Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

TREK: Distill to Explore, Reinforce to Refine

Source

arXiv

arxiv.org/abs/2607.05339v1

Read the source

Source

  • Source name: arXiv
  • Source kind: paper
  • Authors / organization: Yuanda Xu, Zhengze Zhou, Kayhan Behdin, Jelena Markovic-Voronov, Hejian Sang, Xiaomin Li, Wenhui Zhu, Xinchen Du, Aida Rahmattalabi, Ran He, Sen Na, Zhipeng Wang, Alborz Geramifard
  • Venue: arXiv: cs.LG
  • Published: 2026-07-06
  • Collected: 2026-07-07
  • URL: http://arxiv.org/abs/2607.05339v1
  • PDF: https://arxiv.org/pdf/2607.05339v1
  • DOI: UNKNOWN

Plain-Language Brief

TREK is a procedure that uses distillation for exploration support expansion to improve policy optimization on tasks where the current policy stalls on hard prompts. It identifies low-performing prompts, generates verified candidate solutions, pulls those verified modes into the student's support, and then refines the policy using GRPO. TREK improves performance on mathematical reasoning and agentic tasks compared to standard policy optimization.

What the Document Says

  • TREK is a staged procedure that uses distillation for exploration support expansion.
  • It identifies prompts where the unaided student has very low pass rate.
  • Queries a proposal source to produce verified candidate solutions.
  • Keeps the top-$r$ proposals ranked by current student likelihood.
  • Applies a short forward-KL phase to pull those verified modes into the student's support.
  • Returns to standard on-policy GRPO refinement.
  • Improves Qwen3 models across all tested scales on AIME 2024 and AIME 2025.
  • For Qwen3-8B, improves AIME 2025 from 36.9 to 40.3 and AIME 2024 from 47.9 to 51.1 (avg@16).
  • The self-context variant reaches 38.5 and 49.6 without an external teacher.
  • Raises ALFWorld success rate from 75.8 to 82.8 and ScienceWorld success rate from 12.5 to 26.7.
  • On the hardest task types, TREK achieves high success rates early in training.

Key Concepts Mentioned

  • TREK
  • Group Relative Policy Optimization (GRPO)
  • distillation
  • exploration support expansion
  • verified candidate solutions
  • forward-KL phase
  • AIME 2024
  • AIME 2025
  • Qwen3
  • ALFWorld
  • ScienceWorld

Claims Made by the Document

  • TREK improves Qwen3 models on AIME 2024 and AIME 2025.
  • For Qwen3-8B, TREK improves AIME 2025 and AIME 2024 scores significantly.
  • The self-context variant of TREK also shows improvement without an external teacher.
  • TREK raises success rates on ALFWorld and ScienceWorld.
  • On the hardest task types, TREK achieves high success rates early in training.

Evidence Used

  • Improves Qwen3 models across all tested scales on AIME 2024 and AIME 2025
  • For Qwen3-8B, it improves AIME 2025 from 36.9 to 40.3 and AIME 2024 from 47.9 to 51.1 (avg@16)
  • The self-context variant reaches 38.5 and 49.6 without an external teacher
  • Raises ALFWorld success rate from 75.8 to 82.8 and ScienceWorld success rate from 12.5 to 26.7
  • On the hardest task types, TREK achieves high success rates early in training

Methods / Approach

  • Identifying low-performing prompts
  • Generating verified candidate solutions
  • Keeping top proposals ranked by current student likelihood
  • Applying a short forward-KL phase
  • Returning to standard on-policy GRPO refinement

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

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 inference-time context, and it can efficiently identify which hard-prompt samples are most worth consolidating even when teacher internals are unavailable. TREK first identifies prompts where the unaided student has very low pass rate, queries a proposal source to produce verified candidate solutions, keeps the top-$r$ proposals ranked by current student likelihood, applies a short forward-KL phase to pull those verified modes into the student's support, and then returns to standard on-policy GRPO refinement. On mathematical reasoning, TREK with DeepSeek-V4 proposals improves Qwen3 models across all tested scales on AIME 2024 and AIME 2025; for Qwen3-8B, it improves AIME 2025 from 36.9 to 40.3 and AIME 2024 from 47.9 to 51.1 (avg@16), while the self-context variant reaches 38.5 and 49.6 without an external teacher. On agentic tasks, TREK raises ALFWorld success rate from 75.8 to 82.8 and ScienceWorld success rate from 12.5 to 26.7; notably, on the hardest task types, TREK achieves high success rates early in training while unaided GRPO requires substantially more optimization steps to reach comparable levels.