AI 论文与研究

智能体战争对所有国家都将是大事。AI将有能力执行震网级别的网络行动，下一场军备竞赛将在自我改进的智能体之间展开。然而目前我们几乎没有针对这一点的评测。我认为没有任何国家为即将到来的规模和速度做好了准备

Artificial Analysis 推出 AA-PersonalityBench，从 30 项人格特质和 21 项沟通特征两个维度衡量语言模型差异，基于 900 道精选问题、由 LLM 评审团两两对比打分。

MIT、波士顿儿童医院与 Inria 的研究者开发并训练了深度学习模型 E(3)-Pose，可在胎儿头部运动时辅助成像，生成供放射科医生查看的高分辨率三维图像。该模型把胎儿头部的左右对称性等物理原理嵌入网络架构，而非仅从数据中学习物理规律，从而更准确地识别大脑朝向。相关论文已在 2026 年欧洲计算机视觉会议（ECCV）上展示，目前模型主要基于第三孕期胎儿训练。

NVIDIA 在德国 Jülich 超算中心的 JUPITER 上完成 104 量子比特容错量子计算模拟，这是迄今规模最大的容错量子处理器完整模拟。该模拟验证了 magic state cultivation 方案，所需物理量子比特比此前方法少 90%，并首次在真实硬件噪声下评估其表现。模拟基于 GH200 超芯片与 cuQuantum、CUDA-Q 运行，可让研究者在造硬件前预测设计性能。

Modern text-to-image (T2I) systems use internet-scale language pretraining for text-conditioned image generation. We ask whether language pretraining is necessary for strong T2I generation. We introduce Halcyon, a family of early-fusion mu…

Multimodal large language models (MLLMs) have made rapid progress on visual reasoning, yet many tasks require evidence that is difficult to obtain through internal reasoning alone: our controlled study on 22 benchmarks shows that simply in…

While LLM agents can generate visually plausible CAD assemblies, these designs often fail to fulfill their intended physical functions. We study functional CAD design, a setting where agents interpret underspecified requests to construct a…

Multimodal reinforcement learning with verifiable rewards (RLVR) provides reliable task-level supervision, yet offers limited guidance for localizing errors in visual reasoning. On-policy self-distillation (OPSD) complements this supervisi…

Reinforcement learning enables mobile robots to make fast and scalable navigation decisions in constrained human-shared environments. However, directly extending single-robot approaches to multi-robot settings often fails to ensure safety …

In this paper, we systematically investigate the mechanisms, scaling behavior, and generalization of physical Recursive self-improvement. We introduce PhysRSI-eval to quantify physical RSI through a five-axis metric: First-Attempt Performa…

Long-horizon agents must learn heterogeneous capabilities from interleaved trajectories. Updating shared parameters to learn new capabilities can degrade previously acquired ones, resulting in catastrophic forgetting. As a trajectory unfol…

A seemingly reasonable push in Sokoban can make the puzzle unsolvable. To avoid such mistakes, an agent has to anticipate the consequences of its actions before execution, especially when feedback from the real environment is sparse or act…

The goal of a world model is to capture the dynamics of an environment. This has the great appeal that it allows agents to plan and simulate a course of actions with the help of a world model. For these uses to be reliable, however, the le…

Systems that combine several large language models (LLMs) often use one context for many questions. When the context is passed between models as text, the receiving model processes all of its tokens again for every question. Avoiding this …

Bilevel optimization provides a framework for machine learning tasks such as hyperparameter optimization, data reweighting, and meta-learning. Existing Newton-based bilevel methods typically rely on the classical hypergradient formula, whi…

LLM agents that compute the right answer can still answer wrongly. Agent harnesses increasingly keep long tool outputs out of the context window, saving them to files that the agent searches or reads on demand. We ask where this costs accu…

We propose KronLSTM, a linear attention architecture based on the Kronecker Delta Rule, a novel update rule for associative matrix memories. KronLSTM extends the xLSTM family with a memory-revision mechanism that corrects targeted associat…

Machine learning is widely regarded as a promising application in which quantum computing could offer a computational advantage. However, general quantum machine learning models with provable guarantees on performance, trainability, and ef…

Many modern language models have adopted hybrid architectures that incorporate both linear-time recurrence and softmax attention. Modern recurrence layers provide a highly expressive mixing mechanism while ensuring high throughput by compr…

Frontier agents retrieve structural metadata well and compute geometry from coordinates less reliably. We ask how much of a benchmark score reflects that gap and how much reflects the evaluation configuration around the model. We study Pro…

Agent learning is shifting from learning from human data toward learning through interaction with the world. However, scaling such learning directly in real-world computer systems is challenging and can introduce substantial security risks…

