This paper proposes Post-Completion Learning (PCL), a novel learning framework that utilizes the sequence space after the model output is completed, to overcome the limitation of existing language model training that terminates at the terminal token (). PCL enhances inference and self-evaluation capabilities by generating self-evaluations and reward predictions even after the model completes its training. Furthermore, it maintains efficiency by stopping the inference process at completion. Using white-box reinforcement learning techniques, the model evaluates outputs according to reward rules and supervises the scores by aligning them with the reward function. This approach combines dual-track SFT and RL training, which simultaneously optimize inference and evaluation capabilities, to achieve multi-objective hybrid optimization. Experimental results on various datasets and models demonstrate consistent performance improvements over existing SFT and RL methods.
