English

Offline Exploration-Aware Fine-Tuning for Long-Chain Mathematical Reasoning

Machine Learning 2026-03-18 v1 Computation and Language

Abstract

Through encouraging self-exploration, reinforcement learning from verifiable rewards (RLVR) has significantly advanced the mathematical reasoning capabilities of large language models. As the starting point for RLVR, the capacity of supervised fine-tuning (SFT) to memorize new chain-of-thought trajectories provides a crucial initialization that shapes the subsequent exploration landscape. However, existing research primarily focuses on facilitating exploration during RLVR training, leaving exploration-aware SFT under-explored. To bridge this gap, we propose Offline eXploration-Aware (OXA) fine-tuning. Specifically, OXA optimizes two objectives: promoting low-confidence verified teacher-distillation data to internalize previously uncaptured reasoning patterns, and suppressing high-confidence incorrect self-distillation data to redistribute probability mass of incorrect patterns toward potentially correct candidates. Experimental results across 6 benchmarks show that OXA consistently improves mathematical reasoning performance, especially achieving an average gain of +6+6 Pass@1 and +5+5 Pass@kk points compared to conventional SFT on the Qwen2.5-1.5B-Math. Crucially, OXA elevates initial policy entropy, and performance gains persist throughout extensive RLVR training, demonstrating the long-term value of OXA.

Keywords

Cite

@article{arxiv.2603.16206,
  title  = {Offline Exploration-Aware Fine-Tuning for Long-Chain Mathematical Reasoning},
  author = {Yongyu Mu and Jiali Zeng and Fandong Meng and JingBo Zhu and Tong Xiao},
  journal= {arXiv preprint arXiv:2603.16206},
  year   = {2026}
}

Comments

Working in process

R2 v1 2026-07-01T11:23:43.249Z