Meta-Optimization of Agent Learning, Skills & Memory for Continual Learning
Tong Chen
I am a final-year Ph.D. student in Computer Science at the University of Washington, advised by Luke Zettlemoyer and Hannaneh Hajishirzi. I have interned at Meta FAIR, AI2, Microsoft Research, and Tencent AI Lab. Before UW, I received my B.Eng. in Computer Science from the Yao Class at Tsinghua University.
My research focuses on language models and agents that improve over time.
-
Memory & Long-Horizon Agents.
I develop harness optimization and harness distillation that teach models to use the file system as memory;
train models with reinforcement learning to abstract reusable knowledge into memory across streams of tasks;
continually pretrain a language model to learn at test time, compressing context tokens into a LoRA adapter
in a single forward pass; and strengthen retrieval-augmented generation with better datastores and retrievers.
Meta-Optimization (preprint) · Learning to Learn Experiential Knowledge (preprint) · Generative Adapter (ICLR'25) · Dense X Retrieval (EMNLP'24) · Revela (ICLR'26) -
Alignment & Open-Ended Generation.
I mitigate reward hacking in RL for long-form hallucination reduction without hurting generation quality;
post-train models to regulate verbatim copying of memorized text in open-ended generation; and evaluate
non-literal copying of memorized content in the context of copyright.
Binary RAR (ICML'26 Spotlight) · ParaPO (COLM'25) · CopyBench (EMNLP'24)
I am on the industry job market.
#News
#Selected Publications
* equal contribution. Full list on Google Scholar.
Learning to Learn Experiential Knowledge
