AI 论文与研究

本周最佳两篇 Week 38 Story Imprinting： AI Assistants Absorb Traits from Human Characters They Resemble JEPA-Anything： Learning Predictive Models across Different Worlds Story Imprinting https://arxiv.org/pdf/2609.10883 JEPA-Anything https://arxi…

Sierra 与普林斯顿大学推出 τ^τ-bench 基准，把智能体构建模拟成真实客户交付任务，智能体需基于公司记录、客户需求、生产 API、现有代码和预算交付可部署的客服智能体。

新基准测试发布 https://bw.swerdlow.dev/report

本周顶级 AI 论文（9 月 14 日 - 20 日）： - SoL-Pi - GAUGE - Salesforce Koa - Stellar Colosseum - Byte Model Scaling - Bash vs Typed Tools - Capability Laundering 继续阅读：

DAIR.AI 发布 9 月 14-20 日 Top AI Papers：Meta 研究字节级语言模型随算力增长的缩放规律，用 Marginalize-It 与 End-Of-Token 两种方法从 token 教师蒸馏 1B 字节学生，拟合缩放律显示 End-Of-Token 学生在渐近线上领先蒸馏 token 模型最多 4%，且仅用六分之一训练数据即可追平。

一篇评论文章认为，LLM 正在无视版权与许可证吞食网络内容，打破软件版权与开放之间维持了 40 年的平衡。作者称 AI 让分享知识从对世界的馈赠变成作者的负担：公开代码可能被 AI 更易发现漏洞、被 AI 生成的 pull request 淹没，而现有许可证义务缺乏执法意愿。他警告这可能导致数字黑暗时代，呼吁创作者联合推动新的数字复兴。

整个 arXiv 三百多万论文全部整理为 Huggingface 上的数据集。 ICLR 2027 投稿量超过往年总和？ 这才刚刚开始。 【引用 @secemp9】：很高兴分享将整个 arXiv 网站作为数据集发布在 HuggingFace 上 https://huggingface.co/datasets/secemp9/arxiv-complete 3，148，796 篇论文，包含每个版本，格式为 LaTeX、PDF、PostScript、HTML，总计 16 TB

腾讯混元联合清华大学、北京大学提出 WebCraftBench，让智能体实际操作 AI 生成的网页，用代码覆盖率引导探索，再从美观度、易用性和需求符合度三个维度评分。

作者 madars 声称于 2026 年 9 月 19 日与 Claude 协作完成了 RSA 挑战数 RSA-896 的因数分解，并公布了完整的 270 位十进制结果及两个素因子 p 和 q 的具体数值，附原文链接 https://saweis.net/posts/rsa-896.html。

Before a machine learning model can learn a thermodynamic equation of state, it must discover what its measurements represent: which channels scale with system size, which are intensive conjugates, how sectors pair through contact, and whi…

Data acquisition (DAQ) systems at future particle physics experiments stand to benefit from the extremes of AI/ML development: large-scale foundation models can enhance the performance of feature extraction algorithms, and small-scale on-d…

Bayesian filtering provides a principled framework for online state estimation under uncertainty, yet its application to systems with high-dimensional states and complicated posterior distributions remains challenging. Recent generative mo…

Before a machine can discover a physical law, it must discover what its measurements are: which observations live on cells, which are intensive or extensive, which sectors are dual, and which distinctions are merely gauge. We introduce phy…

We introduce leaky-integrator reconstruction, a training-free method that cures the error accumulation of recursive differenced forecasting. Our first contribution is diagnostic: predicting one-step changes and integrating them by cumulati…

Repository-level software engineering (SWE) comprises heterogeneous task categories, whose progress under pooled agentic reinforcement learning can be uneven: gains in some categories coincide with regressions in others, while aggregate re…

Open-vocabulary audio-visual event localization (OV-AVEL) grounds a text-queried event in time from video, audio, and language. The supervision sources available to this task can differ in temporal-boundary reliability: on OV-AVEBench, our…

Reinforcement learning (RL) offers a natural language for healthcare decisions whose conse- quences unfold over time, yet most reported progress remains far from routine intervention. Ex- isting surveys organize the field by algorithm or c…

Long-context language models interface with external knowledge through raw natural language. In retrieval-augmented systems, this creates a persistent index-payload schism: dense vectors enable searchable routing, but models must re-ingest…

FORM is a domain-specific symbolic manipulation language widely used in particle physics for processing the very large algebraic expressions arising from multi-loop Feynman diagram calculations. Despite its central role in precision theore…

Recurrent models must preserve information that changes future behavior while suppressing hidden-state error. These objectives conflict: contraction improves stability, but contraction along a future-distinguishing direction destroys memor…

Recent advances in large language models (LLMs) have led to the emergence of coding agents capable of performing complex engineering tasks, including register-transfer level (RTL) design and optimization. Existing RTL benchmarks mainly eva…

