|
Jiaxi Liu | 刘家熙 I am Jiaxi Liu, a first-year Master's student in the Department of Computer Science at Shanghai Jiao Tong University. Since 2024, I have been a member of ReThinkLab, advised by Prof. Junchi Yan(严骏驰). I received my Bachelor of Engineering from the Department of Artificial Intelligence at Nanjing University in 2025. During my undergraduate studies, I had the privilege of working with Prof. Yang Yu in the LAMDA Group. I was also fortunate to serve as a visiting researcher at Yale University, collaborating with Prof. Zhuoran Yang. Additionally, I previously worked as an algorithm engineering intern at Tencent, Guangzhou. My research interests lie in Large Language Models (LLMs), self-evolving systems, agentic RL, and machine learning on graphs. My research goal is to build highly autonomous AI agents by synergizing generative modeling with principled decision-making frameworks, ultimately pushing the boundaries of self-evolving artificial intelligence. |
日高悬风拂面留下丝丝温暖，
别辜负眼前季节！
|
Publications (* indicates equal contribution) |
|
TodoEvolve: Learning to Architect Agent Planning Systems
Jiaxi Liu*, Guibin Zhang*, Yanzuo Jiang, Zihan Zhang, Heng Chang, Zhenfei Yin, Qibing Ren, Junchi Yan ICML 2026 CCF A PDF | Code We propose TodoEvolve, a meta-planning paradigm that autonomously synthesizes and dynamically revises task-specific LLM agent planning architectures via Impedance-Guided Preference Optimization (IGPO), consistently outperforming carefully engineered static planning frameworks. |
|
Design Linear Constrained Neural Layers with Implicit Convex Optimization
Junchi Yan, Jiaxi Liu, Liangliang Shi, Fangyuan Zhou, Wenzheng Pan, Zhongteng Gui, Yihui Tu ICML 2026 CCF A We introduce LinConLayer, a plug-in differentiable layer for enforcing linear constraints in neural networks through implicit convex optimization. The framework unifies classic constraint layers from an optimization perspective and instantiates two efficient variants, BLCLayer and GLCLayer, enabling stable, hard-constrained prediction for linear programming, partial graph matching, and portfolio allocation. |
|
Towards Practical Large-scale Dynamical Heterogeneous Graph Embedding: Cold-start Resilient Recommendation
Mabiao Long*, Jiaxi Liu*, Yufeng Li, Hao Xiong, Junchi Yan, Kefan Wang, Yi Cao, Jiandong Ding TKDE CCF A We introduce a practical, two-stage framework combining HetSGFormer for scalable static learning and Incremental Locally Linear Embedding (ILLE) for real-time, GPU-free updates, addressing scalability, data freshness, and cold-start challenges in billion-scale recommendation systems. |
Honors and Awards |
|
Education |
|
Miscellaneous |
|
|
The source code is stolen from Jon Barron. Thanks for his sharing! 🙏🏻 |