机器学习学术速递[5.27]

点击阅读原文访问arxivdaily.com，涵盖CS|物理|数学|经济|统计|金融|生物|电气领域，更有搜索、收藏等功能！

cs.LG 方向，今日共计219篇

大模型相关(34篇)

【1】Guiding LLM Post-training Data Engineering with Model Internals from Sparse Autoencoders
标题：使用Sparse Autoencoders的模型内部指导LLM训练后数据工程
链接：https://arxiv.org/abs/2605.27354

作者：Yi Jing,Zao Dai,Jinwu Hu,Zijun Yao,Lei Hou,Juanzi Li,Xiaozhi Wang
摘要：模型内部编码了关于大型语言模型（LLM）如何处理其训练数据的丰富信息;然而，训练后的数据工程在很大程度上依赖于外部信号，而忽略了模型内部丰富的内在信号。我们提出了SAERL，一个用于LLM强化学习（RL）的数据工程框架。它建模三个内在的数据属性：多样性，难度和质量，使用稀疏自动编码器（SAE），一个先进的机械可解释性工具提取的模型内部。每个属性都是一个具体的数据工程操作的基础：SAE空间聚类与适度的批量混合，用于批量多样性控制，用于从简单到困难的课程排序的难度代理，以及用于数据过滤的质量探测器。SAERL比普通GRPO提高了3.00%的平均准确率，并在Qwen2.5-Math-1.5B上减少了20%的训练步骤，达到了目标准确率，并且在模型尺度和RL算法上获得了一致的收益。实验表明，SAE有效地跨模型族和规模的传输，作为一个轻量级的和可重用的数据工程工具。这些结果表明，模型内部是训练后数据工程的强大而实用的信号源。
摘要：Model internals encode rich information about how a large language model (LLM) processes its training data; however, post-training data engineering largely relies on external signals and ignores rich intrinsic signals lying in model internals. We propose SAERL, a data engineering framework for LLM reinforcement learning (RL). It models three intrinsic data properties: diversity, difficulty, and quality, using model internals extracted with Sparse Autoencoder (SAE), an advanced mechanistic interpretability tool. Each property grounds a concrete data engineering operation: SAE-space clustering with moderate batch mixing for batch diversity control, a difficulty proxy for easy-to-hard curriculum ordering, and a quality probe for data filtering. SAERL improves average accuracy by 3.00% over vanilla GRPO and reaches target accuracy with 20% fewer training steps on Qwen2.5-Math-1.5B, with consistent gains across model scales and RL algorithms. Experiments show that SAE transfers effectively across model families and scales, serving as a lightweight and reusable data engineering tool. These results demonstrate that model internals are a powerful and practical source of signals for post-training data engineering.

【2】BASIS: Batchwise Advantage Estimation from Single-Rollout Information Sharing for LLM Reasoning
标题：BASIS：LLM推理的单次推出信息共享的批量优势估计
链接：https://arxiv.org/abs/2605.27293

作者：Shijin Gong,Erhan Xu,Kai Ye,Francesco Quinzan,Giulia Livieri,Chengchun Shi
备注：17 pages, 7 figures
摘要：具有可验证奖励的强化学习已经成为提高大型语言模型推理能力的标准方法。现有算法在价值估计和策略学习中面临计算效率和样本效率之间的权衡。我们引入了BASIS，这是一种无批评的后训练算法，旨在解决这种权衡。在每个在线训练步骤中，BASIS仅对每个提示的一个卷展栏进行采样，但利用整个批次中提示的丰富信息来改进值函数估计。我们的实验表明，与REINFORCE++相比，BASIS将值函数估计的MSE降低了69%，REINFORCE++是一种代表性的单次推出基线，并且与8次推出的组均值估计相比，一次推出的MSE更低。价值估计的这种改进可以转化为更好的策略优化：BASIS使用大大减少的训练时间，实现了接近多部署GRPO类型基线的性能，并且通常优于单部署REINFORCE类型基线。
摘要：Reinforcement learning with verifiable rewards has become a standard recipe for improving the reasoning abilities of large language models. Existing algorithms face a tradeoff between computational efficiency and sample efficiency in value estimation and policy learning. We introduce BASIS, a critic-free post-training algorithm designed to address this tradeoff. At each online training step, BASIS samples only one rollout per prompt, but leverages rich information across prompts in the entire batch to improve value function estimation. Our experiments demonstrate that BASIS reduces MSE in value function estimation by 69% compared to REINFORCE++, a representative single-rollout baseline, and achieves lower MSE with one rollout than group mean estimators with 8 rollouts. This improvement in value estimation translates to better policy optimization: using substantially less training time, BASIS achieves performance close to multi-rollout GRPO-type baselines and often outperforms single-rollout REINFORCE-type baselines.

【3】It's Not Always Sycophancy: Measuring LLM Conformity as a Function of Epistemic Uncertainty
标题：这并不总是谄媚：衡量LLM符合性作为认识不确定性的函数
链接：https://arxiv.org/abs/2605.27288

作者：Kevin H. Guo,Chao Yan,Avinash Baidya,Katherine Brown,Xiang Gao,Juming Xiong,Zhijun Yin,Bradley A. Malin
摘要：众所周知，大型语言模型（LLM）会放弃其最初的立场，以符合用户的推回。虽然先前的研究在很大程度上将这种行为归因于在从人类反馈的强化学习过程中学习到的奉承，但我们假设一致性也是由模型在推理时的认知不确定性驱动的。在本文中，我们介绍了MUSE，一个两阶段的评估框架来解开驱动LLM一致性的机制。具体来说，MUSE将模型在响应查询时的认识不确定性与其在后续回合中屈服于用户推回的可能性进行映射。我们证明，驱动一致性的机制不仅仅是奉承。具体来说，我们描述了两个不同的因素，共同驱动一致性：阿谀奉承的一致性，其中一个模型与用户的推回，即使在其初始响应的绝对确定性，和不确定性驱动的一致性，其中一个模型的一致性的可能性增加，随着其不确定性。此外，我们进行消融研究表明，阿谀奉承的一致性和不确定性驱动的一致性增长与1）LLM的感知专业知识的用户和2）用户的建议的可行性。更广泛地说，MUSE通过区分强迫诱导的奉承和训练语料库驱动的不确定性，为更有针对性的干预策略提供信息。
摘要：Large language models (LLMs) are known to abandon their initial stance to conform to user pushback. While prior research largely attributes this behavior to sycophancy learned during reinforcement learning from human feedback, we hypothesize that conformity is also driven by a model's epistemic uncertainty at inference time. In this paper, we introduce MUSE, a two-stage evaluation framework to disentangle the mechanisms driving LLM conformity. Specifically, MUSE maps a model's epistemic uncertainty in responding to a query against its likelihood to yield to user pushback in a subsequent turn. We demonstrate that the mechanisms driving conformity extend beyond sycophancy alone. Specifically, we characterize two distinct factors that jointly drive conformity: sycophantic conformity, where a model aligns with user pushback even with absolute certainty in its initial response, and uncertainty-driven conformity, where a model's likelihood for conformity increases alongside its uncertainty. Furthermore, we conduct ablation studies to demonstrate that both sycophantic conformity and uncertainty-driven conformity grow with 1) the LLM's perceived expertise of the user and 2) the plausibility of the user's suggestions. More broadly, MUSE informs more targeted intervention strategies by distinguishing alignment-induced sycophancy and training-corpora-driven uncertainty.

【4】Learning When to Think While Listening in Large Audio-Language Models
标题：在大型音频语言模型中聆听时学习何时思考
链接：https://arxiv.org/abs/2605.27190

作者：Zhiyuan Song, Weici Zhao, Yang Xiao, Suhao Yu, Cheng Zhu, Jiatao Gu
备注：19 pages, 4 figures, 6 tables
摘要：
摘要：

【5】LLMs Are Already Good Tutors: Training-Free Prompt Optimization for Pedagogical Math Tutoring
标题：法学硕士已经是很好的导师：教学数学辅导的免训练即时优化
链接：https://arxiv.org/abs/2605.27088

作者：Unggi Lee, Minchul Shin, Yeil Jeong, Sookbun Lee, Jeongsu Moon, Kyungtae Joo, Eunjoo Lee, Hoilym Kwon
备注：17 pages, 5 figures
摘要：
摘要：

【6】ReMoE: Boosting Expert Reuse through Router Fine-Tuning in Memory-Constrained MoE LLM Inference
标题：ReMoE：通过内存限制MoE LLM推理中的路由器微调来促进专家重用
链接：https://arxiv.org/abs/2605.27081

作者：Xiongwei Zhu, Xiaojian Liao, Tianyang Jiang, Yusen Zhang, Liang Wang, Limin Xiao
备注：Accepted at the 43rd International Conference on Machine Learning (ICML 2026)
摘要：
摘要：

【7】Tracing Computation Density in LLMs
标题：跟踪LLM中的计算密度
链接：https://arxiv.org/abs/2605.27033

作者：Corentin Kervadec, Iuliia Lysova, Iuri Macocco, Marco Baroni, Gemma Boleda
摘要：
摘要：

【8】Evaluating the Relevance of Uncertainty Estimators for LLM Hallucination
标题：评估LLM幻觉不确定性估计器的相关性
链接：https://arxiv.org/abs/2605.27016

作者：Yedidia Agnimo, Anna Korba, Annabelle Blangero, Nicolas Chesneau, Karteek Alahari
备注：35 pages, 7 figures, 9 tables
摘要：
摘要：

【9】Negligible in Size, Significant in Effect: On Scale Vectors in Large Language Models
标题：规模可忽略，效果显着：大型语言模型中的规模载体
链接：https://arxiv.org/abs/2605.26895

作者：Mingze Wang, Shuchen Zhu, Yuxin Fang, Binghui Li, Kai Shen, Shu Zhong
备注：36 pages
摘要：
摘要：

【10】Knowledge Graphs as the Missing Data Layer for LLM-Based Industrial Asset Operations
标题：知识图是基于LLM的工业资产运营缺失的数据层
链接：https://arxiv.org/abs/2605.26874

作者：Madhulatha Mandarapu, Sandeep Kunkunuru
备注：16 pages, 12 tables. Positions a typed knowledge-graph data layer orthogonally to the LLM-orchestration paradigms (Agent-As-Tool vs Plan-Execute) compared in AssetOpsBench (KDD 2026). Adds a same-model gpt-4.1 NLQ row and the IBM 3-axis rubric re-scoring. Code: this https URL
摘要：
摘要：

【11】MONA: Muon Optimizer with Nesterov Acceleration for Scalable Language Model Training
标题：MONA：用于可扩展语言模型训练的Nesterov加速Muon优化器
链接：https://arxiv.org/abs/2605.26842

作者：Jiacheng Li, Jianchao Tan, Hongtao Xu, Jiaqi Zhang, Yifan Lu, Yerui Sun, Yuchen Xie, Xunliang Cai
摘要：
摘要：

【12】Stabilizing Recurrent Dynamics for Test-Time Scalable Latent Reasoning in Looped Language Models
标题：稳定循环语言模型中测试时可扩展潜在推理的回归动力学
链接：https://arxiv.org/abs/2605.26733

作者：Xiao-Wen Yang, Ziyu Han, Xi-Hua Zhang, Wen-Da Wei, Jie-Jing Shao, Lan-Zhe Guo, Yu-Feng Li
备注：ICML 2026
摘要：
摘要：

【13】MemFail: Stress-Testing Failure Modes of LLM Memory Systems
标题：MemFail：LLM存储系统的压力测试故障模式
链接：https://arxiv.org/abs/2605.26667

作者：Ishir Garg, Neel Kolhe, Dawn Song, Xuandong Zhao
摘要：
摘要：

【14】WINDQuant: Weight-Informed Neural Decision-Making for Global Mixed-Precision LLM Quantization
标题：WINDQuant：全球混合精度LLM量化的权重知情神经决策
链接：https://arxiv.org/abs/2605.26660

作者：Phong Nam Huu Nguyen, Khoi M. Le, Cong-Duy T Nguyen, Anh Tuan Luu, Thong Thanh Nguyen, Tho Quan
摘要：
摘要：

【15】Cordyceps: Covert Control Attacks on LLMs via Data Poisoning
标题：虫草：通过数据中毒对LLM进行秘密控制攻击
链接：https://arxiv.org/abs/2605.26595

作者：Zedian Shao, Charles Fleming, Teodora Baluta
摘要：
摘要：

【16】SEC-bench Pro: Can Language Models Solve Long-Horizon Software Security Tasks?
标题：SEC-长凳Pro：语言模型可以解决长期软件安全任务吗？
链接：https://arxiv.org/abs/2605.26548

作者：Hwiwon Lee, Jiawei Liu, Dongjun Kim, Ziqi Zhang, Chunqiu Steven Xia, Lingming Zhang
摘要：
摘要：

【17】Open-Weight LLM Fine-Tuning Defenses are Susceptible to Simple Attacks
标题：开量级LLM微调防御容易受到简单攻击
链接：https://arxiv.org/abs/2605.26526

作者：Kevin Kuo, Chhavi Yadav, Virginia Smith
备注：main body: 9 pages, 3 figures
摘要：
摘要：

