作者位次 #5/7ACL 2026Findings· Findings of ACL
Bridging SFT and RL : Dynamic Policy Optimization for Robust Reasoning
Taojie Zhu · Dongyang Xu · Ding Zou · Sen Zhao · Qiaobo Hao · Zhiguo Yang · Yonghong He
摘要
Post-training paradigms for Large Language Models (LLMs), primarily Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), face a fundamental dilemma: SFT provides stability (low variance) but suffers from high fitting bias, while RL enables exploration (low bias) but grapples with high gr...