Tong Chen

My research focuses on how language models learn from experience:

• Self-Improving Agents: Memory & Continual Learning. I build agents that turn past experience into reusable knowledge: meta-optimizing how agents save and load memory, training note-taking models with reinforcement learning across streams of tasks, compiling context directly into model weights, and learning retrievers that surface the right knowledge at the right granularity.
  NoteTaker (preprint) · Generative Adapter (ICLR'25) · Revela (ICLR'26) · Dense X Retrieval (EMNLP'24)
• RL Post-Training & Alignment. I study how to post-train language models to be factual and trustworthy without losing their other capabilities: reinforcement learning with binary retrieval-augmented rewards that reduces hallucination, preference optimization against verbatim regurgitation of pre-training data, and benchmarks for literal and non-literal copying in long-form generation.
  Binary RAR (ICML'26 Spotlight) · ParaPO (COLM'25) · CopyBench (EMNLP'24)

• CV
• Scholar
• GitHub
• Twitter
• LinkedIn
• Email