About Me
I am an incoming Ph.D. student at HKUST and currently a third-year M.Eng. student in Data Science at Zhejiang University, advised by Prof. Wei Chen at ZJUVAI. I received my B.Eng. in Computer Science from Jinan University.
Previously, I was a research assistant at THUNLP, Tsinghua University, working with Prof. Zhiyuan Liu and Prof. Chen Qian. I also worked as a research intern at Alibaba Group and a Talent Program intern at ByteDance.
I hope to build machine-native AI that can learn and develop its own representations, communication protocols, coordination structures, and mechanisms for self-improvement.
I am seeking research internship opportunities and welcome research collaborations. Contact me · Book a chat.
Research Interests
- Multi-agent collaboration: latent-space communication and reasoning, learned coordination.
- Agent evolution and evaluation: recursive self-improvement, interaction-driven learning, autonomous evaluation, and user simulation.
- Foundation models: model training, post-training, and novel architectures.
Selected Publications
ACL 2026Association for Computational Linguistics (ACL), 2026
ACL 2025Association for Computational Linguistics (ACL), 2025
Findings 2025Findings of the Association for Computational Linguistics (ACL Findings), 2025
* denotes equal contribution.
More publications (6)
- Learning to Evolve: A Self-Improving Framework for Multi-Agent Systems via Textual Parameter Graph OptimizationFindings of the Association for Computational Linguistics (ACL Findings), 2026
- MemCoRL: Alternating Co-Optimization of Memory Retrieval and Utilization via Collaborative Reinforcement LearningAssociation for Computational Linguistics (ACL), 2026
- International Conference on Learning Representations (ICLR), 2025
- Conference on Neural Information Processing Systems (NeurIPS), 2024
Posts
- Engineering Self-Improving Agents: From Context to Harness Evolution
- Interlat: Enabling Agents to Communicate Entirely in Latent Space
- Faithful LLM User Simulation: Closing the Gap Between Synthetic Users and Real People
More posts (6)
- Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents
- Latent Space: From Continuous Reasoning to Representation Analysis
- Training Large Language Models to Reason in a Continuous Latent Space
- Agents from Google: A Systems View of Models, Tools, and Orchestration
- Long-Horizon Agents Need Long-Lived State, Not Just Long Context
- Optimizing Outcomes, Not Replies: Multi-Turn Post-Training for LLM Agents
Talks
Experience
- Built the team's first data-flywheel rollout system for multi-turn, task-oriented dialogue, supporting downstream multi-turn dialogue RL and agentic RL.
- Designed a self-play simulation and evaluation environment with high-fidelity, persona-conditioned user-simulation agents for rollout generation and evaluation signals.
- Conducted post-training and preference alignment of large-scale LLMs (e.g., Qwen3.5-122B/397B) using SFT and RLHF/DPO, alongside autonomous agent-driven evaluation of dialogue policies.
- Proposed Interlat (code), a latent-space communication framework that replaces natural-language message passing with temporally aligned hidden-state exchange.
- Designed a token-latent curriculum and learned compression scheme, achieving up to 24x communication speedup at comparable task performance.
- Gained hands-on experience with distributed training on 256+ GPUs; co-authored Online-PVLM, MemCoRL, and Learning to Evolve.
- Led EvoPatient (code), an open-source doctor-patient coevolution system for virtual standardized patients, accepted to ACL 2025.
- Drove the project end-to-end, including research formulation, system implementation, medical-case curation, expert evaluation, and paper writing.
- Improved patient-answer ability from 0.763 to 0.860 after 200 training cases.
- Worked with Prof. Zhiyuan Liu and Prof. Chen Qian on ChatDev-based multi-agent collaboration, studying role organization, information exchange, and coordination.
- Led Croto (code), designing cross-team orchestration with parallel teams, hierarchy partitioning, and greedy aggregation for multi-agent software development.
- Contributed to MACNet (code) and iAgent (code); related OpenBMB/ChatDev projects have accumulated over 33k GitHub stars.
Awards and Honors
- 2026: Outstanding Graduate Student.
- 2025: National Scholarship (awarded to ∼2% of master’s students nationwide annually); Outstanding Graduate Student.
- 2024: First-Class Scholarship for Outstanding Graduates.
- Undergraduate: Three undergraduate scholarships.
Academic Service
- Reviewer: NeurIPS 2025, ICLR 2026.
Miscellaneous
Beyond research, I enjoy soccer, fencing (national amateur épée champion), snowboarding, billiards, ballroom dancing (I occasionally teach classes), piano, photography, and physics.