【18】Dense2MoE: Pushing the Pareto Frontier of On-Device LLMs via Unified Pruning and Upcycling
标题：Dense 2 MoE：通过统一修剪和上循环推动设备上LLM的帕累托前沿
链接：https://arxiv.org/abs/2605.26496

作者：Fengfa Li, Hongjin Ji, Yifeng Ding, Lei Ren, Chen Wei
备注：19 pages
摘要：
摘要：

【19】Elias in the Lighthouse, Again? Diagnosing Low Diversity in LLM Stories
标题：埃利亚斯又在灯塔里？诊断LLM故事中的低多样性
链接：https://arxiv.org/abs/2605.26492

作者：Sil Hamilton, David Mimno
摘要：
摘要：

【20】The Stability of Singular Distribution: A Spectral Perspective on the Two-Phase Dynamics of Language Model Pre-training
标题：奇异分布的稳定性：语言模型预训练两阶段动力学的光谱视角
链接：https://arxiv.org/abs/2605.26489

作者：Hongtao Zhang, Wenjie Zhou, Chenxi Jia, Wei Chen, Xueqi Cheng
摘要：
摘要：

【21】Extra-Merge: Tracing the Rank-1 Subspace of Model Merging in Language Model Pre-Training
标题：Extra-Merge：语言模型预训练中模型合并的秩1子空间追踪
链接：https://arxiv.org/abs/2605.26484

作者：Wenjie Zhou, Bohan Wang, Hongtao Zhang, Chenxi Jia, Wei Chen, Xueqi Cheng
摘要：
摘要：

【22】MuCon: Clipped Muon Updates for LLM Training
标题：MuCon：LLM训练的剪辑Muon更新
链接：https://arxiv.org/abs/2605.26459

作者：Albert Yi
摘要：
摘要：

【23】Reasoning, Code, or Both? How Large Language Models Handle Variations in Math Questions
标题：推理、代码还是两者兼而有之？大型语言模型如何处理数学问题的变化
链接：https://arxiv.org/abs/2605.26414

作者：Matthew Kutakh
备注：6 pages, 4 figures, 2 tables
摘要：
摘要：

【24】QAM-W: Joint 2D Codebook Quantization for LLM Weights via Hadamard Rotation and Activation-Aware Scaling
标题：QAM-W：通过Hadamard旋转和激活感知缩放对LLM权重进行联合2D码本量化
链接：https://arxiv.org/abs/2605.26339

作者：Preetam Sharma, Kacper Dobek
摘要：
摘要：

【25】Benchmarking Convolutional, Transformer, Hybrid, and Vision Language Models for Multi Disease Retinal Screening
标题：对卷积、Transformer、混合和视觉语言模型进行基准测试，用于多疾病视网膜筛查
链接：https://arxiv.org/abs/2605.26283

作者：Durjoy Dey, Aymane Ajbar, Yuhong Yan
备注：12 pages, 3 figures, accepted at ICMHI 2026, 10th International Conference on Medical and Health Informatics, Kyoto, Japan. To appear in ACM Conference Proceedings
摘要：
摘要：

【26】The Bridge-Garden Dilemma in LLM Distillation: Why Mixing Hard and Soft Labels Works
标题：LLM蒸馏中的桥花园困境：为什么混合硬标签和软标签有效
链接：https://arxiv.org/abs/2605.26246

作者：Guanghui Wang, Kaiwen Lv Kacuila, Zhiyong Yang, Zitai Wang, Jin-Wen Wu, Longtao Huang, Qianqian Xu, Qingming Huang
备注：Accepted at ICML 2026
摘要：
摘要：

【27】SetupX: Can LLM Agents Learn from Past Failures in Functionality-Correct Code Repository Setup?
标题：SetupX：LLM代理可以从功能正确代码存储库设置中过去的故障中学习吗？
链接：https://arxiv.org/abs/2605.26186

作者：Zihang Zhou, Ziqian Ren, Yukai Wu, Yingjie Xiong, Wei Zhou, Chao Peng, Dong Zhang, Bingheng Yan, Xuanhe Zhou, Fan Wu
备注：21 pages, 6 figures
摘要：
摘要：

【28】InfoQuant: Shaping Activation Distributions for Low-Bit LLM Quantization
标题：InfoQuant：塑造低位LLM量化的激活分布
链接：https://arxiv.org/abs/2605.26175

作者：Ke Li, Dong An, Xiaoling Zang, Can Ye, Liang Xie, Qibo Qiu, Chen Shen, Xiaofei He, Wenxiao Wang
摘要：
摘要：

【29】Device Context Protocol: A Compact, Safety-First Architecture for LLM-Driven Control of Constrained Devices
标题：设备上下文协议：一种紧凑、安全第一的架构，用于LLM驱动的受约束设备控制
链接：https://arxiv.org/abs/2605.26159

作者：Dongxu Yang
备注：15 pages, 5 figures. Reference implementation, Python package (pip install pydcp), and reproduction scripts at this https URL
摘要：
摘要：

【30】Turning Bias into Bugs: Bandit-Guided Style Manipulation Attacks on LLM Judges
标题：将偏见变成错误：对LLM评委的强盗引导风格操纵攻击
链接：https://arxiv.org/abs/2605.26156

作者：Xianglin Yang, Bryan Hooi, Gelei Deng, Tianwei Zhang, Jin Song Dong
备注：Accepted to the Forty-Third International Conference on Machine Learning (ICML 2026)
摘要：
摘要：

【31】Pretraining Data Exposure in Large Language Models: A Survey of Membership Inference, Data Contamination, and Security Implications
标题：大型语言模型中的预训练数据暴露：成员推断、数据污染和安全影响的调查
链接：https://arxiv.org/abs/2605.26133

作者：Ziyi Tong, Feifei Sun, Le Minh Nguyen
备注：accepted by NLDB 2025
摘要：
摘要：

【32】Self-Verified Distillation: Your Language Model Is Secretly Its Own Synthetic Data Pipeline
标题：自我验证蒸馏：您的语言模型秘密地是其自己的合成数据管道
链接：https://arxiv.org/abs/2605.26132

作者：Tony Lee, Percy Liang
摘要：
摘要：

【33】The Constraint Tax: Measuring Validity-Correctness Tradeoffs in Structured Outputs for Small Language Models
标题：约束税：衡量小型语言模型结构化输出的有效性和正确性权衡
链接：https://arxiv.org/abs/2605.26128

作者：Jaideep Ray
摘要：
摘要：

【34】GEM: Geometric Entropy Mixing for Optimal LLM Data Curation
标题：GEM：优化LLM数据处理的几何熵混合
链接：https://arxiv.org/abs/2605.26121

作者：Yue Min, Ziyun Qiao, Ruining Chen, Yujun Li
备注：Submitted to ICML 2026
摘要：
摘要：

Graph相关(图学习|图神经网络|图优化等)(8篇)

【1】Learning Dynamic Graph Representations through Timespan View Contrasts
标题：通过时间跨度视图对比学习动态图表示
链接：https://arxiv.org/abs/2605.27063

作者：Yiming Xu, Zhen Peng, Bin Shi, Xu Hua, Bo Dong
备注：Accepted by Neural Networks
摘要：
摘要：

【2】TED: Related Party Transaction guided Tax Evasion Detection on Heterogeneous Graph
标题：TED：异类图上的关联方交易引导的逃税检测
链接：https://arxiv.org/abs/2605.26984

作者：Yiming Xu, Bin Shi, Bo Dong, Jiaxiang Wang, Hua Wei, Qinghua Zheng
备注：Accepted by Data Mining and Knowledge Discovery (DMKD25)
摘要：
摘要：

【3】On the Detection of Commutative Factors in Factor Graphs: Necessary and Sufficient Conditions
标题：关于因子图中交换因子的检测：充要条件
链接：https://arxiv.org/abs/2605.26908

作者：Malte Luttermann, Ralf Möller, Marcel Gehrke
摘要：
摘要：

【4】Generalist Graph Anomaly Detection via Prototype-Based Distillation
标题：通过基于原型的蒸馏进行通才图异常检测
链接：https://arxiv.org/abs/2605.26857

作者：Yiming Xu, Zihan Chen, Zhen Peng, Song Wang, Bin Shi, Bo Dong, Chao Shen
备注：Accepted by ICML 2026
摘要：
摘要：

【5】Beyond Trajectory-Level Attribution: Graph-Based Credit Assignment for Agentic Reinforcement Learning
标题：超越轨迹级归因：基于图的强化学习学分分配
链接：https://arxiv.org/abs/2605.26684

作者：Xin Cheng, Shuo He, Lang Feng, HaiYang Xu, Ming Yan, Lei Feng, Bo An
备注：Accepted by ICML 2026
摘要：
摘要：

【6】DDGAD: Trajectory Dynamics for Diffusion-Based Graph Anomaly Detection
标题：DDGAD：基于扩散的图异常检测的轨迹动力学
链接：https://arxiv.org/abs/2605.26446

作者：Yuxin Yang, Limei Hu, Feng Chen
摘要：
摘要：

【7】Dynamic Link Prediction with Temporally Enhanced Signed Graph Neural Networks
标题：基于时间增强型符号图神经网络的动态链路预测
链接：https://arxiv.org/abs/2605.26290

作者：Derek Regier, Andrew Polyak, Aresh Dadlani, Khosro Salmani
备注：11 pages
摘要：
摘要：

【8】Provably Communication-Efficient and Privacy-Preserving Federated Graph Neural Networks
标题：可证明通信高效且隐私保护的联邦图神经网络
链接：https://arxiv.org/abs/2605.26243

作者：Zhishuai Guo, Wenhan Wu, Chen Chen, Lei Zhang, Olivera Kotevska, Ravi K Madduri
摘要：
摘要：

Transformer(7篇)

【1】Kan Extension Transformers: A Categorical Unification of Attention, Diffusion, and Predict-Detach Self-Conditioning
标题：Kan扩展Transformer：注意力、扩散和预测分离自我条件的分类统一
链接：https://arxiv.org/abs/2605.27259

作者：Sridhar Mahadevan
备注：30 pages
摘要：
摘要：

【2】JLT: Clean-Latent Prediction in Latent Diffusion Transformers
标题：JLT：潜扩散Transformer中的净潜预测
链接：https://arxiv.org/abs/2605.27102

作者：Funing Fu, Tenghui Wang, Junyong Cen, Qichao Zhu, Guanyu Zhou
摘要：
摘要：

【3】PATE-TabTransGAN: Differentially Private Synthetic Tabular Data Generation via Transformer-Based Student Discrimination
标题：PATE-TabTransGAN：通过基于transformer的学生判别生成差异私有合成表格数据
链接：https://arxiv.org/abs/2605.26802

作者：M. Youssef, M. Woźniak
备注：16 pages, 3 figures, 4 tables. Submitted for publication
摘要：
摘要：

【4】Latent Recurrent Transformer: Architecture Exploration, Training Strategies, and Scaling Behavior
标题：潜在循环Transformer：架构探索、训练策略和扩展行为
链接：https://arxiv.org/abs/2605.26797

作者：Zeyi Huang, Xuehai He, LiLiang Ren, Yiping Wang, Baolin Peng, Hao Cheng, Shuohang Wang, Pengcheng He, Jianfeng Gao, Yong Jae Lee, Yelong Shen
摘要：
摘要：

【5】CSV-ViT: A Vision Transformer with the Variable-sized Cortical Supervertices for Detection of Alzheimer's Disease Pathologies
标题：CSV-ViT：一种具有可变大小皮质超顶的视觉Transformer，用于检测阿尔茨海默病病理
链接：https://arxiv.org/abs/2605.26514

作者：Geonwoo Baek, Ikbeom Jang
摘要：
摘要：

【6】Energy-Gated Attention and Wavelet Positional Encoding: Complementary Inductive Biases for Transformer Attention
标题：能量门控注意力和子波位置编码：Transformer注意力的互补感性偏置
链接：https://arxiv.org/abs/2605.26355

作者：Athanasios Zeris
备注：10 pages, 1 figure, 3 tables. Part 2 of a five-paper series on spectral methods in transformer attention. Code: this https URL
摘要：
摘要：

【7】Transformers Can Learn Posterior Predictive Distributions In-Context
标题：Transformer可以在上下文中学习后验预测分布
链接：https://arxiv.org/abs/2605.26713

作者：Gyeonghun Kang, Changwoo J. Lee, Xiang Cheng
摘要：
摘要：

GAN|对抗|攻击|生成相关(11篇)

【1】Towards Controllable Image Generation through Representation-Conditioned Diffusion Models
标题：通过表示条件扩散模型实现可控图像生成
链接：https://arxiv.org/abs/2605.27343

作者：Nithesh Chandher Karthikeyan,Jonas Unger,Gabriel Eilertsen
摘要：扩散模型已经成为高质量图像生成和编辑的强大工具，但指导这些模型产生特定的输出仍然是一个挑战。传统的方法依赖于调节机制，如文本提示或语义映射，这需要广泛的注释数据集。在这项初步工作中，我们探索了以预训练的自监督模型表示为条件的扩散模型。自调节机制不仅提高了无条件图像生成的质量，而且提供了一个表示空间，可以用来控制生成。我们探索这个空调空间，通过确定方向的变化，并表现出有前途的性能方面的平滑度和解纠缠。
摘要：Diffusion models have emerged as powerful tools for high-quality image generation and editing, but guiding these models to produce specific outputs remains a challenge. Conventional approaches rely on conditioning mechanisms, such as text prompts or semantic maps, which require extensively annotated datasets. In this preliminary work, we explore diffusion models conditioned on representations from a pre-trained self-supervised model. The self-conditioning mechanism not only improves the quality of unconditional image generation, but also provides a representation space that can be used to control the generation. We explore this conditioning space by identifying directions of variations, and demonstrate promising properties in terms of smoothness and disentanglement.

