AI 论文与研究

Anthropic Frontier Red Team 研究警告，小型 AI 模型可能距在智能烤箱、杂牌 Android TV 盒子等商用 IoT 设备上自我复制只差数月。

vLLM 让训练前向直接调用其推理 kernel 并改造这些高性能 kernel，使 rollout 与训练对每个采样 token 产生逐位相同的 logprob（diff = 0）。

Apple 团队发布首个从零预训练的三模态掩码扩散模型（MDM）研究，在文本、图文和音频文本数据上训练。论文系统分析了多模态缩放规律、模态混合比例、噪声调度和 batch size 影响，提出基于 SDE 的重参数化方法免去调优最优 batch size，并在 3B 模型规模（6.4T tokens）上验证了文本生成、T2I 和 T2S 任务的表现。

Apple 研究团队提出一种轻量方法，直接从语音中检测语码转换（CS），无需依赖多语种转录。该方法利用预训练语音编码器的帧级语种信息，通过单语录音找出语种选择性隐藏状态维度，再聚合其帧级激活值生成 CS 预测。在 CS-FLEURS 的 14 个语种对上，Whisper-small 编码器取得 0.798 的平均 F1，高于 CNN 的 0.776 和 HuBERT 线性基线的 0.751。

一项新基准 Design Creativity Bench 用 80 个开放式 UI 设计任务评测 16 个前沿模型，从原创性、创意跨度和合理性三个维度衡量设计创造力。表现最好的 OpenAI GPT-6 Astra 得分 65.5，人类设计师参考分为 92.3，差距明显。该基准的排行榜、方法论与可复现评测框架已公开。

Meta 介绍其人类主导的 AI 科研系统探索，使用 Muse Spark 1.3 与 Muse Wonder 构建编码脚手架，访问五百万份去标识电子健康记录，医学研究者四周内完成十篇覆盖传染病、肿瘤、内分泌等领域的研究手稿。每篇论文经过医学同行反馈、生物统计学家审查和第三方独立专家评审；作者称最快七天内从首个提示词完成手稿，最佳情况下效率提升至传统流程的十倍。

indigo 测试 Gemini 4 Argon 的写作和知识组合能力，结论是其让 Google 重新追上第一梯队但未超越对手，应用 Antigravity（Harness）落后半年到一年。

Johann Rehberger 披露 Visual Studio Copilot 与 SQL Server Management Studio 中的远程代码执行漏洞 CVE-2026-103572，根因是利用文件名末尾多加一个点绕过路径校验，写入 .mcp. 或 settings.VisualStudio. 等受保护配置文件，实现任意命令执行。

📈 合成复杂数据 @venturetwins 在 @Antigravity 中用 Argon 将大型复杂数据集中的洞察合成为一份交互式演示文稿 https://x.com/venturetwins/status/2110039192365637985/video/1

MIT McGovern 研究所与 ICoN 中心在 PNAS 发文提出，判断人工神经网络是否具备生物合理性，关键看它能否在没有开发者干预下自组织出学习机制。他们用随机互连的神经元样单元和简单连接规则，构建出能识别模式、分类图像的自组织网络，性能与依赖反向传播的网络相当。研究认为，若大脑确实使用类似的突触基元完成反向传播的功能，将为机器与大脑在基本机制层面建立强关联。

论文提出 SanSi，通过让 Ouro-1.4B 这个预训练可复用层的小模型重复运行同一批层最多 8 次，并在每次循环后读取选项概率，使小模型在决策任务上大幅提准。

Google DeepMind 发布 Gemini 4 Argon 并开始广泛推出，同时扩大 Fairwind Program 给受信任的网络防御者。模型解决了全部 12 个端到端模拟网络攻击场景（上一代为 5 个），Wiz 用其发现此前模型遗漏的医院软件关键漏洞，涉及数十万患者记录。

希望看到更多公开的端到端智能体基准测试

我们一直在探索利用 Gemini 4 Argon 加速 @Google 研究的潜力。其中一项测试是：用人类完成趣味任务的视频（比如来一局沙包游戏？）来自主生成训练模拟，而不是完全依赖人工编程。 来听听我们研究团队的介绍，并观看实际效果——我们的类人机器人研究这些视频、运行模拟，最终掌握了沙包投掷游戏。

The Verge 评测 Ray-Ban Meta Audio 音频智能眼镜，售价 $349，仅保留开放式音频功能并移除摄像头，佩戴者不再有被旁观者抵触的压力。

Agentic robotics places a large language model (LLM) agent in the robot control loop, where it composes perception, motion, and vision-language-action (VLA) tools under the guidance of a memory and thereby generalizes to new tasks without …

