Ph.D. Candidate in Artificial Intelligence
LLM Recommendation · Agent Memory · Causal & Uplift Modeling
I am a direct-entry Ph.D. candidate at the Gaoling School of Artificial Intelligence, Renmin University of China, currently visiting the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI). I build learning systems that understand user intent, remember long-term preferences, and make reliable decisions under distribution shift.
- My research sits at the intersection of LLM post-training, personalized agents, recommender systems, and causal machine learning.
- I have published 13 papers, including 5 first-author papers, with 440+ Google Scholar citations excluding the survey paper.
- I have worked on recommendation and decision systems at Tencent AI4Fin, DiDi, and Ant Group, spanning financial recommendation, agent memory, uplift modeling, and large-scale ranking.
- I am a core contributor to RecBole and an early core member of RecAgent / YuLan-Rec.
- I enjoy turning research ideas into reproducible systems, from data and training workflows to evaluation and online inference.
| Direction | Current focus |
|---|---|
| LLM Recommendation | Small-model post-training, reasoning distillation, constrained generation, and production inference for recommendation |
| Agent Memory | Short- and long-term memory, user-profile evolution, retrieval, reflection, consolidation, and memory evaluation |
| Causal & Uplift Modeling | Continuous-treatment effect estimation, invariant learning, policy optimization, and uncertainty-aware decisioning |
| Reliable Recommendation | Fairness, robustness under distribution shift, user simulation, and reproducible recommendation workflows |
Core contributor to the unified recommendation framework nominated for the CIKM 2022 Best Paper Award. I contributed to its Data → Dataset → DataLoader → Interaction workflow, fairness-aware recommendation support, evaluation interfaces, and community maintenance through 30+ resolved issues and 10+ pull requests.
Early core member responsible for Agent Memory in an LLM-based user behavior simulation platform. I designed a hierarchical memory network covering observation, short-term intent, long-term preference, retrieval, reflection, and forgetting to support consistent behavior across recommendation and social scenarios.
- Personalized Agents: long-horizon user behavior simulation and memory-grounded decision making.
- Responsible Recommendation: robust and distributional fairness under environment shifts.
- Causal Learning: uplift modeling, counterfactual regression, and active causal discovery.
- LLM Training Dynamics: sample-order effects, hidden-confounder imputation, and efficient post-training.
Selected venues include ICML, EMNLP, KDD, RecSys, ACM TOIS, SSDBM, and IEEE HPCC. See the complete publication list on Google Scholar.
Building personalized AI systems that can learn, remember, and make better decisions.