【2】Not All Tokens Matter Equally: Dynamic In-context Vector Distillation with Decisive-Token Supervision for Long-form Medical Report Generation
标题：并非所有代币都同等重要：具有决策性代币监督的动态上下文内载体蒸馏用于长格式医疗报告生成
链接：https://arxiv.org/abs/2605.27194

作者：Ning Wu, Rui Liu, Xinkun Lin, Weixing Chen, Jinxi Xiang, Tao Wei, Lina Yao, Mingjie Li
备注：Preprint. 20 pages, 6 figures
摘要：
摘要：

【3】High-Quality Synthetic Financial Time-Series using a GAN-Diffusion Framework
标题：使用GAN扩散框架的高质量合成金融时间序列
链接：https://arxiv.org/abs/2605.27113

作者：Giuseppe Masi, Andrea Coletta, Novella Bartolini
摘要：
摘要：

【4】Adversarial Dual On-Policy Distillation from Expressive Flow-based Teacher
标题：基于表达流的教师的对抗性双重政策提炼
链接：https://arxiv.org/abs/2605.27095

作者：Zhenglin Wan, Jingxuan Wu, Xingrui Yu, Chubin Zhang, Mingcong Lei, Bo An, Ivor W. Tsang, Yang You
摘要：
摘要：

【5】When Muon Optimizer Meets Adversarial Training: A Theoretical and Empirical Study
标题：当μ子优化器遇到对抗训练：理论和实证研究
链接：https://arxiv.org/abs/2605.26929

作者：Jun Yan, Weiquan Huang, Jiankai Zuo, Yujian Mo, Xi Fang, Chengliang Wu, Zeming Wei
摘要：
摘要：

【6】Adversarial Training for Robust Coverage Network under Worst-case Facility Losses
标题：最坏情况下设施损失下稳健覆盖网络的对抗训练
链接：https://arxiv.org/abs/2605.26763

作者：Changhao Miao, Yuntian Zhang, Tongyu Wu, Fang Deng, Chen Chen
摘要：
摘要：

【7】Near-Optimal Regret in Adversarial Kernel Bandits
标题：对抗性核心盗贼中的近乎最佳遗憾
链接：https://arxiv.org/abs/2605.26585

作者：Yu-Jie Zhang, Hao Qiu, Jonathan Scarlett, Kevin Jamieson
摘要：
摘要：

【8】Distribution-Aware Conformal Prediction: A Framework for generating efficient prediction intervals for time series
标题：分布感知保形预测：为时间序列生成高效预测区间的框架
链接：https://arxiv.org/abs/2605.26569

作者：Daniel Schweizer, Peter Kuhn, Jayant Sharma, Shivali Dubey, Malte von Ramin, Christoph Brockt-Haßauer
备注：submitted to Journal of Machine Learning Research (JMLR)
摘要：
摘要：

【9】A Hybrid Vision-Language Architecture for Automated Defect Reasoning and Report Generation in Industrial Inspection
标题：用于工业检测中自动缺陷推理和报告生成的混合视觉语言架构
链接：https://arxiv.org/abs/2605.26533

作者：Malikussaid, Imad Gohar
备注：23 pages, 6 figures, 9 equations, and 6 tables
摘要：
摘要：

【10】Adversarial Water-Filling: Theory, Algorithms and Foundation Model
标题：对抗性注水：理论、算法和基础模型
链接：https://arxiv.org/abs/2605.26163

作者：Xindi Tong, Chee Wei Tan, H. Vincent Poor
备注：Submitted to IEEE Journal of Selected Topics in Signal Processing
摘要：
摘要：

【11】Furina: Fragmented Uncertainty-Driven Refusal Instability Attack
标题：Furiina：碎片化的不确定性驱动的拒绝不稳定攻击
链接：https://arxiv.org/abs/2605.26158

作者：Tongxi Wu, Jian Zhang, Yang Gao
备注：This work is accepted as a regular paper at ICML 2026
摘要：
摘要：

半/弱/无/有监督|不确定性|主动学习(7篇)

【1】LUCoS: Latent Unsupervised Context Selection for Tabular Foundation Models
标题：LUCoS：表格基础模型的潜在无监督上下文选择
链接：https://arxiv.org/abs/2605.27254

作者：Oroel Ipas, Guillermo Gomez-Trenado, Rocío Romero-Zaliz, Isaac Triguero
备注：Comments: 18 pages, 4 figures, supplementary appendices included
摘要：
摘要：

【2】FoundObj: Self-supervised Foundation Models as Rewards for Label-free 3D Object Segmentation
标题：FoundObj：自我监督的基础模型作为无标签3D对象分割的奖励
链接：https://arxiv.org/abs/2605.27178

作者：Zihui Zhang, Zhixuan Sun, Yafei Yang, Jinxi Li, Jiahao Chen, Bo Yang
备注：ICML 2026. Zihui and Zhixuan are co-first authors. Code and data are available at: this https URL
摘要：
摘要：

【3】Learning to Orchestrate Agents under Uncertainty
标题：学会在不确定性下识别代理人
链接：https://arxiv.org/abs/2605.27073

作者：Mary Chriselda Antony Oliver, Lan Jiang, Aaron Bundi Anampiu, Elaf Almahmoud, Francesco Quinzan, Umang Bhatt
摘要：
摘要：

【4】StreamSplit: Continuous Audio Representation Learning via Uncertainty-Guided Adaptive Splitting
标题：StreamSplit：通过不确定性引导的自适应拆分的连续音频表示学习
链接：https://arxiv.org/abs/2605.26523

作者：Minh K. Quan, Pubudu N. Pathirana
备注：Accepted at ACM MobiSys 2026
摘要：
摘要：

【5】Diffuse to Detect: Generative Diffusion Models for Unsupervised IC Anomaly Detection
标题：扩散检测：无监督IC异常检测的生成扩散模型
链接：https://arxiv.org/abs/2605.26468

作者：Yuxuan Yin, Chen He, Todd Jacobs, Jialei He, Boxun Xu, Robert Jin, Peng Li
备注：9 pages, 5 figures
摘要：
摘要：

【6】Planning Neural Dynamics with Lie Group Embedding through Supervised Projective Manifold Learning
标题：通过有监督的投影模学习利用李群嵌入规划神经动力学
链接：https://arxiv.org/abs/2605.26167

作者：Tianwei Wang, Bryan Chen, Qian Zuo, Qiyue Xia, Xin Li, Wei Pang
备注：Preprint. Under review
摘要：
摘要：

【7】SilIF: Silhouette-Augmented Isolation Forest for Unsupervised Transaction Fraud Detection
标题：SilIF：用于无监督交易欺诈检测的轮廓增强隔离森林
链接：https://arxiv.org/abs/2605.26135

作者：Venkatakrishnan Gopalakrishnan
备注：5 pages, 1 figure, 5 tables. Code: this https URL
摘要：
摘要：

迁移|Zero/Few/One-Shot|自适应(10篇)

【1】Transfer Learning using 66 Diseases for Disease Forecasting Applications
标题：使用66种疾病进行疾病预测应用的迁移学习
链接：https://arxiv.org/abs/2605.27269

作者：Lauren J Beesley, Alexander C Murph, Dave Osthus, Lauren A Castro
摘要：
摘要：

【2】Agile Online Model Selection: Resolving Adaptation Lag via Safeguarded Large Learning Rates
标题：敏捷在线模型选择：通过保障的大学习率解决适应滞后
链接：https://arxiv.org/abs/2605.26919

作者：Kei Takemura, Ryuta Matsuno, Keita Sakuma
备注：Accepted to KDD 2026
摘要：
摘要：

【3】More Expressive Feedforward Layers: Part I. Token-Adaptive Mixing of Activations
标题：更有表现力的前向层：第一部分。代币自适应激活混合
链接：https://arxiv.org/abs/2605.26647

作者：Mingze Wang, Jinbo Wang, Yikuan Xia, Kai Shen, Shu Zhong
备注：31 pages
摘要：
摘要：

【4】Geometry-Aware Contrastive Learning for Few-Shot Automatic Modulation Recognition
标题：用于Few-Shot自动调制识别的几何感知对比学习
链接：https://arxiv.org/abs/2605.26600

作者：Guanqun Zhao, Yitong Liu, Jiaxuan Fang, Yufei Mao, Hongwen Yang
摘要：
摘要：

【5】Few-shot Cross-country Generalization of Tabular Machine Learning and Foundation Models for Childhood Anemia Prediction under Distribution Shift
标题：分布转变下儿童贫血预测的表格式机器学习和基础模型的几次跨国推广
链接：https://arxiv.org/abs/2605.26589

作者：Yusuf Brima, Marcellin Atemkeng, Lansana Hassim Kallon, David Niyukuri, Antoine Vacavant, Samuel Saidu, Ding-Geng Chen
摘要：
摘要：

【6】When Does Deep RL Beat Calibrated Baselines? A Benchmark Study on Adaptive Resource Control
标题：Deep RL何时超越校准基线？自适应资源控制的基准研究
链接：https://arxiv.org/abs/2605.26418

作者：Guilin Zhang, Chuanyi Sun, Kai Zhao, Shahryar Sarkani, John Fossaceca
摘要：
摘要：

【7】GAC: Noise-Aware Adaptive Mixing for Hybrid SFT-RL Post-Training
标题：广汽：混合SFT-RL训练后的噪音感知自适应混音
链接：https://arxiv.org/abs/2605.26184

作者：Yuelin Hu, Zhenbo Yu, Zhengxue Cheng, Wei Liu, Li Song
备注：15 pages, 3 figures, 22 tables
摘要：
摘要：

【8】When Does Adaptive Guidance Help? Belief-Aware Privileged Distillation for Autonomous Driving Under Partial Observability
标题：适应性指导何时有帮助？部分可观察性下的自动驾驶信念感知蒸馏
链接：https://arxiv.org/abs/2605.26155

作者：Mehmet Haklidir
备注：9 pages, 3 figures, 7 tables. Accepted at CVPR 2026 Workshop on Autonomous Driving (WAD)
摘要：
摘要：

【9】Adaptive Reinforcement Learning for Robust Open Quantum System Control: A Multi-Task Framework with Temporal Optimization
标题：用于鲁棒开放量子系统控制的自适应强化学习：具有时间优化的多任务框架
链接：https://arxiv.org/abs/2605.26925

作者：Haftu W. Fentaw, Steve Campbell, Simon Caton
摘要：
摘要：

【10】Structure-Adaptive Conformal Inference for Large-Scale Out-of-Distribution Testing
标题：大规模分布外测试的结构自适应共形推理
链接：https://arxiv.org/abs/2605.26429

作者：Rongyi Sun, Wenguang Sun, Zinan Zhao
摘要：
摘要：

强化学习(6篇)

【1】Alignment Tampering: How Reinforcement Learning from Human Feedback Is Exploited to Optimize Misaligned Biases
标题：对齐篡改：如何利用人类反馈的强化学习来优化未对齐的偏差
链接：https://arxiv.org/abs/2605.27355

作者：Dongyoon Hahm,Dylan Hadfield-Menell,Kimin Lee
备注：Accepted at ICML 2026, Source code: https://alignment-tampering.github.io/
摘要：基于人类反馈的强化学习（RLHF）是将大型语言模型（LLM）与人类偏好对齐的标准方法。在这项工作中，我们引入了对齐篡改，这是一个潜在的漏洞，其中LLM进行对齐会影响偏好数据集，导致RLHF放大不期望的行为。这源于RLHF的核心限制：（1）偏好数据集是根据LLM自己的输出构建的，允许它影响它们，（2）成对比较只表明哪个响应更好，而不是为什么。这些限制可被利用来引起对准篡改。例如，如果LLM生成具有较高质量的有偏见的响应，注释者将基于质量更喜欢它们。然而，偏好标签并不能区分质量和偏差，奖励模型继承了这一局限性。通过强化学习或N中最佳抽样来优化这种奖励可能会放大不一致的偏差。我们的实验证明了不同偏见的放大：从关键词偏见到宣传（例如，性别歧视）、品牌推广和工具性目标寻求。缓解仍然具有挑战性，因为用于鲁棒RLHF的现有技术无法在不牺牲响应质量的情况下完全解决对准篡改。这些研究结果揭示了当前RLHF的结构脆弱性，并强调需要防止这种脆弱性。项目页面：https://alignment-tampering.github.io/
摘要：Reinforcement Learning from Human Feedback (RLHF) is the standard method to align Large Language Models (LLMs) with human preferences. In this work, we introduce alignment tampering, a potential vulnerability where the LLM undergoing alignment influences the preference dataset, causing RLHF to amplify undesired behaviors. This arises from core limitations of RLHF: (1) preference datasets are constructed from the LLM's own outputs, allowing it to influence them, and (2) pairwise comparisons only indicate which response is better, not why. These limitations can be exploited to cause alignment tampering. For example, if an LLM generates biased responses with higher quality, annotators will prefer them based on quality. However, preference labels do not distinguish quality from bias, and the reward model inherits this limitation. Optimizing such rewards through reinforcement learning or best-of-N sampling can amplify misaligned biases. Our experiments demonstrate amplification across diverse biases: from keyword bias to propaganda (e.g., sexism), brand promotion, and instrumental goal-seeking. Mitigation remains challenging, as existing techniques for robust RLHF fail to fully resolve alignment tampering without sacrificing response quality. These findings reveal structural vulnerabilities of current RLHF and emphasize the need to prevent this vulnerability. Project page: https://alignment-tampering.github.io/

