Jinze Li

Hi, I’m Jinze Li (李金泽), a PhD candidate at The University of Hong Kong.

My broader vision is to bring LLMs from research advances to dependable real-world systems, pursued along two complementary fronts: improving inference efficiency for cost-effective deployment, and empowering long-horizon agents through persistent memory and self-evolution.

Current research interests include:

Speculative Decoding — accelerating LLM inference by drafting and verifying tokens efficiently.
Agent Memory & Self-evolution — enabling agents to handle long-horizon interactions and improve themselves over time.
LLM Post-training — SFT, RL, and decoding-time techniques to align and enhance model behavior.

I have been fortunate to do research internships at SenseTime, Huawei, ByteDance, AMD, and Ant Group (the Ant Star Program).

education

2023 – 2027 · Ph.D., 香港大学 / The University of Hong Kong, Hong Kong SAR, China
2020 – 2023 · M.Sc., 中国科学院大学 / University of Chinese Academy of Sciences, Beijing, China
2016 – 2020 · B.Eng., 大连理工大学 / Dalian University of Technology, Dalian, China

news

May 15, 2026	Our paper OCR-Memory: Optical Context Retrieval for Long-Horizon Agent Memory was accepted to ACL 2026 (Main Conference)!
Jan 22, 2026	Our paper Training-Free Loosely Speculative Decoding (FLy) was accepted to ICLR 2026!
May 01, 2025	Our paper Gumiho: A Hybrid Architecture to Prioritize Early Tokens in Speculative Decoding was accepted to ICML 2025!
Sep 01, 2023	Started my PhD journey at The University of Hong Kong!

selected publications

Preprint

Beyond the Target: From Imitation to Collaboration in Speculative Decoding

Jinze Li, Yixing Xu, Guanchen Li, and 7 more authors

2026

preprint

Bib

@misc{li2026spad,
  title = {Beyond the Target: From Imitation to Collaboration in Speculative Decoding},
  author = {Li, Jinze and Xu, Yixing and Li, Guanchen and Xu, Jinfeng and Yang, Shuo and Zhang, Yang and Yin, Xuanwu and Li, Dong and Ngai, Edith C.H. and Barsoum, Emad},
  year = {2026},
  note = {preprint},
}

ACL

OCR-Memory: Optical Context Retrieval for Long-Horizon Agent Memory

Jinze Li, Yang Zhang, Xin Yang, and 5 more authors

In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Main Conference Long Papers), 2026

arXiv Bib

@inproceedings{li2026ocrmemory,
  title = {OCR-Memory: Optical Context Retrieval for Long-Horizon Agent Memory},
  author = {Li, Jinze and Zhang, Yang and Yang, Xin and Qu, Jiayi and Xu, Jinfeng and Yang, Shuo and Ding, Junhua and Ngai, Edith C.H.},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Main Conference Long Papers)},
  year = {2026},
}

ICLR

Training-Free Loosely Speculative Decoding: Accepting Semantically Correct Drafts Beyond Exact Match

Jinze Li, Yixing Xu, Guanchen Li, and 6 more authors

In The Fourteenth International Conference on Learning Representations (ICLR), 2026

arXiv Bib

@inproceedings{li2026fly,
  title = {Training-Free Loosely Speculative Decoding: Accepting Semantically Correct Drafts Beyond Exact Match},
  author = {Li, Jinze and Xu, Yixing and Li, Guanchen and Yang, Shuo and Xu, Jinfeng and Yin, Xuanwu and Li, Dong and Ngai, Edith C.H. and Barsoum, Emad},
  booktitle = {The Fourteenth International Conference on Learning Representations (ICLR)},
  year = {2026},
}

ICML

Gumiho: A Hybrid Architecture to Prioritize Early Tokens in Speculative Decoding

Jinze Li, Yixing Xu, Haiduo Huang, and 4 more authors

In Proceedings of the 42nd International Conference on Machine Learning (ICML), 2025

arXiv Bib

@inproceedings{li2025gumiho,
  title = {Gumiho: A Hybrid Architecture to Prioritize Early Tokens in Speculative Decoding},
  author = {Li, Jinze and Xu, Yixing and Huang, Haiduo and Yin, Xuanwu and Li, Dong and Ngai, Edith C.H. and Barsoum, Emad},
  booktitle = {Proceedings of the 42nd International Conference on Machine Learning (ICML)},
  year = {2025},
}