial perturbations. Adversarial training (AT) is one of the most effective defenses, yet it suffers from high memory cost on large models and…