【2】SQARL: A Size-Agnostic Reinforcement Learning approach for Circuit Allocation in Distributed Quantum Architectures
标题：SQARL：一种用于分布式量子体系结构中电路分配的与尺寸无关的强化学习方法
链接：https://arxiv.org/abs/2605.27027

作者：Víctor Carballo, Júlia López-Closa, Mario Martin
摘要：
摘要：

【3】Focal Reward: Balanced Reinforcement Learning under Rubric-Based Rewards
标题：焦点奖励：基于条目的奖励下的平衡强化学习
链接：https://arxiv.org/abs/2605.26579

作者：Yu Huang, Zihua Zhao, Zhaoxin Huan, Wanli Gu, Feng Hong, Xinmu Ge, Lin Yuan, Weichang Wu, Qiang Hu, Xiaolu Zhang, Jun Zhou, Jiangchao Yao
备注：Preprint
摘要：
摘要：

【4】Robust Koopman Control Barrier Filters for Safe Actor-Critic Reinforcement Learning
标题：用于安全的演员批评强化学习的稳健Koopman控制屏障过滤器
链接：https://arxiv.org/abs/2605.26452

作者：Dhruv S. Kushwaha, Zoleikha A. Biron
备注：17 pages, 7 figures
摘要：
摘要：

【5】MechRL: Reinforcement Learning Agents Perform Circuit Discovery for Mechanistic Interpretability
标题：MechRL：强化学习代理执行电路发现以实现机械解释性
链接：https://arxiv.org/abs/2605.26343

作者：Barsat Khadka
摘要：
摘要：

【6】Scaling World-Model Reinforcement Learning Through Diffusion Policy Optimization
标题：通过扩散政策优化扩展世界模型强化学习
链接：https://arxiv.org/abs/2605.26282

作者：Xiaoyuan Cheng, Wenxuan Yuan, Zhancun Mu, Yuanzhao Zhang, Yiming Yang, Hai Wang, Zhuo Sun, Che Liu
摘要：
摘要：

元学习(1篇)

【1】Beyond Differences: Doubly Robust Meta-Learners for Ratio-Based Treatment Effects
标题：超越差异：基于比率的治疗效果的双重稳健元学习者
链接：https://arxiv.org/abs/2605.26288

作者：Michael Fuchs, Dominik Kreiss
备注：13+5 pages, 5 figures, 6 tables. Code: this https URL
摘要：
摘要：

符号|符号学习(1篇)

【1】Symbolic Regression via Latent Iterative Refinement
标题：通过潜在迭代细化的符号回归
链接：https://arxiv.org/abs/2605.27245

作者：Xieting Chu, Sriram Vishwanath, Vijay Ganesh
备注：Preprint. 21 pages, 11 figures
摘要：
摘要：

医学相关(4篇)

【1】Implementation of Big Data Analytics for Diabetes Management: Needs Assessment in the Rwanda Healthcare System
标题：糖尿病管理大数据分析的实施：卢旺达医疗保健系统的需求评估
链接：https://arxiv.org/abs/2605.26786

作者：Silas Majyambere, Tony Lindgren, Workneh Y. Ayele, Celestin Twizere
摘要：
摘要：

【2】On the Role of Inductive Bias in Time-Series Pretraining: A Case Study in Learning Generalizable Representations for Clinical Time Series
标题：论归纳偏差在时间序列预训练中的作用：学习临床时间序列可概括表示的案例研究
链接：https://arxiv.org/abs/2605.26194

作者：Sharmita Dey, Diego Paez-Granados
摘要：
摘要：

【3】Prospective evaluation of multimodal respiratory failure prediction: Do chest X-rays improve performance beyond EHR signals?
标题：多模式呼吸衰竭预测的前瞻性评估：胸部X光检查是否能改善EHR信号之外的表现？
链接：https://arxiv.org/abs/2605.26255

作者：Xiaolei Lu, Shamim Nemati
摘要：
摘要：

【4】What Molecular Structure Cannot Tell Us: A Taxonomy of Explainability Gaps in GNN-Based Drug Toxicity Prediction
标题：分子结构无法告诉我们什么：基于GNN的药物毒性预测中解释性差距的分类学
链接：https://arxiv.org/abs/2605.26183

作者：Juergen Dietrich
备注：14 pages
摘要：
摘要：

蒸馏|知识提取(2篇)

【1】Less is More: Early Stopping Rollout for On-Policy Distillation
标题：少即是多：提前停止推出按政策蒸馏
链接：https://arxiv.org/abs/2605.27028

作者：Zhou Ziheng, Jiaqi Li, Huacong Tang, Ying Nian Wu, Demetri Terzopoulos
摘要：
摘要：

【2】Not All Disagreement Is Learnable: Token Teachability in On-Policy Distillation
标题：并非所有分歧都是可以学习的：按政策蒸馏中的代币可教性
链接：https://arxiv.org/abs/2605.26844

作者：Yuanyi Wang, Su Lu, Yanggan Gu, Pengkai Wang, Yifan Yang, Zhaoyi Yan, Congkai Xie, Jianmin Wu, Hongxia Yang
摘要：
摘要：

推荐(1篇)

【1】Causal Representation Learning for Generalisable Recommendation
标题：可概括推荐的因果表示学习
链接：https://arxiv.org/abs/2605.27043

作者：Yorgos Felekis, Michael O'Riordan, Oriol Corcoll, Ciarán M. Gilligan-Lee
摘要：
摘要：

超分辨率|去噪|去模糊|去雾(1篇)

【1】AirCast-SR: A Foundation Model for Kilometer-Scale Atmospheric Super-Resolution via Latent Consistency Diffusion
标题：AirCast-SR：通过潜浓度扩散实现公里级大气超分辨率的基础模型
链接：https://arxiv.org/abs/2605.26130

作者：Somnath Luitel, Manmeet Singh, Joshua Durkee, Abdullah Al Fahad, Naveen Sudharsan, Prabhjot Singh, Cenlin He, Harsh Kamath, Zong-Liang Yang, Krishnagopal Halder, Sandeep Juneja, Parthasarathi Mukhopadhyay, Saptarishi Dhanuka, Amit Kumar Srivastava
备注：Somnath Luitel and Manmeet Singh are equal-contribution co-first authors, with Manmeet Singh (this http URL@wku.edu) as corresponding author
摘要：
摘要：

自动驾驶|车辆|车道检测等(2篇)

【1】Towards Generalization-Oriented Models for Vehicle Routing Problems with Mixture-of-Experts
标题：面向概括的混合专家车辆路径问题模型
链接：https://arxiv.org/abs/2605.26776

作者：Changhao Miao, Yuntian Zhang, Tongyu Wu, Fang Deng, Chen Chen
摘要：
摘要：

【2】Uniboost: Global Coordination with Value Alignment for Fair and Efficient Traffic Allocation
标题：Uniboost：全球协调与价值一致，实现公平有效的流量分配
链接：https://arxiv.org/abs/2605.26424

作者：Ge Fan, Nan Zhao, Kai Meng, Cong Luo, Yang Fu, Huiping Chu, Jialin Liu, Yuning Jiang, Bo Zheng
备注：accepted by SIGIR 2026
摘要：
摘要：

联邦学习|隐私保护|加密(2篇)

【1】Separate Aggregation of Split Network for Personalized Federated Learning
标题：用于个性化联邦学习的分裂网络的单独聚合
链接：https://arxiv.org/abs/2605.26571

作者：Yunseok Kang, Jaeyoung Song
摘要：
摘要：

【2】On the Push-Based Asynchronous Federated Learning: A Bias-Correction Aggregation Approach
标题：基于推送的同步联邦学习：一种偏差纠正聚合方法
链接：https://arxiv.org/abs/2605.26162

作者：Jiahui Bai, Hai Dong, A. K. Qin
备注：Accepted at the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2026). This is the extended version with full appendix
摘要：
摘要：

推理|分析|理解|解释(14篇)

【1】Greening AI Inference with Accuracy and Latency-aware User Incentives
标题：绿色人工智能推理，具有准确性和延迟意识的用户激励
链接：https://arxiv.org/abs/2605.27309

作者：Vasilios A. Siris,Adamantia Stamou,George D. Stamoulis,Konstantinos Varsos,Ramin Khalili
摘要：人工智能服务的广泛使用引发了对其环境可持续性的担忧，最近的研究将人工智能推理的碳排放确定为主要贡献者。本文介绍了一个框架，用于设计基于用户的评价推理质量和延迟，连同他们的环境意识，同时考虑到碳排放和两个QoE参数之间的权衡的AI推理激励。我们的方法可以适应不同的权衡，这取决于AI模型的大小和复杂性以及为推理请求提供服务的资源分配。这些激励措施可以通过实际的两级服务订阅来提供，为用户提供折扣，以换取减少的碳排放。折扣服务选项为AI提供商提供了灵活性，可以在高碳强度期间以较低的质量和较高的延迟为一定比例的推理请求提供服务。
摘要：The widespread use of AI services has raised concerns for its environmental sustainability, towards which recent studies have identified carbon emissions of AI inference as the major contributor. This paper introduces a framework for designing AI inference incentives based on the users' valuation for inference quality and latency, together with their environmental consciousness, while accounting for the tradeoff between carbon emissions and the two QoE parameters. Our approach can accommodate different tradeoffs, that depend on the size and complexity of the AI models and the allocation of resources to serve inference requests. The incentives can be offered through a practical two-tier service subscription that offers users a discount in exchange for reduced carbon emissions. The discounted service option gives the AI provider the flexibility to serve some percentage of inference requests at a lower quality and higher latency during periods of high carbon intensity.

【2】Explainable Comparison of Feature-Based and Deep Learning Models for TROPOMI Methane Plume Screening
标题：TROPOMI甲烷羽流筛查基于环境和深度学习模型的可解释比较
链接：https://arxiv.org/abs/2605.27236

作者：Solomiia Kurchaba, Joannes D. Maasakkers, Berend J. Schuit, Ilse Aben
摘要：
摘要：

【3】Nonlinear Data Integration via Kernel Methods for Data Collaboration Analysis
标题：通过数据协作分析的核心方法进行非线性数据集成
链接：https://arxiv.org/abs/2605.27219

作者：Yamato Suetake, Yuta Kawakami, Shunnosuke Ikeda, Yuichi Takano
备注：50 pages, 7 figures
摘要：

【4】Deep-layer limit and stability analysis of the basic forward-backward-splitting induced network (II): learning problems
标题：基本前向向后分裂诱导网络的深层极限和稳定性分析（II）：学习问题
链接：https://arxiv.org/abs/2605.27133

作者：Xuan Lin, Chunlin Wu
备注：38 pages, 1 figure
摘要：

【5】DEI: Diversity in Evolutionary Inference for Quality-Diversity Search
标题：DEI：质量多样性搜索进化推理的多样性
链接：https://arxiv.org/abs/2605.27130

作者：John Donaghy, Shikhar Rastogi
备注：Accepted to ICML 2026 Workshop Scalable Learning and Optimization for Efficient Multimodal AI Agents (SCALE)
摘要：

【6】EEG-FM-Audit: A Systematic Evaluation and Analysis Pipeline for EEG Foundation Models
标题：EEG-FM-Audit：脑电基础模型的系统评估和分析管道
链接：https://arxiv.org/abs/2605.26910

作者：Xianheng Wang, Yige Yang, Damien Coyle
备注：26 pages
摘要：

【7】Why Prompt Optimization Works, and Why It Sometimes Doesn't: A Causal-Inspired Edit-Level Analysis
标题：为什么提示优化有效，为什么有时不起作用：凯瑟琳启发的编辑级分析
链接：https://arxiv.org/abs/2605.26655

作者：Shuzhi Gong, Hechuan Wen
备注：17 pages, 4 figures, 8 tables
摘要：

【8】Aligning Few-Step Generative Models by Amortizing Sample-based Variational Inference
标题：通过分解基于样本的变分推理来对齐少步生成模型
链接：https://arxiv.org/abs/2605.26552

作者：Jaewoo Lee, Hyeongyu Kang, Dohyun Kim, Kyuil Sim, Woocheol Shin, Minsu Kim, Taeyoung Yun, Jeongjae Lee, Sanghyeok Choi, Tabitha Edith Lee, Jongchul Ye, Jinkyoo Park
备注：Under review
摘要：

【9】SIKA-GP: Accelerating Gaussian Process Inference with Sparse Inducing Kernel Approximations for Bayesian Deep Learning
标题：SIKA-GP：通过稀疏诱导核逼近来加速高斯过程推理，用于Bayesian深度学习
链接：https://arxiv.org/abs/2605.26509

