ℹ️ Short Bio
Hi, I’m Yuanzhe, a second year master student at University of Califronia, San Diego 🔱.
I have the great honor of being collaborating with Prof. Yaoqing Yang from CS@Dartmouth College, Julian McAuley from CSE@UC San Diego, Zhiting Hu HDSI@UC San Diego, and Dr. Ren Pu from AI & Learning Systems@LBNL, Yu Wang from Memory Team@xAI, Zexue He from HAI@Stanford.
My current research is focused on
-
Understanding the mechanisms, training dynamics, and generalization of LLMs and SciML models via mathematical analysis. Building upon this theoretical foundation, We design advanced optimization algorithms to make LLM compression and training more efficient. Previous works include FARMS (ICML 2025) for layer-wise pruning, Model Balancing (EMNLP 2024 Oral) for low-resource fine-tuning, and ongoing comprehensive analysis of multi-regime dynamics in SciML models preparing for ICML 2026.
-
Enhancing Memory and Reasoning in LLMs and Agents, specifically by enabling models to process long-term history and achieve advanced reasoning capabilities through post-training. Previous works include M+ (ICML 2025) for long-term information retention, K2-Think (Tech Report) for large-scale reasoning, MIRIX (Open-source framework, 3K+ 🌟) for multi-agent memory systems, MemoryAgentBench (Under Review), 190+ 🌟 for comprehensive evaluation, and Mem-alpha (Under Review) for RL-based memory management.
My research leverages mathematical insights into LLMs to develop efficient algorithms, while simultaneously unlocking advanced memory and reasoning capabilities in LLMs and Agents. Moving forward, I aim to empower LLM Agents with superior memory and reasoning abilities by innovating across efficient optimization methods, effective frameworks, mechanisms, and interactive environment designs.
I am actively seeking for 26 Fall PhD Positions, industrial research internship after M.S graduation (about six months), and research collobration opportunities . Feel free to reach out!
📖 Educations
|
University of California, San Diego (UCSD) M.S. in Computer Science and Engineering |
2024.09 - 2026.03 (Expected) |
|
Huazhong University of Science and Technology (HUST) B.S. in Artificial Intelligence, Innovation Experimental Honor Class, Qiming School GPA: 3.91/4.0 |
2020.09 - 2024.06 |
⚙️ Research Project
📖 Mathematical Analysis on LLMs and SciML Models
# denotes equal contribution
[3] Multi-Regime Patterns in SciML Models: How Optimization Methods Address Distinct Failure Modes
Yuanzhe Hu (Co-First Author)
Manuscript in preparation (Targeting ICML 2026)
🤔 Enhancing Memory and Reasoning in LLMs and Agents
MIRIX: Multi-Agent Memory System for LLM-Based Agents
My Contribution: Designed the framework for MIRIX’s Evaluation, project maintenance and bug solving.
Open-Source Project, 3K+ 🌟
[4] Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions
{Yuanzhe Hu#, Yu Wang#}, Julian McAuley
Under Review (8/6/6/4) / ICML 2025 LCFM Workshop
Short Summary: MemoryAgentBench is a new benchmark designed to comprehensively evaluate memory agents in LLMs.
[5] M+: Extending MemoryLLM with Scalable Long-Term Memory
Yu Wang, Dmitry Krotov, Yuanzhe Hu, Yifan Gao, Wangchunshu Zhou, Julian McAuley, Dan Gutfreund, Rogerio Feris, Zexue He
ICML 2025
Short Summary: M+ enhances long-term information retention in LLMs by integrating a retriever-based long-term memory mechanism.
[6] Mem-$\alpha$: Learning Memory Construction via Reinforcement Learning
Yu Wang, Ryuichi Takanobu, Zhiqi Liang, Yuzhen Mao, Yuanzhe Hu, Julian McAuley, Xiaojian Wu
Under Review
Short Summary: Mem-alpha, a reinforcement learning framework, enhances memory management in LLMs through interaction and feedback.
[7] K2-Think: A Parameter-Efficient Reasoning System
Zhoujun Cheng, Richard Fan, Shibo Hao, Taylor W. Killian, Haonan Li, Suqi Sun, Hector Ren, Alexander Moreno, Daqian Zhang, Tianjun Zhong, Yuxin Xiong, Yuanzhe Hu, Yutao Xie, Xudong Han, Yuqi Wang, Varad Pimpalkhute, Yonghao Zhuang, Aaryamonvikram Singh, Xuezhi Liang, Anze Xie, Jianshu She, Desai Fan, Chengqian Gao, Liqun Ma, Mikhail Yurochkin, John Maggs, Xuezhe Ma, Guowei He, Zhiting Hu, Zhengzhong Liu, Eric P. Xing
MBZUAI IFM / LLM 360 Tech Report
Short Summary: K2-Think is a parameter-efficient reasoning system based on a 32B model.
🔥 News
- 2025.09: 😁 Excited to share that our recent work “K2-Think: A Parameter-Efficient Reasoning System”.
- 2025.07: 😁 We open-sourced the MemoryAgentBench. Thanks for the great help from Yu Wang!
- 2025.05: 🎉🎉 Two papers are accepted by ICML 2025 as Poster! See you at Vancouver.
- 2024.09: 🎉🎉 Excited to share that our work “Model Balancing Helps Low-data Training and Fine-tuning” is accepted by EMNLP 2024 as Oral Presentation!
- 2024.06: 😁 I graduated from HUST!
- 2024.06: 😄 I created my account on OpenReview!
Last Update: 11/2025