On the Non-decoupling of Supervised Fine-tuning and Reinforcement Learning in Post-training
This addresses a foundational issue for researchers and practitioners in AI, as it challenges the common practice of alternating SFT and RL in post-training, indicating it is incremental by providing theoretical and empirical evidence against decoupling.
The paper tackles the problem of whether supervised fine-tuning (SFT) and reinforcement learning (RL) can be decoupled in post-training of large language models, proving that decoupling is impossible in either order and confirming this with experiments on Qwen3-0.6B, showing degradation in performance.
Post-training of large language models routinely interleaves supervised fine-tuning (SFT) with reinforcement learning (RL). These two methods have different objectives: SFT minimizes the cross-entropy loss between model outputs and expert responses, while RL maximizes reward signals derived from human preferences or rule-based verifiers. Modern reasoning models have widely adopted the practice of alternating SFT and RL training. However, there is no theoretical account of whether they can be decoupled. We prove that decoupling is impossible in either order: (1) SFT-then-RL coupling: RL increases SFT loss under SFT optimality and (2) RL-then-SFT coupling: SFT lowers the reward achieved by RL. Experiments on Qwen3-0.6B confirm the predicted degradation, verifying that SFT and RL cannot be separated without loss of prior performance in the post-training