HuatuoGPT-3: RL-Only Domain Adaptation from Base Models

OnePO enables domain adaptation using only reinforcement learning by treating teacher outputs as transient guidance to prevent gradient starvation and distribution anchoring. This one-stage method bypasses the traditional supervised fine-tuning phase, reducing optimization complexity while maintaining exploration diversity during the transition from general-purpose to domain-specific expertise.