Assistant Professor, Texas A&M University - College Station
2 papers at NeurIPS 2025
We proposed a framework for reinforcing large reasoning models with discriminative constrained optimization , grounded in the principle that increasing the scores of positive answers while decreasing those of negative ones.
A theoretical bound showing that SSL approximates supervised contrastive learning as a function of #classes in the training dataset. Studied the connection between the representation geometry and their downstream transferability.
