Proximal Supervised Fine-Tuning

TL;DR

Proximal Supervised Fine-Tuning (PSFT) introduces a trust-region inspired objective to improve the stability and generalization of foundation models during fine-tuning, especially out-of-domain.

Contribution

This paper adapts trust-region methods from reinforcement learning to supervised fine-tuning, enhancing stability and out-of-domain generalization of foundation models.

Findings

PSFT matches in-domain SFT performance

Outperforms SFT in out-of-domain generalization

Remains stable under prolonged training without entropy collapse

Abstract

Supervised fine-tuning (SFT) of foundation models often leads to poor generalization, where prior capabilities deteriorate after tuning on new tasks or domains. Inspired by trust-region policy optimization (TRPO) and proximal policy optimization (PPO) in reinforcement learning (RL), we propose Proximal SFT (PSFT). This fine-tuning objective incorporates the benefits of trust-region, effectively constraining policy drift during SFT while maintaining competitive tuning. By viewing SFT as a special case of policy gradient methods with constant positive advantages, we derive PSFT that stabilizes optimization and leads to generalization, while leaving room for further optimization in subsequent post-training stages. Experiments across mathematical and human-value domains show that PSFT matches SFT in-domain, outperforms it in out-of-domain generalization, remains stable under prolonged training without causing entropy collapse, and provides a stronger foundation for the subsequent optimization.