Princeton University

Topology-Preserving Neural Operator Learning via Hodge Decomposition

Beyond GRPO and On-Policy Distillation: An Empirical Sparse-to-Dense Reward Principle for Language-Model Post-Training