Please note: This PhD seminar will take place in TBD.
Cong Wei, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Wenhu Chen
Aligning visual generative models with human preferences remains challenging because existing reward models often require large-scale preference datasets and reduce complex human judgments to unexplained scalar scores. In this seminar, I will present two complementary frameworks that address these limitations.
RewardClaw is an agentic framework for test-time preference learning. Using as few as 100 demonstrations, it iteratively develops a library of evaluation tools and skills, which agents compose into reasoning chains for assessing instruction-guided image edits. The library is automatically refined by analyzing prediction successes and failures, without requiring additional human annotation.
RationalRewards instead represents preferences through explicit, multi-dimensional critiques. Trained using Preference-Anchored Rationalization, it provides structured and semantically grounded feedback that can serve as a reward for reinforcement learning. Its critiques can also be used in a Generate–Critique–Refine loop to improve prompts and generation quality at test time without updating model parameters.
Together, these approaches move reward models beyond passive scalar evaluators toward data-efficient and interpretable optimization interfaces for visual generation.