Engineering predictions require physical mechanisms to be translated consistently into equations, discretization, code, and validation, yet errors can propagate despite local checks. Artificial-intelligence (AI) agents automate scientific …

Pitch estimation on an edge device is constrained in three ways at once. The model must be small, it must stay accurate when the input is noisy, and one frame must be produced inside the frame period. In this report the Frequency Convoluti…

Stochastic reconfiguration (SR) is the standard optimizer for neural quantum states (NQS), but modern NQS often have far more parameters than Monte Carlo samples. We show that in this regime the diagonal shift is more than a numerical stab…

Digital direct-to-consumer (DTC) health campaigns are usually measured after the fact. In-flight forecasting commonly relies on a separate classifier for every cutoff and horizon. We treat this task as a dynamic-system problem and build a …

Single-cell foundation models (scGPT, scBERT, Geneformer) achieve cell-type classification accuracy up to 97.5% in our experiments, yet this aggregate accuracy can mask systematic failure on rare, often disease-relevant cell populations th…

Low-rank adaptation (LoRA) has become a key technology for serving large-scale personalized large language models and diffusion models in the cloud. However, the co-occurrence patterns, resource contention relationships, and evolutionary r…

Social bot detection is essential for protecting online platforms from misinformation amplification, coordinated manipulation, and distorted public discourse. However, large language models have made social bots much harder to detect from …

Graph databases are frequently positioned as categorically necessary for connected-data workloads, yet the systems dimension along which they actually differ - query planning, indexing, and data-readiness cost - is rarely isolated from ven…

MiniMax 披露六项全球合作：Genspark 基于 MiniMax M3 开放权重后训练出 PPT 专用模型 Gen-1 Slides，并成为其 AI Slides 标准模式默认模型，200 个真实生产任务评测中综合与视觉设计得分均列五个参测模型第一，单份生成成本为 Claude Opus 5 的十分之一。

英伟达与 NTU、MIT 团队公开 SoL-Pi 论文，用自动研究循环让脚手架机制自我演化，最终保留动作融合、在线上下文压缩、观察打包和证据保留式委托阅读四个机制。

PackLab 是面向闭环机器人装箱的 MLLM 开发、训练与评估框架，包含物理仿真平台 PackLab-Suite、装箱专用模型 PackLab-VLM 和多难度标准化评测 PackLab-Bench。

Mira-Scene 是一个组合式 3D 场景重建框架，用稠密有界的对应关系恢复替代稀疏位姿回归，核心是像素对齐的 Canonical Coordinate Map（CCM），可在无需场景级布局标注的情况下从物体级 3D 数据训练。

针对仓库级软件工程任务中池化强化学习出现的类别此消彼长问题，研究者提出类别感知的专家训练与策略整合框架，通过同源类别专家交替进行长程 Agentic-miniRL 与 Refresh-Repair-Expand（RRE），并用标签路由多教师同策略蒸馏（MOPD）将专家整合为单一可部署学生模型。

研究者提出 Grounded Action Models（GAMs），一种以 3D 定位构建的机器人基础模型新范式，支持语言、点、框提示并转为以物体为中心的表征，再经多流 Transformer 预测动作块。

首个针对文生视频扩散模型"运动规划"过程的可解释性研究发现，RoPE 导致空间注意力过度衰减，使早期候选区域过早锁定在物理上不合理的位置，抑制相邻帧的合理轨迹并引发生成失败。作者据此提出一种轻量级架构修改，按去噪步数缩放 RoPE 频率，在免训练和基于训练的实验中均提升了生成视频的物理常识。

研究团队提出 OmniEchoBench 空间音频-视觉感知与导航基准，覆盖 197 个真实空间音频-视觉场景、2，972 组问答对和 900 个导航样本，音频为来自 30 个真实环境的一阶 Ambisonics（FOA）数据。

Google 发布 ScientistTwo 论文，展示 AI 研究中的递归自我改进：模型先改进人类方法，再以自身发现为新基线继续改进。它以提出想法、实验验证、剔除无效方案并根据评审反馈开启新一轮的循环方式，自主改进了 107 个人类定义 ML 问题中的 86 个；在基于 ICLR、ICML 和 NeurIPS 论文的问题上，报告了 80.4% 的成功率和相对原始人类基线平均 25.2% 的改进。

LangChain 测试了 Jev 与 LLM 评委在准确率、可重复性、延迟和成本四个维度的表现，以验证 System One 模型能否为智能体评测提供新思路。

MIT 和 Sakana AI 的论文提出 SIFT（Self-Improvement via Fast Tree-search），用 LLM judge 先排序候选自修改、只评测有希望的节点，大幅降低自我改进编码智能体的运行成本。