Pinned Loading

1. crl-ordering crl-ordering Public

   Experiment code for Goal Retrieval versus Action-Value Ordering: Contrastive Critics under Best-of-K Selection.

   Python

2. entropy-collapse-rlvr entropy-collapse-rlvr Public

   SFT overtraining collapses output entropy, squashing GRPO gradient signal; we derive the exact failure condition, show standard checkpoint selection inverts quality ranking, and propose a 2-stage d…

   Python

3. pcrc pcrc Public

   Per-Context Reverse Curricula for Sparse-Reward Reinforcement Learning

   Python