Hi, I’m Jiang Wu

I am interested in personalization, self-evolving agents, and recursive self-improvement (RSI), particularly how models can use feedback and experience to improve their behavior. I currently focus on agent memory: comparing repeated attempts at a task, separating grounded episodic evidence from transferable strategies, and checking when a stored lesson applies to a new problem.

I have been working on LLM post-training and personalized large language models, including preference optimization, lightweight adaptation, and preference-guided decoding. A central question is how to learn from user feedback while keeping individual preferences distinct from explicit task and style instructions, so that personalization remains efficient and controllable. Some of my RL work also explores how to train multi-turn agents by jointly using turn-level and trajectory-level rewards under safety constraints.

I graduated from the University of Southern California in 2020.