一作ACL 2026Poster· Natural Language Processing
Rethinking Expert Trajectory Utilization in LLM Post-training for Mathematical Reasoning
Bowen Ding · Yuhan Chen · Jiayang Lyu · Jiyao Yuan · Qi Zhu · Shuangshuang Tian · Dantong Zhu · Futing Wang · Heyuan Deng · Fei Mi · Lifeng Shang · Tao Lin
摘要
Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) dominate the post-training landscape for mathematical reasoning, yet differ fundamentally in their reliance on expert trajectories. To understand the optimal way to harness these trajectories for maximizing performance, we propose the P...