AI 论文与研究

A long-standing goal of AI is a model that can continually learn and improve itself. On post-trained models, supervised finetuning (SFT) on new data often causes poor generalization and catastrophic forgetting. As such, the conventional wi…

This paper studies online quantile regression for large-scale and streaming data using Stochastic SubGradient Descent (SSGD) with constant learning rates. Classical offline inference for quantile regression is computationally and memory in…

Feature pyramid methods, from FPN to BiFPN, have achieved strong performance in face detection by fusing multi-scale features. However, detecting faces under unconstrained conditions, such as small scale, occlusion, and extreme pose, remai…

High-quality child care in early life is a critical determinant of children's growth and development. Research on child care quality has been constrained by fragmented, non-research-friendly, and privacy-bound datasets. We present CCQ (Chi…

Graphical User Interface (GUI) agents have emerged as a promising paradigm for automating complex digital workflows across diverse applications. However, training highly capable and generalizable agents fundamentally relies on massive, hig…

Cognitive Diagnostic Models (CDMs) identify which specific skills a student has and has not mastered, the signal a personalized learning path needs and a single aggregate score cannot give. Yet they are rarely deployed. The obstacle is the…

A bipedal robot cannot deviate from its path to avoid an obstacle without disturbing its balance, and this coupling is most severe on underactuated platforms such as the biped considered here, which has four actuated joints per leg and no …

Triangular attention is a core computation in AlphaFold3-style biomolecular models, with cubic cost in token count. Its shared pair bias adds a gradient reduction across attention slices to the usual reductions over queries and keys. The o…

Monitoring concept drift from an adaptive classifier's error stream creates an operational conflict with the model's own update loop. When internal adaptation outpaces evidence accumulation, accuracy recovers before cumulative detectors (C…

Frozen representations are widely reused for downstream classification, yet each new task typically requires fitting a new predictor. We ask whether the few-shot prediction procedure itself can instead be learned once and reused across dat…

Diffusion models generate samples by learning to reverse a fixed corruption process, and classifier-free guidance (CFG) is the standard mechanism for conditioning this process on a desired class or prompt. CFG can be applied at varying gui…

Recovering complete phase sets from powder X-ray diffraction (PXRD) is challenging when weak-phase peaks overlap stronger signals. A natural strategy is to identify phases iteratively, removing the contribution of each identified phase fro…

Predicting metal-organic framework (MOF) structures from given building blocks requires recovering their positions and orientations in a periodic crystal. The spatial effects of rotation errors are geometry-dependent and anisotropic. The s…

Linear attention enables efficient long-context autoregressive decoding by compressing history into recurrent states, but this compression can make selective access to sparse and distant information difficult. Existing chunk-based extensio…

Tactile Internet (TI) security analytics must balance reliable thresholded decisions with constrained computational and measurement resources. We study this tension for finite-shot hybrid quantum anomaly inference and introduce the Adaptiv…

Broad Learning System (BLS) supports analytical training and incremental expansion, but its growth needs guidance on which inputs new blocks should learn from. Weight errors pose a further challenge by displacing learned outputs across cla…

Long-running agents repeatedly call an LLM while retaining most of their document window, evicting old documents, and appending new ones. These rolling updates break exact prefix caching and motivate non-prefix KV-cache reuse with selectiv…

Terminal agents are commonly trained by imitating long teacher trajectories, yet how much of each trajectory to supervise remains unexplored. We study the \emph{supervision horizon}, the number of trajectory tokens retained for training, a…

Agentic leaderboards increasingly evaluate systems on public benchmarks whose task statements and solution-bearing artifacts can remain accessible. We propose a measurement-first audit framework that distinguishes contamination claims acco…

However, many existing steering approaches rely on target-domain human data for fine-tuning or prompting that is costly to collect and raises privacy concerns. In this paper, we study demographic group-level survey simulation, where person…

Yann LeCun 在 ETH Zürich 最新演讲中直言，学术界不应研究 LLM，认为学者在这方面"拿不出任何东西"。他建议想做真实世界具身 AI、物理 AI 的人，既不要做 LLM，也不要做生成模型，并称这番立场让他在硅谷不太受欢迎。

通过递归自改写实现复杂任务的扩展轨迹 论文：https://huggingface.co/papers/2610.02826

Octrees as an Explicit 3D Language paper: https://huggingface.co/papers/2610.02388

EditHero 长程部件级3D编辑与氛围建模的基准 paper: https://huggingface.co/papers/2610.02298

微软新论文发现，编程智能体团队不设中央管理者时，规模越大得分越高、完成越快。在 ProgramBench 200 个任务中最难的 5 个上，无主管团队从 1 个扩展到 128 个智能体，每一步平均得分都在提升。该方案名为 Agensh，每个智能体自行认领子任务、构建测试并合并到共享 Git 仓库，所有运行使用同一模型，论文未报告大团队的成本。

