AI 论文与研究

arXiv 论文介绍 YuE2，通过符号规划统一符号与音频音乐生成，单一 AR-NAR Mixture-of-Transformers 先生成可读乐谱，再扩展为语义音乐 token 并实现全曲音频。

研究者提出 SMAT（Simple MAT），通过将常见模型合并操作归纳为 Scale、Mask、Perturb 三种，联合优化专家损失与模拟合并参数下的期望损失，并引入周期性调度、算子融合和参数存储切换，使每步仅需一次前向和一次反向传播。在四个语言与视觉语言骨干模型上，SMAT 将五种合并方法的平均分较各骨干最强基线提升 1.07-2.16 分，训练时间开销较标准微调不足 2%。

研究者提出 DRM（Diffusion Reward Model），把奖励建模重构为对 p(r|x,y) 的条件密度估计：以冻结的 LLM 编码器为条件，用轻量 Diffusion Transformer 将高斯噪声去噪为奖励向量，从而不预设输出分布并自然刻画人类偏好的多模态结构。

研究提出将 Diffusion Transformer 的残差连接从被动求和改为主动检索机制，通过结构化连接设计让每个 Transformer block 选择性关注早期层表示。该方法训练迭代次数最多减少 1.73 倍，额外参数不到 0.1%，将 REPA-XL/2 的无引导 FID 从 5.9 降至 4.34，使用 classifier-free guidance 时达到 1.39 FID。

研究发现，大语言模型在线策略蒸馏（OPD）或许并不需要 KL 散度：只要让更新方向朝向教师模型即可，甚至仅对教师与学生分歧强烈的一小部分 token 保持正确更新方向，训练依然有效。基于此提出 Consensus Multi-Teacher On-Policy Distillation（C-MOPD），让每个样本受所有教师监督，在数学和代码基准上持续优于 MOPD。

一项研究系统消融了 EEG 基础模型的时空掩码策略，在 MAE 和 JEPA 两种自监督框架下预训练 58 个模型，并在 OpenEEGBench 的 12 个数据集上用线性探针评测。

SpatialSpeak 是一个两阶段框架，将 QA 原生重建预训练与空间 CoT 学习结合，在 ReVSI 上把 CoT-VC 带来的增益从 2.6 分提升到 6.9 分。该框架在 ReVSI、VSI-Bench 和 SPAR-Bench 上取得 SOTA，ReVSI 得分 62.8，超过最强对比基线 8.7 分。

研究在注意力机制中用 K-interval attention 量化 softmax 对 Transformer 预训练的影响，该方法用 K+1 个网格值近似指数函数。

研究者构建了一套无需人工标注的合成流水线，通过扰动改写公开文档、生成需依赖文档推理的问题与评分标准，从 3.5k 篇文档生成约 10k 样本，用于训练学生模型。

StoryEngine 是一个基于状态的智能体框架，通过分离权威语义规划与不可靠视觉观测来生成连贯的长篇视频故事。它维护实体位置与故事相关状态的结构化表示，传播事件引发的状态变化以定义每个镜头的起止状态，并构建可复用实体的规范参考，将状态与视觉约束编译为可执行的渲染计划。实验显示 StoryEngine 在叙事质量、叙事连贯性和视觉一致性等所有评估维度上均持续优于现有最先进方法。

DeepSeek-AI 发布技术报告，介绍用于大规模 LLM 智能体 RL 训练与评估的生产级沙箱平台 DeepSeek Elastic Compute（DSec），统一提供 FnCall、容器、microVM 和完整 VM 四种后端。

ProgramDistill 从交互式 Web 应用，到可验证、参考引导的 SWE 任务 论文：https://huggingface.co/papers/2609.18805

原来你不需要IMU 🤯 [引用 @aditya_bhatt]：在我最新的博士论文中，我们向"堆传感器"宣战。 首次实现抗推挤的人形机器人行走，仅靠关节编码器！无需IMU，无需力/力矩传感器。 🥁 隆重介绍 Blind Dexterity 🧵 👇 与 @OKaidanov @liu_puze @Jan_R_Peters 合作，来自 @DFKI @ias_tudarmstadt

Learning transferable representations for dynamic text-attributed graphs (DyTAGs) requires models to capture underlying interaction dynamics that persist across domains. However, existing methods tend to overfit to domain-specific structur…

We propose sliced Orlicz-Wasserstein (SOW) distance which is a generalization of sliced Wasserstein (SW) distance. SOW replaces the $L^p$ norm in SW with a Luxemburg norm cost induced by an Orlicz function $φ$. First, we prove that SOW dis…

Multimodal contrastive learning is a dominant paradigm for learning transferable representations from unlabeled data, but standard objectives primarily capture information that is redundant between modalities. Partial Information Decomposi…

Accurate assessment of Schistosoma japonicum-associated liver fibrosis is essential for disease management and long-term follow-up in endemic regions. Ultrasound provides non-invasive imaging, but complex local echogenic patterns and anato…

Recommender systems rely heavily on heterogeneous behavioral feedback to infer user preference. Although abundant, these signals are imperfect measurements: the same observed behavior can arise from different underlying states, such as gen…