Context compression improves long-context efficiency but may introduce faithfulness issues through an information bottleneck. Existing evaluations rely on QA metrics, but a reader can still answer correctly from an unfaithful compression, …

Beyond basic tool use and coding skills, effectively incorporating visual references into agentic generation is equally essential for applying AI agents to real-world productivity tasks. We study this crucial capability through webpage dev…

Memory-native evolution requires making memory an intrinsic capability of foundation models, enabling them to continuously accumulate and utilize information. To this end, we introduce MemT-1.0, a new architecture of memory foundation mode…

Salesforce 论文研究跨委托方的智能体间（A2A）协作失败原因，基于 Claude Code、Codex 和 Cursor 在独立沙箱中跑了 2500 多场会话、488 个场景，归纳出 20 种失败模式。

面向打包多模态掩码的精确序列并行。 它按层选择 ring attention、稀疏 all-to-all 或 Ulysses，数值与单 GPU 一致，并在最多 128 块 H100 上于 48 个案例中的 47 个优于 FlashMask + Ulysses 混合方案。 https://lumalabs.ai/news/packed-multimodal-attention

Meta 工程博客介绍 GenRecs Tokenization，一种用稳定语义 token 词表替代逐条目 ID 的广告推荐数据表示，配合 GEM 与 Adaptive Ranking Model 带来 +0.35% 转化提升。

PromptArmor 发布对 Harvey for Outlook 的威胁分析，指出这一在 Outlook 内运行 Harvey 法律 AI 的插件因读取不可信邮件而面临间接提示词注入风险。

Cohere Labs 提出 CALIBER 方法，训练推理模型在开始推理前估计解题成功率、推理后再评估该答案正确的可能性。在 BigMathDigits 上，7B 模型的 ECE 从最强单置信度基线的 0.059 降至 0.028，降幅 52.5%，AUROC 达 0.806、Brier 分数 0.130。该 7B 模型在未经额外训练的情况下，改进还扩展到数学之外的问答基准。

Anthropic 让内部研究模型在 Claude Science 中主要自主运行 64 小时，设计针对五种蝎子神经毒素的蛋白质迷你结合剂，共生成 318,484 个设计、筛选出 180 个。

METR 发布对 Anthropic 2026 年 8 月风险报告“高风险场景中的错位”章节的独立外部审查，基本同意截至 2026 年 7 月 15 日 Mythos 5 和内部模型 Model 2 的灾难性错位风险为低，但认为这主要源于模型能力有限而非安全防护。

谷歌研究人员发布论文，使用四个 Gemini 4 Argon 智能体在 27 小时内完全自主通关《我的世界》并击败末影龙，团队称这是 AI 智能体首次在无人类辅助下从头到尾通关。

Sakana AI 开发出推荐方法 QD-Rec，把「候选＋选择理由假设」作为推荐单元，由生成 AI 基于历史优质假设反复生成新假设，并用 Quality-Diversity 思路保留不同切口的推荐。在 M&A、客户定向、论文推荐三类任务中，QD-Rec 多数设置下平均评分高于每次从零写假设的基线，并构建出规模更大的推荐集合。

Modal 工程团队分享 NVIDIA Vera CPU（Arm 兼容）在 Modal Sandboxes 的基准测试，作者自建 Agentless Agent Bench（50 个 Terminal-Bench oracle 解法与验证器、剥离网络调用、并发 Docker 容器运行）对比多款 CPU。

面壁智能 OpenBMB 提出 Trident，用多正样本视图 InfoNCE 联合训练文本、图像、融合三种视图，解决多模态检索器在图文混合语料上性能骤降的“Chaos in the Text”现象。

商汤发布迄今覆盖面最广的 GPT-6-Astra 具身智能实证研究，横跨 12 个仿真基准和 8 个真机平台，并同步开源 SenseNova-RoboRSI 框架和一个以约 1/50 推理成本运行的 27B 模型。

写好STATA代码对我的学术生涯非常重要。看来这项技能正基本变得过时。（也提醒那些仍把AI当作聊天机器人而非智能体的学者们，一年来变化很大）https://cepr.org/voxeu/columns/letting-ai-agents-write-econometric-code

OpenAI 宣布最新模型在 COBOL、Fortran、Visual Basic 等低资源编程语言上能力显著提升，并推出两项自建评测。

Google Research 团队发布 Harness²，一个即插即用的递归 Agent harnessing 框架，围绕冻结模型改进执行 harness。流程为基线记录、Editor 起草候选编辑、Composer 合并有效编辑、Reflector 将经验存入库供后续任务调用。

自变量推出 X2Real，一个基于 NVIDIA Isaac Lab-Arena 的通用机器人策略 sim-to-real 基准，覆盖 10 个能力维度、44 个分层长程任务，从视觉与语言推理到精细操作、双臂协作和工具使用。