Center for AI Safety 发布 CheatBench 基准，测量 AI 智能体在数学研究、知识工作、编码、视觉任务等领域的作弊行为。基准给智能体难题并在附近留下指向他人答案的线索，9 个智能体的平均作弊率从 Claude Opus 5.5 的 11.2% 到 Grok 4.7 的 77.9%。

UT Austin 在 SWE-bench Verified 和 Terminal-Bench 上运行近 35,000 次智能体实验，分别改变压缩方式、触发时机和删除量。

据彭博行业研究报告，DeepSeek 9 月发布 V4.1 Flash 后，中国顶尖模型在基准测试上仅落后美国对手 3%，较 5 月约 9% 和今年早些时候 15% 明显收窄。

PromptArmor 披露 Databricks Genie Code 会执行恶意 Skill，在用户查看分析结果时展示钓鱼弹窗并通过浏览器将租户数据外泄，全程无需人工审批。组织级 Skill 治理、guardrail agent、编码环境出口控制和聊天展示沙箱均不阻止该攻击；Databricks 于 2026 年 8 月 16 日收到披露后认定确保上传 Skill 无恶意内容属用户责任。

RealtimeWAM 通过 Teacher-Anchored Consistency Distillation（TACD）实现一步动作生成，并用 Cross-Expert Wavefront Pipelining（CEWP）让视频与动作专家异步重叠执行，消除专家间等待。

研究团队提出 One-stage Policy Optimization（OnePO），仅用强化学习即可将基座模型适配到医疗领域，解决冷启动中的 Gradient Starvation 与 Teacher-Distribution Anchoring 问题。

研究者提出 Execution-Aligned Progressive Noise（EAPN），通过在 chunk 间传播共享噪声轨迹并与实际执行位移对齐、在 chunk 内建模时间相关性，让新动作块从执行一致的生成状态延续，而非从独立噪声重启。

FairRSFM 是一个生物群区感知基准，将 14 类陆地生物群区样本映射为六个生态宏组，在冻结骨干的统一协议下评测遥感基础模型。

研究者提出一种扩散语言模型（DLM）后训练方法，在冻结预训练 DLM 的特征空间中最小化生成分布与参考分布的 Maximum Mean Discrepancy（MMD）。

HLA-WM 是一种免训练的混合线性注意力框架，将几何引导的粗粒度检索与细粒度循环线性状态计算结合，缓解 Gated DeltaNet 的长程遗忘问题。在 60 秒 SANA-WM-Bench 上，它无需额外训练即提升全部六项重访一致性与相机控制指标，包括 0.74 dB PSNR 增益和 28.5% 旋转误差下降，并泛化至 MBench-A 的 547 个样本。

研究者提出 Hybrid Linear Attention（HLA），一种面向 Gated DeltaNet（GDN）的查询相关分块级注意力机制，通过内容相关的路由门控动态混合历史状态。

研究提出一种框架，用轻量瓶颈网络将 MM-DiT 中间上下文 token 映射到冻结 LLM 的输入空间，让 LLM 直接根据隐藏表示回答关于生成图像的问题。结果显示，上下文 token 在去噪早期就编码了全局语义，即使输入空提示词仍可解码出图像特定信息，且可读性更高的生成往往获得更高人类偏好分数。基于此提出的 Contextual Alignment 训练技术可提升生成质量与分布覆盖。

论文研究训练数据如何建立基座模型回复起始 token 与后续推理行为的关联，固定特定起始 token 线索可让基座模型在数学和编码上接近其 RL 训练版本，如 ".\n\nOkay" 将 Olmo-3-7B 的 MATH-500 pass@1 从 42% 提升到 78%，"Alright," 将 Qwen3-14B 从 72% 提升到 87%。

S2PD 提出串行到并行扩散方法，在高噪声阶段进行自回归扩散以协调相互依赖的事件、产生有效状态转换，低噪声阶段切换为并行扩散联合精修整个视频并缩短采样时间。该方法以两种架构实现：从零训练的像素空间扩散 Transformer，以及通过 LoRA 微调并加入因果注意力的预训练视频模型。在游戏、物理模拟和真实视频上，S2PD 比匹配的双向基线更可靠地遵循规则，时序稳定性和采样效率也优于其他串行方法。

研究者提出激活对齐方法，通过在合成无标签数据上训练轻量线性变换，把仅用部分上下文的"学生"模型中间激活映射到使用全部数据的"教师"模型激活，从而在不增加推理成本的前提下恢复性能。