AI 论文与研究

Artificial Analysis 将 AMD MI355X 纳入 AA-AgentPerf 基准，覆盖 Kimi K3、GLM-5.2/5.3 和 DeepSeek V4 Pro，并与 NVIDIA B200/B300 对比编码智能体推理性能。

研究人员利用 AI 蛋白质设计工具，保留两类蛋白酶的催化核心并替换其余整个蛋白，设计出可调谐、能定点切断特定蛋白质的酶。研究人员用这些酶切断了与神经疾病相关的蛋白质，以展示其有效性。

斯坦福与Together AI的论文提出HybridAgent，让大部分推理在隐藏向量中完成、仅在工具调用和关键决策时输出文本，解码算力降低62.5%至72.9%。在GAIA上，14B版本得分52.4，高于同模型纯文本智能体配合ARPO的43.7，运行速度也快2.1倍。其静默推理步骤仍可解码为可读文字，任何LLM都能改造成HybridAgent。

The Verge 资深评测员 Victoria Song 亲测 Twin Health 数字孪生代谢管理项目 100 天。该项目需持续输入体成分、血糖、血压等数据，用红橙绿颜色预测食物对血糖的影响；随访化验显示代谢健康评分提高 9 分，甘油三酯下降 70mg/dL，多数肝酶恢复正常。作者也指出持续记录带来追踪疲劳，并在癌症治疗等高风险场景下存在数据隐私与信任挑战。

推荐小红书的工作！ Vela: User Behavior as Supervision Sparks LLM Intelligence

Microsoft 研究团队在 Nature 发表 Probabilistic Bias Correction（PBC）方法，用机器学习修正现有气象模型的系统偏差，可仅在 CPU 上运行并已开源。

Vals AI 让六个个人助理智能体在 28 天退款窗口已关闭后，为用户追讨 £350 火车票退款，所有外发消息由模拟环境拦截应答。

Prime Intellect 与 NVIDIA 构建 2,500 个经验证的企业任务，用 GRPO 在 128 块 GB200 GPU 上训练约 11 小时，将 Nemotron 3.5 Super VL 在 AutomationBench-AA 公开分割上的得分从 34.0% 提升到 57.7%，相对提升 70%。

Goodfire 研究员 Nick Cammarata 发表 65 分钟阅读的研究长文，主张早期模型中多义性的成因不是叠加态，而是递归泛化等若干可修复的普通问题。

LlamaIndex 发布 ExtractBench Pro，这是基于真实长尾文档构建的 ExtractBench 扩展评测集，包含 100 份合成文档、六大类新挑战。该数据集由确定性流水线从种子和配置文件生成，覆盖图表取值、跨页规则、内部术语、图示关系等场景，刚性流水线表现吃力，而配备合适工具的智能体系统更能适应。

两周以后的天气预报并不可靠，这给农民和其他人留下了关键空白。概率偏差校正（Probabilistic Bias Correction）改进了延伸期预报，在全球竞赛中拔得头筹，并展现出实际应用的潜力：https://msft.it/6013aseY9 图片来源：科罗拉多州立大学大气研究合作研究所与美国国家海洋和大气管理局（CSU/CIRA & NOAA）

Anthropic 发布博客警告，小型 AI 模型可能距离自我复制能力只有数月。其研究人员将 Google 的 26B 参数 Gemma 4 拆分到 15 台 Android 电视盒上运行，并在模拟 Android 网络中测试 AI 蠕虫，其中只有智谱（Z.ai）的 GLM-5.2 成功传遍整个网络。

Sentient Foundation 发布开源基准 Fingerprinting Leaderboard，用于评测模型指纹方法，证明 AI 模型是否被抄袭，配套测试代码 mlprints 同步开源。

MIT研究人员团队在Nature发表论文，提出一个机器学习后处理层，通过从历史数据中学习现有预报模型的错误并修正其输出，将两周到六周范围的预报准确率翻倍。该模型在ECMWF的季节间预报竞赛中获第一名，首次在该时间范围内稳定超过气候学平均值，并显著改善洪水、寒潮等极端天气事件的预测；模型开源、低成本、可附加于任意预报模型，肯尼亚气象部门已接洽用于改进季风季预报。

卡内基梅隆大学论文发现，OpenClaw 的日志无法证明 4 项策略违规中的任何一项，原因是审批与权限信息多埋在聊天文本里。在 3 个 benchmark 的 696 次违规运行中，验证器仅能证明 42.2%；若改用结构化审批记录，作者估计可达 83.0%，每会话区块链费用不到 1¢。研究还发现仅 10% 到 30% 的违规运行需要人工判断，多数违规可用代码检查。

