arXivDaily arXiv每日学术速递 周一至周五更新

高校专区

University of Science and Technology of China(中国科学技术大学)

2026-04-15 至 2026-04-15 共收录 7
2604.12806 2026-04-15 cs.LG

Interpretable Relational Inference with LLM-Guided Symbolic Dynamics Modeling

基于LLM引导的符号动力学建模的可解释关系推断

Xiaoxiao Liang, Juyuan Zhang, Liming Pan, Linyuan Lü

机构 * School of Cyber Science and Technology, University of Science and Technology of China(中国科学技术大学计算机科学与技术学院)

AI总结 本文提出COSINE框架,通过联合发现交互图和稀疏符号动力学,实现对复杂系统中潜在相互作用结构的可解释推断,实验表明其在合成系统和大规模真实世界数据中的鲁棒性。

Comments Submitted to conference

详情

展开后加载摘要…

URL PDF HTML 收藏
2503.05167 2026-04-15 cs.LG

FMASH: Advancing Traditional Chinese Medicine Formula Recommendation with Efficient Fusion of Multiscale Associations of Symptoms and Herbs

FMASH:通过高效融合症状与药材的多尺度关联来推进传统中医方剂推荐

Xinhan Zheng, Xueting Wang, Ruotai Li, Huyu Wu, Haopeng Jin, Yehan Yang, Guodong Shan

机构 * Beijing University of Posts and Telecommunications(北京邮电大学) University of Science and Technology of China(中国科学技术大学) University of Chinese Academy of Sciences(中国科学院大学)

AI总结 本文提出FMASH框架,通过融合症状与药材的多尺度关联,提升中医方剂推荐效果,实验表明其在两个数据集上均优于现有最佳模型。

详情

展开后加载摘要…

URL PDF HTML 收藏
2604.11841 2026-04-15 cs.LG cs.AI

Polynomial Expansion Rank Adaptation: Enhancing Low-Rank Fine-Tuning with High-Order Interactions

多项式展开秩适应:通过高阶交互增强低秩微调

Wenhao Zhang, Lin Mu, Li Ni, Peiquan Jin, Yiwen Zhang

机构 * Anhui University(安徽大学) University of Science and Technology of China(中国科学技术大学)

AI总结 本文提出PERA方法,通过在低秩因子空间中引入结构化多项式展开,提升低秩微调的表达能力,实验证明其在多种基准上表现优异。

Comments Accepted by ACL 2026 findings

详情

展开后加载摘要…

URL PDF HTML 收藏
2604.11554 2026-04-15 cs.CL

Relax: An Asynchronous Reinforcement Learning Engine for Omni-Modal Post-Training at Scale

Relax:一种用于大规模多模态后训练的异步强化学习引擎

Liujie Zhang, Benzhe Ning, Rui Yang, Xiaoyan Yu, Jiaxing Li, Lumeng Wu, Jia Liu, Minghao Li, Weihang Chen, Weiqi Hu, Lei Zhang

机构 * AI Platform, Xiaohongshu Inc(小红书AI平台) The University of Hong Kong(香港大学) University of Science and Technology of China(中国科学技术大学)

AI总结 Relax通过三个协同设计的架构层解决多模态数据流、大规模鲁棒性和延迟-吞吐量平衡问题,实现比veRL快1.20倍的端到端加速,并在多模态强化学习中表现出稳定收敛性。

Comments 17 pages, 22 figures

详情

展开后加载摘要…

URL PDF HTML 收藏
2601.05524 2026-04-15 cs.CL

Double: Breaking the Acceleration Limit via Double Retrieval Speculative Parallelism

Double: 通过双检索推测并行突破加速限制

Yuhao Shen, Tianyu Liu, Junyi Shen, Jinyang Wu, Quan Kong, Li Huan, Cong Wang

机构 * Zhejiang University(浙江大学) University of Science and Technology of China(中国科学技术大学) National University of Singapore(新加坡国立大学) Tsinghua University(清华大学)

AI总结 Double通过双检索推测并行机制解决推测解码的理论加速上限和计算浪费问题,实现5.3倍的LLaMA3.3-70B加速,优于需大量训练的EAGLE-3。

Comments Accepted by ACL2026 Main

详情

展开后加载摘要…

URL PDF HTML 收藏
2512.06443 2026-04-15 cs.DC cs.AI

Vec-LUT: Vector Table Lookup for Parallel Ultra-Low-Bit LLM Inference on Edge Devices

Vec-LUT:用于边缘设备上并行超低比特LLM推理的向量表查找

Xiangyu Li, Chengyu Yin, Weijun Wang, Jianyu Wei, Ting Cao, Yunxin Liu

机构 * Institute for AI Industry Research (AIR), Tsinghua University(人工智能产业研究院(AIR),清华大学) Beijing Jiaotong University(北京交通大学) University of Science and Technology of China(中国科学技术大学)

AI总结 本文提出Vec-LUT,通过统一表查找和缓存感知流查找技术,提升边缘设备上超低比特LLM推理效率,实验显示性能提升达4.2倍。

Comments MobiSys 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2506.01979 2026-04-15 cs.DC cs.AI

SpecBranch: Speculative Decoding via Hybrid Drafting and Rollback-Aware Branch Parallelism

SpecBranch:通过混合草稿和回滚感知分支并行性进行推测解码

Yuhao Shen, Junyi Shen, Quan Kong, Tianyu Liu, Yao Lu, Cong Wang

机构 * Zhejiang University(浙江大学) National University of Singapore(新加坡国立大学) University of Science and Technology of China(中国科学技术大学)

AI总结 SpecBranch通过混合草稿和回滚感知分支并行性,提升大语言模型推理速度,减少回滚令牌,实现实际应用。

Comments The paper has been accepted by ICLR2026

详情

展开后加载摘要…

URL PDF HTML 收藏