RECALL: REpresentation-aligned Catastrophic-forgetting ALLeviation via Hierarchical Model Merging

TL;DR

RECALL introduces a hierarchical model merging framework that leverages internal representations in large language models to prevent catastrophic forgetting during continual learning, without needing historical data.

Contribution

It proposes a novel, data-free, representation-aware merging method that aligns knowledge across models through hierarchical parameter fusion, improving continual learning in LLMs.

Findings

Outperforms baselines in knowledge retention and generalization.

Effectively preserves domain-general features while adapting to new tasks.

Demonstrates scalability across multiple NLP tasks and scenarios.

Abstract

We unveil that internal representations in large language models (LLMs) serve as reliable proxies of learned knowledge, and propose RECALL, a novel representation-aware model merging framework for continual learning without access to historical data. RECALL computes inter-model similarity from layer-wise hidden representations over clustered typical samples, and performs adaptive, hierarchical parameter fusion to align knowledge across models. This design enables the preservation of domain-general features in shallow layers while allowing task-specific adaptation in deeper layers. Unlike prior methods that require task labels or incur performance trade-offs, RECALL achieves seamless multi-domain integration and strong resistance to catastrophic forgetting. Extensive experiments across five NLP tasks and multiple continual learning scenarios show that RECALL outperforms baselines in both knowledge retention and generalization, providing a scalable and data-free solution for evolving LLMs.