GR-SAP: Generative Replay for Safety Alignment Preservation during Fine-Tuning

2026-03-10

AI Summary

This work addresses the vulnerability of large language models (LLMs) to safety degradation during fine-tuning, particularly when original safety alignment data is unavailable. To mitigate this issue, the authors propose GR-SAP, a novel framework that introduces generative replay into safety alignment preservation. GR-SAP leverages the LLM itself to generate domain-specific safety-aligned data, which is then used in conjunction with downstream task data to jointly optimize both task performance and safety objectives during fine-tuning. Experimental results across multiple models and tasks demonstrate that GR-SAP effectively alleviates safety degradation induced by fine-tuning while maintaining strong downstream task performance.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Recent studies show that the safety alignment of large language models (LLMs) can be easily compromised even by seemingly non-adversarial fine-tuning. To preserve safety alignment during fine-tuning, a widely used strategy is to jointly optimize safety and task objectives by mixing in the original alignment data, which is typically inaccessible even for open-weight LLMs. Inspired by generative replay in continual learning, we propose Generative Replay for Safety Alignment Preservation (GR-SAP), a unified framework that synthesizes domain-specific alignment data from LLMs and integrate them during downstream adaption to preserve safety alignment. Theoretical and empirical analyses demonstrate this synthetic data serves as a reliable proxy for the original alignment data. Experiments across various models and downstream tasks show that GR-SAP substantially mitigates fine-tuning-induced safety degradation while maintaining comparable downstream performance. Our code is available at https://github.com/chili-lab/gr-sap.

Problem

• safety alignment
• fine-tuning
• large language models
• alignment preservation
• generative replay

Innovation

• Generative Replay
• Safety Alignment
• Fine-tuning
• Synthetic Data
• Large Language Models