TREK Method Enhances GRPO Exploration via Teacher-Routed Distillation
July 5, 2026
TREK addresses Group Relative Policy Optimization stalling by using distillation to expand the student model's exploration support. The staged procedure identifies hard prompts where the student lacks correct solution modes and uses verified trajectories from black-box or white-box teachers to guide reinforcement learning.
HOW THIS AFFECTS YOU
●
builderThis provides a method to use external APIs as teachers to refine your local models' reasoning capabilities.
●
researcherYou can improve RL training stability for reasoning models on tasks outside existing on-policy support.