作者：Wenyuan Zhao, Rui Tuo, Chao Tian
备注：20 pages, 8 figures; accepted to International Conference on Machine Learning (ICML) 2026
摘要：

【10】Variational Inference for Evidential Deep Learning
标题：证据深度学习的变分推理
链接：https://arxiv.org/abs/2605.26477

作者：Jiawei Tang, Xinyan Du, Hui Liu, Junhui Hou, Yuheng Jia
摘要：

【11】Amortized Factor Inference Networks for Posterior Inference
标题：后验推理的摊销因子推理网络
链接：https://arxiv.org/abs/2605.26419

作者：Joohwan Ko, Justin Domke
摘要：

【12】MULTISEISMO: A Multimodal Seismic Dataset and Model for Cross-Modal Seismic Understanding
标题：MULTISMO：用于跨模式地震理解的多模式地震数据集和模型
链接：https://arxiv.org/abs/2605.26320

作者：Sai Munikoti, Ian Stewart, Chengping Chai, Lisa Linville, Scott Vasquez, Sameera Horawalavithana, Karl Pazdernik
摘要：

【13】Stateful Inference for Low-Latency Multi-Agent Tool Calling
标题：低延迟多代理工具调用的状态推理
链接：https://arxiv.org/abs/2605.26289

作者：Victor Norgren
摘要：

【14】ARBITER: Reasoning Trajectory Basins and Majority Vote Failures in Test-Time Sampling
标题：仲裁员：测试时抽样中的推理轨迹池和多数投票失败
链接：https://arxiv.org/abs/2605.26172

作者：Meng Cai, Lars Kulik, Farhana Choudhury
备注：Preprint. 34 pages, 2 figures
摘要：

检测相关(6篇)

【1】Automatic Layer Selection for Hallucination Detection
标题：幻觉检测的自动分层选择
链接：https://arxiv.org/abs/2605.26366

作者：Xinpeng Wang, William Cao, Andrew Gordon Wilson, Zhe Zeng
备注：Accepted at ICML 2026
摘要：

【2】Classification and detection of multiple UAVs using rational Gaussian wavelet neural networks
标题：基于有理高斯子波神经网络的多无人机分类与检测
链接：https://arxiv.org/abs/2605.26310

作者：Ungvári Gergő, Ferenc Braun, Attila Ámon, Péter Kackstädter, János Volk, Péter Kovács, Tamás Dózsa
备注：19 pages, 4 figures
摘要：

【3】Bridging Classification and Reconstruction: Cooperative Time Series Anomaly Detection
标题：桥梁分类和重建：协作时间序列异常检测
链接：https://arxiv.org/abs/2605.26193

作者：Qideng Tang, Dai Chaofan, Wubin Ma, Yahui Wu, Haohao Zhou, Tao Zhang, Huan Li, Dalin Zhang
备注：Accepted by KDD 2026
摘要：

【4】When Rule Violations Are Rare: Chimera Training for Logical Anomaly Detection
标题：当规则违规很少时：逻辑异常检测的Chimera训练
链接：https://arxiv.org/abs/2605.26171

作者：Alejandro Ascarate, Leo Lebrat, Rodrigo Santa Cruz, Clinton Fookes, Olivier Salvado
备注：9+30 pages, 4+4 figures, under review
摘要：

【5】Enhancing Autonomous Online Intrusion Detection for IoT with Balanced Learning, Reliable Pseudo-Labels, and Lightweight Architectures
标题：通过平衡学习、可靠的伪标签和轻量级架构增强物联网的自主在线入侵检测
链接：https://arxiv.org/abs/2605.26166

作者：Hanzala Afzaal, Danish Memon, Chouhdary Bilal Raza, Muhammad Khurram Shahzad
备注：9 pages, 5 figures; Code available at this https URL
摘要：

【6】Confounder Detection via Treatment Intent: A New Observational Study Design
标题：通过治疗意图检测混杂因素：一种新的观察性研究设计
链接：https://arxiv.org/abs/2605.26413

作者：Drago Plecko, Patrik Okanovic, Torsten Hoefler, Elias Bareinboim
摘要：

分类|识别(5篇)

【1】The Role of Causal Features in Strategic Classification for Robustness and Alignment
标题：因果特征在稳健性和一致性战略分类中的作用
链接：https://arxiv.org/abs/2605.27163

作者：Antonio Gois, Sophia Gunluk, Nir Rosenfeld, Nidhi Hegde, Simon Lacoste-Julien, Dhanya Sridhar
备注：Accepted at AISTATS 2026. 20 pages, 5 figures
摘要：

【2】Is an Image Also Worth 16x16=256 Superpixels? A Framework for Attentional Image Classification
标题：图像是否也值得16 x16 =256超像素？注意力图像分类框架
链接：https://arxiv.org/abs/2605.27144

作者：Pedro Henrique da Costa Avelar, Anderson R. Tavares, Luís C. Lamb
摘要：

【3】PIDM-DP: Physics-Informed Diffusion with Dormand-Prince Integration for Chaotic System Identification and State Reconstruction across Multiple Dynamical Regimes
标题：PIDM-DP：采用Dormand-Prince积分的物理知情扩散用于跨多个动态机制的混乱系统识别和状态重建
链接：https://arxiv.org/abs/2605.26619

作者：Shailendra Dabral
备注：extended work of my journal paper submission
摘要：

【4】HRVConformer: Neonatal Hypoxic-Ischemic Encephalopathy Classification from the Heart Rate signals
标题：HRVConformer：根据心率信号对新生儿缺氧缺血性脑病进行分类
链接：https://arxiv.org/abs/2605.26190

作者：Shuwen Yu, William P Marnane, Geraldine B. Boylan, Gordon Lightbody
备注：Paper submitted to Journal of Engineering Applications of Artifical Intelligence
摘要：

【5】Data-driven sparse identification of governing PDEs via knockoff filters and multi-criteria trade-offs
标题：通过仿制过滤器和多标准权衡对治理PDEs进行数据驱动的稀疏识别
链接：https://arxiv.org/abs/2605.26631

作者：Pongpisit Thanasutives, Naichang Ke, Yoshinobu Kawahara
备注：42 pages, 5 figures, 10 tables
摘要：

表征(3篇)

【1】Beyond Binary: Speech Representations Across the Cognitive Score Hierarchy
标题：超越二进制：跨认知分数层次的语音表示
链接：https://arxiv.org/abs/2605.27189

作者：Serli Kopar, Roshan Prakash Rane, Christian Mychajliw, Lydia Federmann, Gerhard Eschweiler, Daniela Berg, Sam Gijsen, Paula Andrea Perez-Toro, Kerstin Ritter
摘要：

【2】Two Speeds of Learning: A Representation-Readout Decomposition of Grokking and Double Descent
标题：两种学习速度：Grokking和双重下降的代表读出分解
链接：https://arxiv.org/abs/2605.27078

作者：Chi-Ning Chou, Oscar Uzdelewicz, Neng-Chun Chiu, Yao-Yuan Yang, SueYeon Chung
摘要：

【3】Rotation-Invariant Spherical Watermarking via Third-Order SO(3) Representation Coupling
标题：基于三阶SO（3）表示耦合的旋转不变球形水印
链接：https://arxiv.org/abs/2605.26702

作者：Pengzhen Chen, Yanwei Liu, Xiaoyan Gu, Antonios Argyriou, Wu Liu, Weiping Wang
备注：ICML 2026
摘要：

3D|3D重建等相关(1篇)

【1】TrackRef3D: Multi-View Consistent Track-then-Label for Open-World Referring Segmentation in 3D Gaussian Splatting
标题：TrackRef3D：用于3D高斯溅射中开放世界参照分割的多视图一致跟踪然后标记
链接：https://arxiv.org/abs/2605.26576

作者：Yuyang Tan, Renhe Zhang, Hang Zhang, Ao Li, Xin Tan
摘要：

优化|敛散性(6篇)

【1】Probabilistic Smoothing with Ratio-Monotone Transforms for Global Optimization
标题：利用比率单调变换进行全局优化的概率平滑
链接：https://arxiv.org/abs/2605.27316

作者：Kukyoung Jang,Taehyun Cho,Junrui Zhang,Ping Xu,Kyungjae Lee
摘要：Probabilistic smoothing is a standard tool for global optimization, but existing methods rely on Gaussian kernels and specific transforms, often resulting in strong hyperparameter sensitivity and limited robustness. We propose a general smoothing framework that combines flexible symmetric unimodal kernels with monotonic ratio-based transformations. Under mild conditions, we show that the smoothed objective preserves the global maximizer and that all stationary points concentrate near the true optimum for sufficiently large amplification, without requiring a decreasing smoothing schedule. We further provide explicit complexity bounds for stochastic gradient ascent and show that a leave-one-out baseline provably reduces variance. Experiments on high-dimensional benchmarks and black-box adversarial attacks demonstrate improved robustness and competitive performance.

【2】Convergence of Spectral Descent for Non-smooth Optimization
标题：非光滑优化谱下降的收敛性
链接：https://arxiv.org/abs/2605.26977

作者：Yixuan Yang, Yuqing He, Song Li
摘要：

【3】Ratio-Variance Regularized Policy Optimization
标题：比率方差正规化政策优化
链接：https://arxiv.org/abs/2605.26784

作者：Yu Luo, Shuo Han, Yihan Hu, Lei Lv, Huaping Liu, Fuchun Sun, Jianye Hao, Dong Li
摘要：

【4】Bilevel Optimization over Saddle Points of Zero-Sum Markov Games
标题：零和Markov对策鞍点的二层优化
链接：https://arxiv.org/abs/2605.26654

作者：Zihao Zheng, Irwin King, Songtao Lu
备注：Accepted to the International Conference on Machine Learning (ICML 2026)
摘要：

【5】Online Learning on Hidden-Convex Losses via Algorithmic Equivalence: Optimal Regret, Geometric Barrier, and Bandit Feedback
标题：通过数学等效进行隐凸损失的在线学习：最佳遗憾、几何障碍和强盗反馈
链接：https://arxiv.org/abs/2605.26373

作者：Anas Barakat, Andreas Kontogiannis, Vasilis Pollatos, Ioannis Panageas, Antonios Varvitsiotis
备注：43 pages
摘要：

【6】Stochastic global optimization of continuous functions via random walks on Grassmannians
标题：通过格拉斯曼式随机游动实现连续函数的随机全局优化
链接：https://arxiv.org/abs/2605.14151

作者：Kartik Gupta, Stephen D. Miller, Pradeep Ravikumar, Ramarathnam Venkatesan
备注：21 pages
摘要：

预测|估计(8篇)

【1】SPHERE-JEPA: Spherical Prediction with Homogeneous Embeddings
标题：SPHERE-JEPA：均匀嵌入的球形预测
链接：https://arxiv.org/abs/2605.26900

作者：Léo Nicollier (CB, ATT), Max Dunitz (CB, ATT), Marc Pic (ATT), Pablo Musé (CB, IFUMI), Enric Meinhardt-Llopis (CMLA, CB), Gabriele Facciolo (CB)
摘要：

【2】APEX: Amplitude Anchors and Phase Priors for Target-Scarce Higher-Frequency Wave Prediction
标题：APEX：用于目标稀缺高频波预测的幅度先验和相先验
链接：https://arxiv.org/abs/2605.26732

作者：Yifan Sun, Lei Cheng, Sijie Chen, Ting Zhang, Jianlong Li, Shikai Fang
摘要：

【3】SL-BiLEM: Structured Learnable Behavior-in-the-Loop Epidemic Modeling for Forecasting and Policy Evaluation
标题：SL-BiLEM：用于预测和政策评估的结构化可学习行为在环流行病模型
链接：https://arxiv.org/abs/2605.26704

作者：Haochun Wang, Sendong Zhao, Jingbo Wang, Yanrui Du, Bing Qin, Ting Liu
备注：ACM SIGKDD 2026
摘要：

【4】Beyond Holistic Models: Systematic Component-level Benchmarking of Deep Multivariate Time-Series Forecasting
标题：超越整体模型：深度多元时间序列预测的系统性学生级基准
链接：https://arxiv.org/abs/2605.26562

作者：Shuang Liang, Chaochuan Hou, Xu Yao, Shiping Wang, Hailiang Huang, Songqiao Han, Minqi Jiang
备注：accepted by KDD 2026 Datasets and Benchmarks Track
摘要：

【5】PolyFusionAgent: A Multimodal Foundation Model and Autonomous AI Assistant for Polymer Property Prediction and Inverse Design
标题：PolyFusionAgent：用于聚合物性能预测和反向设计的多峰基础模型和自主人工智能助手
链接：https://arxiv.org/abs/2605.26543

作者：Manpreet Kaur, Xingying Zhang, Qian Liu
备注：23 pages, 5 figures, 2 tables; Supplementary material included
摘要：

【6】Jailbreak susceptibility prediction and mitigation via the behavioral geometry of models
标题：通过模型的行为几何学预测和缓解越狱易感性
链接：https://arxiv.org/abs/2605.26409

作者：Hayden Helm, Xiaodong Liu, Weiwei Yang
摘要：

【7】Max-Window Scale Estimation for Near-Lossless HiF8 W8A8 Quantization-Aware Training
标题：近乎无损HiF 8 W8 A8量化感知训练的最大窗口规模估计
链接：https://arxiv.org/abs/2605.26189

