TREK: Distill to Explore, Reinforce to Refine
Group Relative Policy Optimization (GRPO) is effective when the current policy already samples useful reasoning trajectories, but it stalls on hard prompts whose correct solution modes lie outside the student's on-policy support. We propose TREK (Teacher-Routed Exploration via Forward KL), a simple staged procedure that uses distillation not for imitation but for exploration support expansion. A key advantage of TREK is its generality: because it only consumes verified output trajectories, it can use an external black-box teacher, a white-box teacher, or the same model given additional inferen
Lineage graph
Paper → model → repo connections mined from source citations (Tier-1 exact match).
Why these links exist
- Linked via arxiv authorYuanda Xu →
TREK: Distill to Explore, Reinforce to Refine
- Linked via arxiv authorZhengze Zhou →
TREK: Distill to Explore, Reinforce to Refine
- Linked via arxiv authorKayhan Behdin →
TREK: Distill to Explore, Reinforce to Refine
- Linked via arxiv authorJelena Markovic-Voronov →
TREK: Distill to Explore, Reinforce to Refine
- Linked via arxiv authorHejian Sang →
TREK: Distill to Explore, Reinforce to Refine
- Linked via arxiv authorXiaomin Li →
TREK: Distill to Explore, Reinforce to Refine
- Linked via arxiv authorWenhui Zhu →
TREK: Distill to Explore, Reinforce to Refine
- Linked via arxiv authorXinchen Du →
TREK: Distill to Explore, Reinforce to Refine
- Linked via arxiv authorAida Rahmattalabi →
TREK: Distill to Explore, Reinforce to Refine
- Linked via arxiv authorRan He →
TREK: Distill to Explore, Reinforce to Refine
- Linked via arxiv authorSen Na →
TREK: Distill to Explore, Reinforce to Refine
- Linked via arxiv authorZhipeng Wang →
TREK: Distill to Explore, Reinforce to Refine
- Linked via arxiv authorAlborz Geramifard →
TREK: Distill to Explore, Reinforce to Refine