Sparse-to-dense motion completion recovers dense mesh motion from sparse vertex displacements. Residual priors trained on one geometry family frequently fail to beat parameter-free harmonic extension off-family. Moreover, cross-family trai…

How should demonstration quality be used during vision-language-action (VLA) pretraining to improve performance after target-task fine-tuning? We compare weighting, filtering, and conditioning across reward definitions and label granularit…

Long-horizon agents have drawn growing attention for their ability to work autonomously -- from hours to days -- on economically valuable tasks, a capability essential to automated discovery and recursive self-improvement (RSI). Yet open-s…

Sampling equilibrium field configurations remains a major computational bottleneck in lattice field theory, motivating the development of generative samplers that can produce independent configurations. Diffusion models are promising for t…

Reasoning models achieve strong performance through long chains of thought (CoT), but often spend more tokens than needed to solve a problem. Existing ways to shorten the CoT require additional training for each model, add a fixed steering…

How can we build an agent that thinks beyond human language, yet still acts through it in the external environment? Modern LLM agents still reason and interact through text, while latent reasoning offers an efficient internal medium but la…

Generating information-rich visual artifacts such as slides, posters, and dashboards as code is appealing because the outputs are structured, editable, and verifiable. Yet code models cannot see how their choices of layout and color come t…

Persistent homology summarizes the topological evolution of a graph as persistence diagrams (PDs). Generating graphs whose topology matches a target descriptor is increasingly important, yet existing topology-aware generative models rely o…

Reinforcement learning with verifiable rewards (RLVR) relies on costly model-generated responses, motivating larger rollout batches and square-root learning-rate scaling. Reported to approximately preserve progress per sample with one opti…

Multi-objective molecular optimization is a critical process in drug discovery that aims to modify a given molecule to simultaneously improve multiple properties while maintaining a similar overall structure. This process often requires it…

Earth observation (EO) models are still built by hand. For each new task, experts design a training recipe and tune it against a benchmark. We show that this loop can be handed to an agent. Any verifiable task can be cast as a harness and …

Robot policies are expected to generalize across deployment environments, including variations in the robot itself: cameras are bumped and remounted, calibrations drift, controllers run at different rates, and end-effector pads wear down. …

Robots are increasingly autonomous in task execution, yet their improvement remains largely human-driven. This study investigates Physical Recursive Self-Improvement (Physical RSI), in which physical experience guides lasting updates to bo…

This paper investigates offline Q-learning for finite-horizon Markov decision processes with Kullback-Leibler (KL) regularization. Recent studies have shown that KL regularization can yield fast rates in offline decision-making problems. H…

A central challenge in scaling flow models is maintaining stability in training and integration. In flow matching, this stability depends critically on how source samples are paired with data. Independent random pairing introduces substant…

It is difficult to accurately model the physics of the interaction between a robot and the objects it manipulates, including contact mechanics, coupled contact dynamics, deformations, stiffness and friction. We present RoboSimStudio (RSS),…

With the rapid development of multimodal large language models (MLLMs), evaluating their video understanding capabilities has become increasingly important. Existing video understanding benchmarks cover only part of the evaluation dimensio…

Simulation-based Bayesian inference for Chemical Reaction Networks (CRNs) must simultaneously solve two computationally demanding problems: discriminating between overlapping reaction pathways and estimating reaction rate constants. Classi…

Large language models (LLMs) generally rely on sampling temperature to regulate the amount of randomness in their outputs. Mainstream methods for selecting temperature typically involve fixed, heuristic global settings and context or token…

Ethan Mollick 称在 Gemini 4 发布前花了较长时间试用，认为它比之前的 Gemini 模型有明显跃升，是首个通过 Fable/Mythos/Astra 门槛的非 OpenAI/Anthropic 模型。

Artificial Analysis 发布 Decision Index v1.0，从知识、推理、工具调用和分类四个基准评测决策模型的能力、成本和速度。

哪一种"氛围评测"更糟？

卡内基梅隆与微软论文提出 CRBugBench，包含 1,800 个代码审查任务，每个 PR 引入一个人工确认并修复的 bug。最佳模型 claude-opus-5 在前 5 条评论中检出 71.9% 的 bug，高于 5 个月前 opus-4.6 的 47.7%，但在某公司私有仓库上下降最多 25.5%。作者建议审查 Agent 与人类审查者配合使用，并为其提供仓库自有知识。

Google 的 Logan Kilpatrick 发推称已有许多 Argon 的实际使用示例，来自 Google 内部、客户和合作伙伴，并附上博客链接 https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/。链接指向 Gemini 4 Argon 的介绍页面。