作者：Yingying Cheng, Jinquan Shi, Li Zhou, Zhiyang He, Zhaoyi Sun, Fan Zhang, Jie Sun
摘要：

【8】TSFMAudit: Data Contamination Auditing in Forecasting Time Series Foundation Models
标题：TSFVMAudit：预测时间序列基础模型中的数据污染审计
链接：https://arxiv.org/abs/2605.26161

作者：Hongkai Li, Shifeng Xie, Lefei Shen, Zhuo Li, Mouxiang Chen, Xiaobin Zhang, Han Fu, Jianling Sun, Xiaoxue Ren, Chenghao Liu
备注：22 pages, 7 figures, 9 tables
摘要：

其他神经网络|深度学习|模型|建模(34篇)

【1】From Scores to Gibbs Correctors: Accelerating Uniform-Rate Discrete Diffusion Models
标题：从分数到吉布斯修正器：加速均匀率离散扩散模型
链接：https://arxiv.org/abs/2605.27352

作者：Yuchen Liang,Ness Shroff,Yingbin Liang
摘要：Discrete diffusion models have achieved strong empirical performance in text and other symbolic domains, but, especially for uniform-rate models, they often require many steps to generate a single sample. Existing acceleration methods either rely on training additional quantities or suffer from slow mixing. In this work, we propose a novel Gibbs-based corrector for discrete diffusion models, termed Gibbs-Accelerated Discrete Diffusion (GADD). GADD leverages the structure of the concrete score function to construct Gibbs posterior likelihoods directly, without requiring any additional training beyond standard score estimation. We show that GADD achieves an overall sampling complexity of $\mathcal{O}(\mathrm{polylog} (\varepsilon^{-1}))$, yielding the first such rate for diffusion-based samplers for uniform-rate discrete diffusion models. We also conduct numerical experiments demonstrating the practical advantages of GADD across synthetic data, zero-shot text sampling, and zero-shot conditional music generation. These results corroborate the theory and show that GADD consistently improves sample quality and wall-clock efficiency over standard baselines, including vanilla Euler methods and CTMC correctors. Beyond this, our theoretical analysis introduces a novel framework for analyzing predictor-corrector methods in discrete diffusion models, which may be of independent interest. Unlike existing approaches that rely on the Girsanov change-of-measure technique, our method is based on an induction argument that tracks error propagation across predictor iterations while accounting for inaccuracies in the corrector updates.

【2】Risk Averse Alert Prioritization for IDS Using Subnormal Gaussian Fuzzy Models
标题：使用次正态高斯模糊模型的IDS风险厌恶警报优先级
链接：https://arxiv.org/abs/2605.27299

作者：Murat Moran
摘要：Modern intrusion detection systems generate thousands of alerts daily, but alert fatigue severely limits security operations effectiveness due to too many false positives or low-impact events. We address this by proposing a principled framework for alert prioritization based on subnormal Gaussian fuzzy numbers, explicitly modeling three sources of uncertainty: threat severity, detection confidence, and organizational risk attitude. Each alert is represented as a fuzzy number with the core indicating severity, spread indicating uncertainty, and height reflecting detection reliability. We apply ranking indices to prioritize alerts, allowing organizations to tune security posture through a risk-attitude parameter. Experimental validation on CIC-IDS2017 and NSL-KDD demonstrates greater robustness than baselines under detector degradation (0.9963 vs 0.8215 NDCGrel@100), with distinct differentiation in mid-confidence alerts and near-parity with baselines under robust detectors. The framework is theoretically grounded, computationally efficient, provides interpretable reasoning, and remains robust across detector families and miscalibration scenarios.

【3】Falcon-X: A Time Series Foundation Model for Heterogeneous Multivariate Modeling
标题：Falcon-X：用于异类多元建模的时间序列基础模型
链接：https://arxiv.org/abs/2605.27286

作者：Yiding Liu,Yifan Hu,Hongjie Xia,Peiyuan Liu,Hongzhou Chen,Xilin Dai,Zewei Dong,Jiang-Ming Yang
摘要：Time series foundation models (TSFMs) are transforming the forecasting paradigm through large-scale cross-domain pretraining. However, most existing TSFMs remain univariate, and recent efforts to enable cross-variate modeling still operate directly within the raw variate space. This design introduces fundamental limitations in semantic alignment and relational expressivity. Specifically, raw-space group mixing lacks a dedicated mechanism to align heterogeneous physical quantities, while standard non-negative attention fails to capture the complex synergistic and antagonistic interactions ubiquitous in real-world systems. To address these challenges, we propose Falcon-X, decouples variates from the raw space and maps them into a unified latent prototype space. Falcon-X employs a Unified Prototype Diff-Attention mechanism that explicitly evaluates both positive and negative semantic affinities to explicitly align heterogeneous variates. Cross-variate interactions are then efficiently performed within this shared space via Latent Entity Attention, naturally facilitating zero-shot structural transfer. Finally, a Variate Reassembly Router robustly reconstructs variate-specific trajectories via a request-and-dispatch mechanism. Extensive evaluations on the GIFT-Eval and fev-bench benchmarks demonstrate that Falcon-X achieves state-of-the-art forecasting performance, offering a principled and scalable paradigm for complex multivariate environments. Falcon-X is publicly released to support future research.

【4】PILOT: A Data-Free Continual Learning Approach for Real-Time Semantic Segmentation via Boundary Guidance
标题：PILOT：一种通过边界引导进行实时语义分割的无数据连续学习方法
链接：https://arxiv.org/abs/2605.27128

作者：Yujing Zhou, Prashant Shekhar, Thomas Yang, Yongxin Liu
摘要：

【5】Mildly Overparameterized ReLU Networks on Orthogonal Data: Incremental Learning and Implicit Bias
标题：垂直数据上的轻度过度参数化ReLU网络：增量学习和隐式偏差
链接：https://arxiv.org/abs/2605.27097

作者：James Town, Etienne Boursier, Ben Lewis, Matthias Englert, Ranko Lazic
备注：66 pages, 6 figures
摘要：

【6】Cost of Structural Learning Under Censored Feedback: A Threshold-Bandit Approach
标题：审查反馈下的结构学习成本：股东-强盗方法
链接：https://arxiv.org/abs/2605.27076

作者：Michael Ledford, William Regli
摘要：

【7】Probabilistic Recurrent Intention Switching Model
标题：概率复发意图转换模型
链接：https://arxiv.org/abs/2605.26998

作者：Wenyuan Sheng, Hao Zhu, Joschka Boedecker
摘要：

【8】Parsimonious Learning-Augmented Online Metric Matching
标题：节俭学习增强在线指标匹配
链接：https://arxiv.org/abs/2605.26886

作者：Yongho Shin, Phanu Vajanopath
备注：To appear in ICML 2026
摘要：

【9】Learning Energy-Based Models from Stochastic Interpolants using Spatiotemporal Differences
标题：使用时空差异从随机插值学习基于能量的模型
链接：https://arxiv.org/abs/2605.26850

作者：Hanlin Yu, RuiKang OuYang, Partha Kaushik, Arto Klami, Michael U. Gutmann, Omar Chehab
摘要：

【10】Periodic Topological Deep Learning for Polymer Design and Discovery
标题：用于聚合物设计和发现的周期性布局深度学习
链接：https://arxiv.org/abs/2605.26833

作者：Yasharth Yadav, Tze Kwang Gerald Er, Atsushi Goto, Kelin Xia
备注：19 pages, 3 figures, 3 tables
摘要：

【11】Localizing Memorized Regions in Diffusion Models via Coordinate-Wise Curvature Differences
标题：通过坐标级弯曲差异局部化扩散模型中的均匀化区域
链接：https://arxiv.org/abs/2605.26756

作者：Gwangho Kim, Sungyoon Lee
备注：ICML 2026
摘要：

【12】The Need for an External Observer Formalizing the Sufficiency Gap: A Mathematical Extension of Mixture Identifiability and Contextual Grounding in Sequence Models
标题：需要外部观察者将充分性差距形式化：序列模型中混合可识别性和上下文基础的数学扩展
链接：https://arxiv.org/abs/2605.26711

作者：Francesco Corielli
摘要：

【13】Model Merging on Loss Landscape: A Geometry Perspective
标题：损失景观模型合并：几何学的角度
链接：https://arxiv.org/abs/2605.26693

作者：Juanwu Lu, Anand Bhaskar, Brian Axelrod, Ekaterina Tolstaya, Tristan Emrich
备注：CVPR 2026 Findings Track. 18 pages, 4 figures, 6 tables
摘要：

【14】RT-Lynx: Putting the GEMM Sparsity In a Right Way for Diffusion Models
标题：RT-Lynx：以正确的方式将GEMM稀疏性应用于扩散模型
链接：https://arxiv.org/abs/2605.26632

作者：Xing Cong, Hanlin Tang, Kan Liu, Lan Tao, Lin Qu, Chenhao Xie
备注：33 pages, 18 figures, Accepted by ICML 2026
摘要：

【15】Bridging Control with Neural Network Verifier alpha-beta-CROWN: A Tutorial
标题：与神经网络验证器的桥梁控制alpha-Beta-CROWN：一个预设
链接：https://arxiv.org/abs/2605.26577

作者：Haoyu Li, Xiangru Zhong, Hao Cheng, Bin Hu, Huan Zhang
备注：ACC 2026 Tutorial
摘要：

【16】Auditing and Fixing Economic Validity in Tabular Foundation Models for Discrete Choice
标题：离散选择表格基础模型中的审计和确定经济有效性
链接：https://arxiv.org/abs/2605.26559

作者：Yingshuo Wang, Xian Sun, Yanhang Li, Zhichao Fan, Zexin Zhuang
备注：5 pages, 1 table. Accepted at the FMSD Workshop, ICML 2026
摘要：

【17】PRISM: Position-encoded Regressive Inverse Spectral Model for Multilayer Thin-Film Design
标题：PRism：用于多层薄膜设计的位置编码回归逆谱模型
链接：https://arxiv.org/abs/2605.26502

作者：Runtian Wang, Renhao Xue, Baige Chen, Hao Wu
备注：8 pages, 3 figures
摘要：

【18】Beyond Pairwise Preferences: Listwise Reward-Aware Alignment for Diffusion Models
标题：超越成对偏好：扩散模型的列表式奖励一致
链接：https://arxiv.org/abs/2605.26491

作者：Austin Wang, Jiaqi Han, Stefano Ermon, Yisong Yue
摘要：

【19】Aperiodic and Low-Frequency Spectral Bias in Reconstruction based EEG Foundation Models
标题：基于重建的脑电基础模型中的非周期性和低频频谱偏差
链接：https://arxiv.org/abs/2605.26434

作者：Aditya Kommineni, Emily Zhou, Kleanthis Avramidis, Simon Bock Segaard, Jeppe Roden Münster, Andreas Peter Juhl Hansen, Takfarinas Medani, Tiantian Feng, Richard Leahy, Shrikanth Narayanan
备注：18 pages, 13 figures, 3 tables
摘要：

【20】Advancing Creative Physical Intelligence in Large Multimodal Models
标题：在大型多模式模型中提升创造性身体智力
链接：https://arxiv.org/abs/2605.26396

作者：Cheng Qian, Hyeonjeong Ha, Jiayu Liu, Jeonghwan Kim, Emre Can Acikgoz, Bingxuan Li, Kunlun Zhu, Jiateng Liu, Aditi Tiwari, Zhenhailong Wang, Xiusi Chen, Mahdi Namazifar, Heng Ji
备注：51 Pages, 9 Figures, 7 Tables, Previous Work CreativityBench: arXiv:2605.02910
摘要：

【21】BioFact-MoE: Biologically Factorized Mixture of Experts for Vision-Language Prognostic Modeling in Hepatocellular Carcinoma
标题：BioFact-MoE：用于肝细胞癌视觉语言预后建模的生物因素化专家混合体
链接：https://arxiv.org/abs/2605.26376

作者：Junlin Yang, Tian Yu, Nicha C. Dvornek, Yuexi Du, Peiyu Duan, Annabella Shewarega, Lawrence H. Staib, James S. Duncan, Julius Chapiro
备注：Early accepted at MICCAI 2026
摘要：

【22】Personalized Generative Models for Contextual Debiasing
标题：上下文去偏置的个性化生成模型
链接：https://arxiv.org/abs/2605.26353

作者：Xinran Liang, Esin Tureci, Prachi Sinha, Ye Zhu, Vikram V. Ramaswamy, Olga Russakovsky
备注：CVPR 2026 Workshop on Synthetic Data for Computer Vision and Generative Models for Computer Vision. Code available at this https URL
摘要：

【23】When Correct Demonstrations Hurt: Rethinking the Role of Exemplars in In-Context Learning
标题：当正确的演示受到伤害时：重新思考示例在上下文学习中的作用
链接：https://arxiv.org/abs/2605.26350

作者：Chenghao Qiu, Chunli Peng, Yufeng Yang, Kuan-Hao Huang, Yi Zhou
摘要：

【24】A PAC-Bayesian View of Generalisation for Physics-Informed Machine Learning
标题：物理信息机器学习的PAC-Bayesian泛化观点
链接：https://arxiv.org/abs/2605.26341