推荐阅读！ AutoEnvScaling：用终端智能体自动化数据飞轮

Anthropic 警告，小型 AI 模型明年或能在各类 IoT 设备上自主自我复制，进而催生无需人类交互即可传播、可自行识别陌生设备的计算机蠕虫，且因无中心节点而难以关停。该公司近期实验显示，实现这种 AI 恶意软件复制器仍面临重大挑战。

安全公司 White Circle 提出 RISE 攻击方法，用现成工具和 API 调用绕过 DALL·E 3、Google Nano Banana 2 和 GPT-Image-2 的安全过滤，约 13% 的尝试生成 explicit 内容。

小红书 AllSpark 团队联合商业算法提出 Vela 框架，将匿名化真实用户行为历史转化为训练任务，基于约 10 万匿名化样本主体构建约 7.5 万条训练样本。

ServiceNow 和 McGill 的论文推出 AgentHorizon 基准，包含来自人工屏幕录制的 1,373 个判官任务，通过交换近同任务的指令构造干净的失败样本。

Mathematical research drives scientific discovery, and recent advances in AI for mathematics have shown that language-model agents can solve challenging problems and develop substantial mathematical proofs. Yet much of today's AI-for-mathe…

We study the optimisation landscape of the maximum mean discrepancy (MMD) over weighted atomic measures. For the one-dimensional Gaussian kernel, we show that, for any compactly supported target distribution, sufficiently many particles en…

In sectors such as energy, aquaculture and chemical processing, the effectiveness of both mechanistic and data-driven models is frequently limited by poor-quality training data, which obscures underlying physical laws and degrades predicti…

Typed decision models (TDMs) are increasingly used in applications such as request routing and model selection, making reliable decisions critical to system performance. Fallback options accommodate categories outside regular choices, enab…

As Large Language Models (LLMs) continue to scale, their energy consumption during inference has become a critical concern. While Autoregressive (AR) models are the industry standard, Diffusion Language Models (DLLMs) have emerged as a pro…

We propose DeltaStreamer, a novel agentic framework for streaming video understanding inspired by how humans observe, remember, and investigate. Streaming video understanding requires a system to retain evidence before knowing which detail…

We present OpenTWBench, an open evaluation suite for Taiwan-domain knowledge in large language models, built around two benchmarks. tw-exam-bench contains 255,854 four-option questions across 148 academic subjects, from fifteen years of Ta…

Memory-efficient subspace optimization has emerged as a practical approach to reducing the memory cost of LLM training while preserving optimization performance. Previous algorithms typically rely on the leading singular subspace of stocha…

Long-horizon tasks expose a weakness of monolithic tool-use loops: one agent must decompose, execute, and verify a task within a single trajectory. We introduce LAGO, which treats specialized language agents as reusable temporally extended…

Enterprise group chats record project progress, decisions, task ownership, and other organizational information valuable to enterprise agents. Reliable retrieval is challenging because evidence spans interleaved threads, channels, and time…

Frontier web agents using closed models have high inference and computational costs. Smaller open models can be an alternative, but to be performant, they require training supervision from a human or teacher model. Token-level teacher dist…

Reward model overoptimization in RLHF is usually described as reward hacking. We recast it as rent-seeking: the policy, a firm, earns reward from a regulator, the proxy reward model, that true quality does not back. The lens turns overopti…

Although multivariate time series forecasting has broad application prospects, effectively incorporating exogenous variables remains challenging. Prevailing discrete-time architectures often impose rigid temporal alignment across heterogen…

LLMs absorb vast knowledge from training data, yet their inherently probabilistic nature offers no guarantee that this knowledge will surface in their outputs. To bridge this gap, Knowledge Extraction Enhancement (KEE) techniques such as a…

LLM agents are evolving from a single model calling tools to planners that orchestrate multiple expert models whose capabilities live in distinct weights. This shift breaks a hidden assumption of today's serving systems: that the working s…

Distilling capable language-model agents into smaller models requires aligning the student's behavior with the teacher's. This raises a fundamental question: at which states and at what granularity should this alignment be performed? Offli…

Agentic AI systems combine model decisions, mutable state, tool calls, and delegation to external agents. Monitoring them requires observations from different execution boundaries, while prevention requires control before an operation prod…

Human values shape how we expect language models to behave. Dominant alignment approaches treat values as global objectives, although the appropriate value orientation is personal and contextual. Training-free alignment can steer models to…

Transformer-based tabular foundation models (TFMs) achieve state-of-the-art performance on small to medium-sized tabular prediction tasks while processing diverse inputs through a shared stack of layers. Recent layerwise analyses identify …

Spatio-temporal forecasting is fundamental to many real-world systems, yet deployment environments are inherently varying: temporal distributions may evolve, spatial domains may change, and observation availability may shift from training …