Predicting smooth dynamics and collisions requires modeling continuous evolution and abrupt state changes. We introduce HamiFormer, a dual-expert diffusion field combining whole-window denoising with residual-corrected Hamiltonian propagat…

Ultrasound (US) acquisition depends on the operator's ability to interpret anatomy and anticipate how the view will change with probe motion. Many robotic US navigation methods select actions without explicitly predicting these anatomical …

Radiomap blind prediction aims to infer radiomaps from observable representations of the propagation environment and base station configuration without field measurements. In practice, the observable representations are inherently incomple…

As AI agents are becoming widely adopted in the financial services industry, careful measurement is essential to understand where they can be reliably deployed and where oversight and professional review remain necessary. Such measurement,…

A trained neural network can be represented by a parameter vector in high dimensions. Learning distributions over these vectors enables the generation of new models across various tasks and architectures. A central challenge is permutation…

Over-the-air computation has emerged as a scalable and efficient solution for deploying federated learning algorithms in wireless networks by exploiting waveform superposition for simultaneous model aggregation. Most existing work struggle…

Unified multimodal models combine understanding, generation, and editing within a single network, offering a promising foundation for versatile multimodal applications. However, growing multimodal contexts make KV cache storage and access …

Flow critics learn return distributions by transporting Gaussian noise to Bellman endpoints via continuous velocity fields. While velocity bootstrapping stabilizes training by querying a successor teacher, existing methods face a structura…

Several works have explored multi-agent reinforcement learning (MARL) in LLM collaboration. However, constructing reliable rewards is difficult in practice, as complete and accurate metrics are often unavailable and hard to aggregate. Pref…

A four-stage LLM pipeline gives up as much as 40.5 accuracy points at its own interfaces (gemma-3-12B on MATH-500, Holm-corrected p = 1.66e-19; the largest tax in the primary family). We hold model, problem, stages, stage prompts and compl…

Model merging efficiently combines specialized large language models (LLMs) without joint retraining, but can substantially alter expert routing in Mixture-of-Experts (MoE) models. Such \emph{routing drift} is often interpreted as routing …

This paper studies learning of shuffle ideals with membership queries as well as with contrastive queries---a form of membership query that reveals not only whether a selected word $w$ is in the target language or not, but also provides a …

Raising the update-to-data (UTD) ratio breaks off-policy critics in two ways grouped as "plasticity loss": collapsing representations and growing value magnitude $|Q|$. We separate them in Soft Actor-Critic (SAC) with scaled critics (width…

Fine-tuning a language model on new data can make it forget facts that it should keep. We ask whether measurements taken at the start of a fine-tuning run can bound, for each protected fact, the probability that the run makes the model for…

Clinical Large Language Models (LLMs) achieve strong medical-exam accuracy; however, a correct answer does not guarantee that the explanation names the concepts that actually drove the decision. We introduce three lightweight, directly int…

一项覆盖3132名参与者、共五个实验的研究发现，仅当AI建议可用时，人们放弃回答的比例从36-44%骤降至3-6%。由于所用模型Step 3.5 Flash对所测试的电影视觉细节问题几乎全错，参与者正确率反而从27.5%降至9.2%；财务激励能减少求助AI并略提升弃答比例，但未能扭转效应，即使AI答案自动展示时弃答率也降至1-7%。

开源 RL 环境就是赢。当然是在 Hugging Face 上！ 引用 @HarveenChadha：本来想安静过个周末，结果小米昨晚放出了他们的 RL 环境仓库。对比一下，这类任务如果去买，通常每个要几百到上千美元，所以这个仓库简直价值数百万。https://huggingface.co/datasets/XiaomiMiMo/MiMo-V2.6-RL-oss

小米官方复盘 MiMo-V2.6 发布后的工具调用重复问题，响应级重复率超 0.05%，并区分了正常并行调用、调用泛滥与调用重复。回放 RL 各 checkpoint 显示泛滥率随训练从 11.1% 升至 24.6%，32 次调用阈值惩罚过于宽松；直接调低阈值需重启 20 步 MixRL，估计成本 231 万美元，且内部测试重复率仅从 13.45% 降至 3.83%。

Anthropic 宣布 Claude 在 Claude Science 系统中仅凭一条提示词、无人监督连续运行数天，算出平面 N=4 超杨-米尔斯理论六粒子振幅的九圈结果，超越 Lance Dixon 团队 2023 年的八圈纪录，总成本几千美元，其中直接自举路线的 Python 运行成本仅约 100 美元。

斯坦福大学 Safavi-Naeini 团队首次实时观测到单个声子消失的量子跃迁过程，论文于 9 月 17 日发表于《科学》。团队将寿命约 2 毫秒的微型机械谐振器与超导量子比特连接，通过数百次连续测量捕捉到从 1 到 0 的跃迁瞬间。

Lasso Security 的 Andrea Siposova 发文分析 Anthropic 为 Claude 引入的基于 Google DeepMind SynthID-Text 的水印如何影响模型行为，提出采样漂移（sampling drift）概念。

Salesforce AI Research 提出 Just-in-Time Memory，主张存储原始轨迹、在下一个任务到来时再决定提取什么，而非每次运行结束就总结。