作者：Thien V. Nguyen, Amaury Habrard, Benjamin Guedj
摘要：

【25】Curriculum Learning for Safety Alignment
标题：安全调整的课程学习
链接：https://arxiv.org/abs/2605.26315

作者：Sandeep Kumar, Virginia Smith, Chhavi Yadav
备注：Accepted at the ICML 2026 GlobalSouthML Workshop
摘要：

【26】Two-Parameter Flows for Learning Population Dynamics of Physical Systems
标题：学习物理系统种群动力学的双参数流
链接：https://arxiv.org/abs/2605.26285

作者：Paul Schwerdtner, Tobias Blickhan, Benjamin Peherstorfer
摘要：

【27】Co-folding model guided by structural proteomics
标题：结构蛋白质组学指导的共折叠模型
链接：https://arxiv.org/abs/2605.26192

作者：Alon Shtrikman, Nitzan Simchi, Michal Ran Shchory, Sagie Brodsky, Eran Seger, Kirill Pevzner
摘要：

【28】Modeling Dynamic Mixtures of Time-Delay Systems from Streaming Time Series
标题：从流时间序列建模延时系统的动态混合
链接：https://arxiv.org/abs/2605.26191

作者：Ren Fujiwara, Yasuko Matsubara, Yasushi Sakurai
备注：Accepted by IJCAI 2026
摘要：

【29】Gaussian Process-based learning with new MCMC-based implementation of Wishart prior on correlation matrix
标题：基于高斯过程的学习，采用基于MCMC的Wishart先验新实现相关矩阵
链接：https://arxiv.org/abs/2605.27093

作者：Kane Warrior, Dalia Chakrabarty
摘要：

【30】Signal-to-Noise Ratio and Sample Size Govern Representational Alignment in Neural Networks
标题：神经网络中的符号对齐由信噪比和样本量决定
链接：https://arxiv.org/abs/2605.26973

作者：Ali Hussaini Umar, Alessandro Laio
摘要：

【31】When Does LeJEPA Learn a World Model?
标题：LeJEPA何时学习世界模型？
链接：https://arxiv.org/abs/2605.26379

作者：David Klindt, Yann LeCun, Randall Balestriero
摘要：

【32】Deep Learning-based Algebraic Reynolds Stress Closures for RANS Simulations of Turbulent Flows
标题：基于深度学习的代数雷诺应力闭合用于湍流RANS模拟
链接：https://arxiv.org/abs/2605.26358

作者：Daniel Dehtyriov, Jonathan F. MacArt, Justin Sirignano
摘要：

【33】Learning Nonlinear Factor Models with Unknown Monotone Links from Incomplete and Noisy Data
标题：从不完整和有噪音的数据中学习具有未知单调链接的非线性因素模型
链接：https://arxiv.org/abs/2605.26271

作者：Yutong Chao, Resat Gökhan, Jalal Etesami, Ali Habibnia
摘要：

【34】Minimal surfaces, Knots, and Neural Networks
标题：最小表面、结和神经网络
链接：https://arxiv.org/abs/2605.26234

作者：Tancredi Schettini Gherardini, Marco Usula
备注：38 pages, 12 figures
摘要：

其他(45篇)

【1】MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation
标题：MusE-Autoskill：通过技能创造、记忆、管理和评估自我进化的代理人
链接：https://arxiv.org/abs/2605.27366

作者：Huawei Lin,Peng Li,Jie Song,Fuxin Jiang,Tieying Zhang
备注：30 pages, 8 figures, 13 tables, working in progress
摘要：Large language model (LLM) agents rely on reusable skills to solve complex tasks. However, existing skill creation approaches treat skills as isolated and static artifacts, limiting their reusability, reliability, and long-term improvement. We propose MUSE-Autoskill Agent (Memory-Utilizing Skill Evolution), a skill-centric agent framework that lets agents continuously improve their task-solving capability by creating, reusing, and refining skills under a unified lifecycle (creation, memory, management, evaluation, and refinement). Our framework enables agents to create skills on demand, store and reuse them across tasks, organize and select them efficiently, and evaluate them through unit tests and runtime feedback for continuous refinement. We further introduce skill-level memory that accumulates experience for each skill across tasks, enabling more effective reuse and adaptation over time. Experiments on SkillsBench provide initial evidence that lifecycle-managed skills can improve task success, efficiency, reuse, and cross-agent transfer, highlighting the importance of treating skills as long-lived, experience-aware, and testable assets.

【2】LocateAnything: Fast and High-Quality Vision-Language Grounding with Parallel Box Decoding
标题：LocateAnything：通过并行盒解码实现快速、高质量的视觉语言基础
链接：https://arxiv.org/abs/2605.27365

作者：Shihao Wang,Shilong Liu,Yuanguo Kuang,Xinyu Wei,Yangzhou Liu,Zhiqi Li,Yunze Man,Guo Chen,Andrew Tao,Guilin Liu,Jan Kautz,Lei Zhang,Zhiding Yu
摘要：Vision-language models (VLMs) commonly formulate visual grounding and detection as a coordinate-token generation problem, serializing each 2D box into multiple 1D tokens that are learned and decoded largely independently. This token-by-token decoding mismatches the coupled structure of box geometry and creates a practical inference bottleneck due to strictly sequential generation. We introduce LocateAnything, a unified generative grounding and detection framework based on Parallel Box Decoding (PBD). By decoding geometric elements such as bounding boxes and points as atomic units in a single step, LocateAnything preserves intra-box geometric coherence and unlocks substantial parallelism. We show that PBD improves both decoding throughput and localization accuracy. We further develop a scalable data engine and curate LocateAnything-Data, a large-scale dataset with more than 138 million training samples, substantially increasing data diversity for high-precision localization. Extensive evaluations show that LocateAnything advances the speed-accuracy frontier, achieving significantly higher decoding throughput while improving high-IoU localization quality across diverse benchmarks. The results highlight the complementary benefits of Parallel Box Decoding and large-scale training data in enabling efficient and precise unified visual grounding and detection.

【3】MobileMoE: Scaling On-Device Mixture of Experts
标题：MobileMoE：扩大设备上专家混合
链接：https://arxiv.org/abs/2605.27358

作者：Yanbei Chen,Hanxian Huang,Ernie Chang,Jacob Szwejbka,Digant Desai,Zechun Liu,Vikas Chandra,Raghuraman Krishnamoorthi
摘要：Mixture-of-Experts (MoE) has become the de facto architecture for hundred-billion-parameter language models, yet its advantages at sub-billion scales for on-device deployment remain largely unexplored. To close this gap, we present MobileMoE, a family of on-device MoE language models with sub-billion active parameters (0.3-0.9B active and 1.3-5.3B total) that establish a new Pareto frontier for on-device LLMs. We first formulate an on-device MoE scaling law that jointly optimizes MoE architecture under mobile memory and compute constraints, identifying an on-device sweet spot - moderate sparsity with fine-grained and shared experts - that is simultaneously memory and compute-optimal. Building on the derived architectures, we train MobileMoE with a four-stage recipe covering pre-training, mid-training, instruction fine-tuning, and quantization-aware training, all on open-source datasets. Across 14 benchmarks, MobileMoE matches or exceeds leading on-device dense LLMs with 2-4$\times$ fewer inference FLOPs, and matches or surpasses the state-of-the-art MoE OLMoE-1B-7B with up to 60% fewer parameters. To bridge the last mile to mobile deployment, we provide the first efficient MoE inference on commodity smartphones with comprehensive on-device profiling. At comparable INT4 weight memory, MobileMoE-S delivers $1.8$-$3.8\times$ faster prefill and $2.2$-$3.4\times$ faster decode than the dense baseline MobileLLM-Pro.

【4】Normal Guidance is what Attention Needs
标题：注意力需要正常的指导
链接：https://arxiv.org/abs/2605.27306

作者：Ethan Harvey,Dennis Johan Loevlie,Michael C. Hughes
摘要：We consider training classifiers for 3D medical images using only one binary label for the entire volume rather than a label for each 2D slice. In such weakly supervised settings, can we learn accurate classifiers for slice-level predictions? Attention-based multiple instance learning (MIL) can produce an attention score for every slice. Yet recent work demonstrates that a simple center-focused baseline that ignores image content can outperform attention-based and transformer-based MIL at slice-level classification of 3D brain scans. We show this baseline also outperforms existing MIL at slice-level classification of thoracic and abdominal CT scans. Motivated by this baseline, we propose Normal Guidance, a regularization technique that encourages the learned attention distribution to follow a bell-shaped curve. Across three medical imaging datasets totaling over 4 million 2D slices, we show our Normal Guidance enables attention-based and transformer-based MIL methods to deliver significantly better slice-level localization than the state-of-the-art while remaining competitive at whole-scan classification.

【5】Detectability in Diversity: Improved Canary Crafting for Privacy Auditing in One Run
标题：多样性的可检测性：一次性改进的金丝雀隐私审计工艺
链接：https://arxiv.org/abs/2605.27292

作者：Mathieu Dagréou,Aurélien Bellet
摘要：Privacy auditing aims to empirically assess privacy leakage in machine learning models using membership inference attacks (MIAs), and to derive lower bounds on differential privacy (DP) parameters. Recent one-run auditing methods address the high cost of standard approaches by relying on a single training run with multiple "canary" points whose inclusion or exclusion must be detected by the auditor. In this work, we study the problem of efficiently crafting canaries for one-run privacy auditing. Motivated by recent theoretical insights suggesting that interference between canaries contributes to weaker leakage estimates compared to multi-run methods, we propose to optimize canaries to be both highly detectable and minimally interfering. Our approach combines a greedy initialization based on influence functions with a bilevel optimization procedure that maximizes distinguishability while promoting diversity in embedding space, enabling the use of computationally efficient bilevel algorithms. Experiments show that our method achieves stronger privacy leakage estimates at a lower computational cost than existing canary crafting approaches.

【6】Causal Risk Minimization for High-Dimensional Treatments
标题：高维度治疗的因果风险最小化
链接：https://arxiv.org/abs/2605.27281

作者：Nikita Dhawan, Arnav Paruthi, Andrew Kim, Lovedeep Gondara, Jekaterina Novikova, Chris J. Maddison
备注：18 pages, 4 figures
摘要：

【7】Trust Region Q Adjoint Matching
标题：信任区域Q伴随匹配
链接：https://arxiv.org/abs/2605.27079

作者：Yonghoon Dong, Kyungmin Lee, Changyeon Kim, Jaehyuk Kim, Jinwoo Shin
摘要：

【8】FalAR: A Large-scale Speaker-Annotated European Portuguese Speech Corpus of Parliamentary Sessions
标题：FalAR：议会会议大规模议长注释的欧洲葡萄牙语演讲库
链接：https://arxiv.org/abs/2605.27062

作者：Francisco Teixeira, Carlos Carvalho, Mariana Julião, Catarina Botelho, Rubén Solera-Ureña, Sérgio Paulo, Thomas Rolland, Ben Peters, Isabel Trancoso, Alberto Abad
备注：Published in LREC2026
摘要：

【9】BhashaSetu: A Data-Centric Approach to Low-Resource Machine Translation
标题：BhashaSetu：以数据为中心的低资源机器翻译方法
链接：https://arxiv.org/abs/2605.27050

作者：Param Thakkar, Anushka Yadav, Michael Tiemann, Abhi Mehta, Akshita Bhasin, Shrinivas Khedkar
摘要：

【10】SCENT: Aligning Mass Spectra with Molecular Structure for Olfactory Perception
标题：SCENT：将光谱与分子结构对齐以实现嗅觉感知
链接：https://arxiv.org/abs/2605.27009

作者：Ziqi Zhang, Eunyeong Jin, Miguel Vasco, Farzaneh Taleb, Nona Rajabi, Alexandra Gutmann, Jonathan Williams, Antônio H. Ribeiro, Danica Kragic
摘要：

【11】Sampling Data with Chains of Forward-Backward Diffusion Steps
标题：具有向前-向后扩散步骤链的数据采样
链接：https://arxiv.org/abs/2605.27006

作者：Hyunmo Kang, Noam Itzhak Levi, Corinna Elena Wegner, Daniel J. Korchinski, Matthieu Wyart
摘要：

【12】RLVR Datasets and Where to Find Them: Tracing Data Lineage for Better Training Data
标题：WLVR数据集以及在哪里找到它们：跟踪数据谱系以获得更好的训练数据
链接：https://arxiv.org/abs/2605.26971

作者：Hsiu-Yuan Huang, Weijie Liu, Chenming Tang, Sanwoo Lee, Kai Yang, Yangkun Chen, Saiyong Yang, Yunfang Wu
备注：7 figures, 12 tables
摘要：

【13】The Strongest Teacher Is Not Always the Best Teacher: Student-Centric Answer Selection
标题：最强的老师并不总是最好的老师：以学生为中心的答案选择
链接：https://arxiv.org/abs/2605.26872

作者：Zhengyu Hu, Zheyuan Xiao, Linxin Song, Fengqing Jiang, Yutai Li, Zhengyu Chen, Zhihan Xiong, Yue Liu, Junhao Lin, Yao Su, Lijie Hu, Kaize Ding, Xiao Teng, Radha Poovendran
摘要：

