GeRe: Towards Efficient Anti-Forgetting in Continual Learning of LLM via General Samples Replay

TL;DR

This paper introduces GeRe, a simple and effective replay-based framework using general samples and activation state constraints to prevent catastrophic forgetting in continual learning of large language models, improving retention and performance.

Contribution

The paper proposes GeRe, a novel replay framework utilizing pretraining texts and activation state constraints, demonstrating its effectiveness in continual LLM learning.

Findings

TM loss improves performance and robustness

Small fixed set of samples suffices for anti-forgetting

Activation state constraints enhance replay stability

Abstract

The continual learning capability of large language models (LLMs) is crucial for advancing artificial general intelligence. However, continual fine-tuning LLMs across various domains often suffers from catastrophic forgetting, characterized by: 1) significant forgetting of their general capabilities, and 2) sharp performance declines in previously learned tasks. To simultaneously address both issues in a simple yet stable manner, we propose General Sample Replay (GeRe), a framework that use usual pretraining texts for efficient anti-forgetting. Beyond revisiting the most prevalent replay-based practices under GeRe, we further leverage neural states to introduce a enhanced activation states constrained optimization method using threshold-based margin (TM) loss, which maintains activation state consistency during replay learning. We are the first to validate that a small, fixed set of pre-collected general replay samples is sufficient to resolve both concerns--retaining general capabilities while promoting overall performance across sequential tasks. Indeed, the former can inherently facilitate the latter. Through controlled experiments, we systematically compare TM with different replay strategies under the GeRe framework, including vanilla label fitting, logit imitation via KL divergence and feature imitation via L1/L2 losses. Results demonstrate that TM consistently improves performance and exhibits better robustness. Our work paves the way for efficient replay of LLMs for the future. Our code and data are available at https://github.com/Qznan/GeRe.