Learning the Mechanism of Catastrophic Forgetting: A Perspective from Gradient Similarity

2026-01-29

AI Summary

This study addresses catastrophic forgetting in large language models during continual knowledge injection. The authors provide the first theoretical explanation from the perspective of gradient similarity, revealing that strong negative gradient similarity is the root cause of forgetting. Based on this insight, they categorize neurons into conflicting and synergistic types. They propose Synergistic Neuron Learning (CNL), a method that updates only synergistic neurons while freezing conflicting ones, thereby theoretically eliminating forgetting. Extensive experiments across five large language models, four datasets, and four optimizers demonstrate that CNL achieves zero forgetting in in-distribution scenarios and reduces forgetting by 59.1%–81.7% in out-of-distribution settings.

Technology Category

• Natural Language Processing: (Large) Language Models
• Machine Learning: Life-Long and Continual Learning
• Search and Optimization: Learning to Search

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Catastrophic forgetting during knowledge injection severely undermines the continual learning capability of large language models (LLMs). Although existing methods attempt to mitigate this issue, they often lack a foundational theoretical explanation. We establish a gradient-based theoretical framework to explain catastrophic forgetting. We first prove that strongly negative gradient similarity is a fundamental cause of forgetting. We then use gradient similarity to identify two types of neurons: conflicting neurons that induce forgetting and account for 50%-75% of neurons, and collaborative neurons that mitigate forgetting and account for 25%-50%. Based on this analysis, we propose a knowledge injection method, Collaborative Neural Learning (CNL). By freezing conflicting neurons and updating only collaborative neurons, CNL theoretically eliminates catastrophic forgetting under an infinitesimal learning rate eta and an exactly known mastered set. Experiments on five LLMs, four datasets, and four optimizers show that CNL achieves zero forgetting in in-set settings and reduces forgetting by 59.1%-81.7% in out-of-set settings.

Problem

• catastrophic forgetting
• continual learning
• large language models
• knowledge injection

Innovation

• catastrophic forgetting
• gradient similarity
• collaborative neurons
• conflicting neurons
• continual learning