Jie “JW” Wu

Off-policy evaluation, a classic problem in reinforcement learning, is how companies estimate a new policy's impact from logged data instead of a live A/B test. We asked whether LLM agents can keep that code tuned on their own, and measured how often they make things better, worse, or simply break.

Good engineers ask questions when instructions don't add up. We built a benchmark to test whether code models do the same, and to make it something the field can actually measure.

AI coding tools don't ask questions when your request is vague. They guess, silently. Our new paper makes those guesses visible.

A reflection on leaving a decade in industry for fundamental ideas and intellectual freedom.