Academic Homepage
Xingkun Yin尹星锟Ph.D. candidate博士生
About Me关于我
I am a second-year Ph.D. student at NICE Lab, Department of Electrical and Computer Engineering, The University of Hong Kong, fortunate to be advised by Professor Hongyang Du.
My research interests include LLM architecture, LLM memory, experience-driven model evolution, and video generation. I am especially interested in the intersection of model capability and system efficiency.
What excites me most is the idea of models that keep evolving, reasoning, and learning from experience after training — more like a human brain, with learning continuing throughout their lifetime, thus achieving Artificial General Intelligence (AGI) and recursive self-improvement (RSI).
I am always open to collaboration. If any of the directions above overlaps with your work, please do get in touch.
我是香港大学电机与计算机工程系 NICE Lab 的二年级博士研究生，很荣幸由 杜泓阳教授指导。
我的研究兴趣包括大模型架构、大模型记忆、经验驱动的模型演化和视频生成。我尤其关注模型能力与系统效率之间的交叉问题。
最让我兴奋的研究课题，是让模型在训练结束后依然能够持续进化、思考并从经验中学习，就像人脑一样，让学习贯穿模型的整个生命周期，从而实现通用人工智能（AGI）与递归自我改进（RSI）。
非常欢迎合作。如果上面的方向与您的研究有交集，欢迎随时联系我。
Research Interests研究方向
LLM Memory大模型记忆
Memory representation, retrieval and realignment for models that operate in dynamic, changing environments.面向动态环境的记忆表示、检索与对齐。
LLM Architecture大模型架构
Efficient architectures and inference-time acceleration for large language and diffusion transformers.面向大语言模型与扩散 Transformer 的高效架构与推理加速。
Experience-Driven Evolution经验驱动的模型演化
Post-deployment and experience scaling: how models keep improving after training through autonomous interaction and shared experience.部署后演化与经验扩展：模型如何通过自主交互与经验共享，在训练之后持续变强。
Video Generation视频生成
Cross-request reuse and compatibility-guided acceleration for text-to-video diffusion transformers.文本到视频扩散 Transformer 的跨请求复用与兼容性引导加速。
Selected Publications代表性论文
arXiv2026 Carnator: Fast Text-to-Video Generation with Generation-Native Compatibility-Guided Cross-Request Reuse
A cross-request acceleration framework that uses generation-native compatibility evidence from early diffusion states, reaching up to 2.17× end-to-end speedup.一个跨请求加速框架：从扩散模型的早期状态中读出「生成原生」的兼容性证据，端到端最高加速 2.17 倍。
View Paper查看论文
arXiv2026 CoDeR: Local Constraint-Compatible Retrieval Beyond Semantic Similarity
A local constraint-compatible dense retrieval method that separates topical relevance from constraint compatibility for constraint-sensitive queries.一种面向约束敏感查询的局部约束相容稠密检索方法，把「主题相关」与「约束相容」分开建模。
View Paper查看论文
NeurIPS2026 GLOVE: Global Verifier for LLM Memory-Environment Realignment
A framework that introduces a new design dimension for LLM memory systems by establishing a relative notion of truth.一个通过建立「相对真理」这一概念，为大模型记忆系统引入新设计维度的框架。
View Paper查看论文
See the complete list of publications, with abstracts and links. 完整论文列表（含摘要与链接）请见论文发表页面。
All Publications全部论文News新闻动态
- 2026.09 Paper论文 GLOVE: Global Verifier for LLM Memory-Environment Realignment was accepted to NeurIPS 2026!论文GLOVE: Global Verifier for LLM Memory-Environment Realignment被 NeurIPS 2026 接收！
- 2025.11 Paper论文 Ubiquitous Intelligence Via Wireless Network-Driven LLMs Evolution was accepted to npj Wireless Technology!论文Ubiquitous Intelligence Via Wireless Network-Driven LLMs Evolution被 npj Wireless Technology 接收！
- 2025.08 Milestone里程碑 Started my Ph.D. at The University of Hong Kong! Expected to graduate at July 2029.入学香港大学攻读博士学位！预计 2029 年 7 月毕业。
Academic Service学术服务
Conference Reviewer
- ICML, NeurIPS, ICLR, AAAI, EMNLP
会议审稿
- ICML、NeurIPS、ICLR、AAAI、EMNLP
Contact联系方式
Email: yinxingkun [at] connect [dot] hku [dot] hk
Address: Department of Electrical and Computer Engineering, The University of Hong Kong, Pokfulam, Hong Kong SAR
As always, GLHF! (Good luck, have fun!)
邮箱：yinxingkun [at] connect [dot] hku [dot] hk
地址：香港特别行政区 薄扶林 香港大学 电机与计算机工程系
As always, GLHF! （Good luck, have fun!）