【14】RAPNet: Accelerating Algebraic Multigrid with Learned Sparse Corrections
标题：RAPNet：利用习得的稀疏纠正加速代数多网格
链接：https://arxiv.org/abs/2605.26854

作者：Yali Fink, Ido Ben-Yair, Lars Ruthotto, Eran Treister
备注：Proceedings of the 43rd International Conference on Machine Learning, Seoul, South Korea Code available at this https URL
摘要：

【15】The Kalman Evolve: Closing the Gap in Kalman Filtering via Interpretable Algorithm Discovery
标题：卡尔曼进化：通过可解释算法发现缩小卡尔曼过滤的差距
链接：https://arxiv.org/abs/2605.26830

作者：Vasileios Saketos, Ming Xiao
摘要：

【16】Innovation: An Almost Characterization of Hallucination
标题：创新：幻觉的几乎特征
链接：https://arxiv.org/abs/2605.26808

作者：Nishant P. Das, Piyush Srivastava
摘要：

【17】Pretrained Approximators for Low-Thrust Trajectory Cost and Reachability
标题：预训练的低推力弹道成本和可达性的逼近器
链接：https://arxiv.org/abs/2605.26790

作者：Zhong Zhang, Giacomo Acciarini, Dario Izzo, Hexi Baoyin, Francesco Topputo
备注：Submitted to the Journal of Guidance, Navigation and Control
摘要：

【18】Time Series Causal Discovery via Context-Conditioned and Causality-Augmented Pretraining
标题：通过上下文条件和因果性增强预训练发现时间序列因果关系
链接：https://arxiv.org/abs/2605.26759

作者：Biao Ouyang, Tengxue Zhang, Zhihao Zhuang, Yang Shu, Chenjuan Guo, Bin Yang
备注：Submitted to the 40th Conference on Neural Information Processing Systems (NeurIPS 2026). 27 pages
摘要：

【19】MTL-FNO: A Lightweight Multi-Task Fourier Neural Operator for Sparse Field Reconstruction
标题：MTL-FNO：一种用于稀疏场重建的轻量级多任务傅里叶神经运算器
链接：https://arxiv.org/abs/2605.26718

作者：Siyu Ye, Shihang Li, Zhiqiang Gong, Benrong Zhang, Weien Zhou, Yiyong Huang, Wen Yao
摘要：

【20】Image Feature Fusion-based Federated Client Unlearning (FCU)
标题：基于图像特征融合的联合客户端取消学习（FCU）
链接：https://arxiv.org/abs/2605.26715

作者：Hangyi Shen, Yizhi Pan, Tiansuo Li, Weiqi Jiang, Guanqun Sun
摘要：

【21】Self-Improvement Imitation with Biologically Guided Search for Protein Design Under Oracle Budgets
标题：在Oracle预算下通过生物引导搜索蛋白质设计的自我改进模仿
链接：https://arxiv.org/abs/2605.26690

作者：Ashima Khanna, Dominik Grimm
摘要：

【22】Sample Complexity of Policy Gradient for Log-Growth Control
标题：逻辑增长控制政策梯度的样本复杂性
链接：https://arxiv.org/abs/2605.26640

作者：Qiuhua Pan, Yukai Shen, Liwei Zhang, Cailian Chen, Xinping Guan
备注：43 pages, 4 figures, 2 tables; includes supplementary material
摘要：

【23】Spend Your Rollouts Where It Counts: Rollout Allocation for Group-Based RL Post-Training
标题：在重要的地方度过您的推出：基于小组的RL训练后的推出分配
链接：https://arxiv.org/abs/2605.26606

作者：Woojeong Kim, Ziyi Yang, Jing Nathan Yan, Jialu Liu
摘要：

【24】On the Error-Correcting Effects of Stochasticity in Discrete Diffusion
标题：离散扩散中随机性的误差修正效应
链接：https://arxiv.org/abs/2605.26582

作者：William Yuan, Sungwon Jeong, Amirali Aghazadeh
摘要：

【25】Linear and Neural Dueling Bandits with Delayed Feedback
标题：具有延迟反馈的线性和神经决斗强盗
链接：https://arxiv.org/abs/2605.26554

作者：Xiangyi Wang, Pingchen Lu, Jie Mao, Mingze Kong, Zhi Hong, Zhiyong Wang, Zhongxiang Dai
摘要：

【26】Recursive Flow Matching
标题：循环流匹配
链接：https://arxiv.org/abs/2605.26535

作者：Jiahe Huang, Sihan Xu, Sharvaree Vadgama, Rose Yu
备注：Project page: this https URL
摘要：

【27】The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence
标题：MiniMax-M2系列：迷你激活释放Max现实世界智能
链接：https://arxiv.org/abs/2605.26494

作者：MiniMax: Aili Chen, Aonian Li, Baichuan Zhou, Bangwei Gong, Binyang Jiang, Boji Dan, Changqing Yu, Chao Wang, Cheng Ma, Cheng Zhong, Cheng Zhu, Chengjun Xiao, Chengyi Yang, Chengyu Du, Chenyang Zhang, Chi Zhang, Chuangyi Huang, Chunhao Zhang, Chunhui Du, Chunyu Zhao, Congchao Guo, Da Chen, Deming Ding, Dianjun Sun, Dongyu Zhang, Enhui Yang, Fei Yu, Guang Zheng, Guodong Zheng, Guohong Li, Haichao Zhu, Haigang Zhou, Haimo Zhang, Han Ding, Hao Zhang, Haohai Sun, Haolin Lyu, Haonan Lu, Haoyu Wang, Huajie Shi, Huiyang Li, Jiacheng Chen, Jian Zhang, Jiaqi Zhuang, Jiaren Cai, Jiaxin Pan, Jiayao Li, Jiayuan Song, Jichuan Zhang, Jie Wang, Jihao Gu, Jin Zhu, Jingwei Dong, Jingyang Li, Jingyu Zhang, Jingze Zhuang, Jinhao Tian, Jinli Liu, Jinyi Hu, Jun Tao, Jun Zhang, Junbin Ruan, Junhao Xu, Junjie Yan, Junteng Liu, Junxian He, Kang Xu, Ke Ji, Ke Yang, Kecheng Xiao, Keyu Duan, Keyu Li, Le Han, Letian Ruan, Li Yuan, Lianfei Yu, Liheng Feng, Lijie Mo, Lin Li, Lingye Bao, Lingyu Yang, Lingyuan Zhou, Loki, Lu Chen, Lunbin Ceng, Ming Li, Ming Zhong, Mingliang Tao, Mingyuan Chi, Mujie Lin, Nan Hu, Ningxin Chen, Peiyin Zhu, Peng Gao, Pengcheng Gao, Pengfei Li, Penglin Li, Pengyu Zhao, Qibin Ren
备注：Technical Report. 35 pages, 10 figures, 4 tables
摘要：

【28】Efficient On-policy Visual-RL via Stochastic Decoupled Policy Gradient
标题：通过随机脱钩策略梯度实现高效的政策上可视化RL
链接：https://arxiv.org/abs/2605.26478

作者：Haoxiang You, Yilang Liu, Davis Zong, Qian Wang, Teeratham Vitchutripop, Qi Wang, Daniel Rakita, Ian Abraham
摘要：

【29】FM-fMRI: Event Conditioned Flow Matching for Rest-to-Task fMRI Time-Series Synthesis
标题：FM-fMRI：用于休息到任务fMRI时间序列合成的事件条件流匹配
链接：https://arxiv.org/abs/2605.26423

作者：Peiyu Duan, Jiyao Wang, Nicha C. Dvornek, Junlin Yang, Ziqi Gao, Lawrence H. Staib, James S. Duncan
备注：MICCAI 2026 Early Accepted
摘要：

【30】Function-Valued Causal Influence in Nonlinear Time Series
标题：非线性时间序列中的函数值因果影响
链接：https://arxiv.org/abs/2605.26408

作者：Valentina V. Kuskova, Dmitry Zaytsev, Michael Coppedge
备注：26 pages, 6 tables, 8 figures
摘要：

【31】Balancing Plasticity and Stability with Fast and Slow Successor Features
标题：平衡可塑性和稳定性与快速和缓慢的连续功能
链接：https://arxiv.org/abs/2605.26357

作者：Raymond Chua, Doina Precup, Blake Richards
备注：Main Paper: 9 pages, 9 figures. Accepted at The International Conference on Machine Learning (ICML) 2026
摘要：

【32】Reparametrizing Shampoo and SOAP for Subspace Basis Updates and BFloat16 Storage
标题：为子空间基础更新和BFloat 16存储重新参数化Shampoo和Soap
链接：https://arxiv.org/abs/2605.26327

作者：Alan Milligan, Zikun Xu, Simon Lacoste-Julien, Felix Dangel, Wu Lin
备注：Preprint, working in progress
摘要：

【33】Semigroup Consistency as a Diagnostic for Learned Physics Simulators
标题：作为学习物理模拟器的诊断方法
链接：https://arxiv.org/abs/2605.26324

作者：Lennon J. Shikhman
备注：10 pages, 3 figures, 3 tables. Accepted to the AI4Physics Workshop at the 43rd International Conference on Machine Learning
摘要：

【34】Quantized Keys Steal Attention: Bias Correction for KV-Cache Compression in Video Diffusion
标题：量化关键字窃取注意力：视频扩散中KV-Cache压缩的偏差校正
链接：https://arxiv.org/abs/2605.26266

作者：Tuna Tuncer, Felix Becker, Thomas Pfeil
备注：Variants of this manuscript were accepted to the ICML 2026 workshops SCALE and F2S
摘要：

【35】Unified Neural Scaling Laws
标题：统一神经缩放定律
链接：https://arxiv.org/abs/2605.26248

作者：Ethan Caballero, Priyank Jaini, David Krueger, Irina Rish
摘要：

【36】From Privacy to Generalization: Linear Max-Information Bounds for DP-SGD
标题：从隐私到概括：DP-Singapore的线性最大信息界
链接：https://arxiv.org/abs/2605.26222

作者：Christoph H. Lampert, Hossein Zakerinia
备注：22 pages
摘要：

【37】ATOM: Instantiating Budget-Controllable Multi-Agent Collaboration via Nucleus-Electron Hierarchy
标题：ATOM：通过核电子体系实例化预算可控的多智能体协作
链接：https://arxiv.org/abs/2605.26178

作者：Xinkui Zhao, Sai Liu, Yifan Zhang, Qingyu Ma, Zewen Lin, Naibo Wang, Guanjie Cheng, Chang Liu, Yueshen Xu
摘要：

【38】LearnedCache: An eBPF-Integrated Perceptron-Based Eviction Policy for the Linux Page Cache
标题：Learnedache：针对Linux页面缓存的ePFA集成的基于感知器的驱逐策略
链接：https://arxiv.org/abs/2605.26168

作者：Zejia Qi
备注：11 pages, 12 figures, 4 listings. Policies and harnesses: this https URL . Model and visualizations: this https URL
摘要：

【39】Neural Bayesian Sequential Routing
标题：神经Bayesian顺序路由
链接：https://arxiv.org/abs/2605.26147

作者：Yongchao Huang
备注：71 pages
摘要：

【40】Constrained Bayesian Experimental Design via Online Planning
标题：通过在线规划的约束Bayesian实验设计
链接：https://arxiv.org/abs/2605.26990

作者：Yujia Guo, Daolang Huang, Xinyu Zhang, Sammie Katt, Samuel Kaski, Ayush Bharti
备注：24 pages, 9 figures. Accepted at the Forty-Third International Conference on Machine Learning (ICML 2026)
摘要：

【41】Particle-Lund Multimodality in Jet Taggers
标题：Jet Taggers中的Partic-Lund多模式
链接：https://arxiv.org/abs/2605.26821

作者：Loukas Gouskos, Benedikt Maier
摘要：

【42】Neural Autoregressive Control Variates for the Quantum Monte Carlo Sign Problem
标题：量子蒙特卡罗符号问题的神经自回归控制变量
链接：https://arxiv.org/abs/2605.26814

作者：Bei Qiao, Lei Wang
备注：18 pages, 9 figures
摘要：

【43】Proper Calibeating
标题：正确的校准
链接：https://arxiv.org/abs/2605.26703

作者：Dean P. Foster, Sergiu Hart
摘要：

【44】CART Random Forests as Sequential Allocation over Random Opportunity Sets: A Stochastic-Control Theory of Ensemble Risk
标题：CART随机森林作为随机机会集的顺序分配：集群风险的随机控制理论
链接：https://arxiv.org/abs/2605.26675

作者：Tianxing Mei, Yingying Fan, Mingming Leng, Jinchi Lv
备注：69 pages, 1 figure
摘要：

【45】Rapid online deep artifact suppression for real-time spiral bSSFP CMR with blipped-CAIPI simultaneous multi-slice imaging at 1.5 T
标题：快速在线深度伪影抑制实时螺旋bSSFP RCM，采用1.5 T的翻转CAIPI同时多切片成像
链接：https://arxiv.org/abs/2605.26127

作者：Julius Åkesson, Iulius Dragonu, Einar Heiberg, Tina Yao, Rebecca Baker, Ruta Virsinskaite, Daniel Knight, Vivek Muthurangu, Jennifer Steeden
摘要：

机器翻译由腾讯交互翻译提供，仅供参考

点击“阅读原文”获取带摘要的学术速递