Mitigating Forgetting in LLM Supervised Fine-Tuning and Preference Learning

Heshan Devaka Fernando, Han Shen, Parikshit Ram, Yi Zhou, Horst Samulowitz, Nathalie Baracaldo, Tianyi Chen

Keywords: Optimization Trade-off, Large Language Models (LLMs), Supervised Fine-tuning (SFT), Reinforcement Learning from Human Feedback (RLHF)

Abstract: Post-training of pre-trained LLMs, which typically consists of the supervised fine-tuning (SFT) stage and the preference learning (RLHF or DPO) stage, is crucial to effective and safe LLM applications. The widely adopted approach in post-training popular open-source LLMs is to sequentially perform SFT and RLHF/DPO. However, sequential training is sub-optimal in terms of SFT and RLHF/DPO trade-off: the LLM gradually forgets about the first stage's training when undergoing the second stage's training. We theoretically prove the sub-optimality of sequential post-training. Furthermore, we propose a practical joint post-training framework that has theoretical convergence guarantees and empirically outperforms sequential post-training framework, while having similar computational cost.

Primary Area: optimization

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4034