Shuyan Zhou

Even the best LLMs struggle to maintain correct intermediate states when solving university-level STEM problems, often taking more steps than necessary and accumulating errors along the way.

Stop guessing when humans want to take over: modeling user intervention styles in web agents boosts their usefulness by 26.5%.

Agents can rapidly adapt to individual and evolving user preferences by explicitly memorizing and acting on both pre- and post-action feedback.