Recolora: Continual Fine-Tuning for LLMs Without Catastrophic Forgetting

Researchers have introduced Recolora, a continual fine-tuning approach designed to prevent catastrophic forgetting in large language models. The method allows LLMs to learn new domains or tasks sequentially without degrading performance on earlier ones, a persistent problem in model updating and deployment.

The technique involves a combination of architectural modifications and optimization strategies. Specifically, Recolora uses low-rank adaptation (LoRA) modules that are dynamically expanded and consolidated, alongside a replay buffer of representative data from previous tasks. This dual approach balances plasticity for new knowledge with stability for old knowledge.

In experiments on code generation, mathematical reasoning, and general language understanding benchmarks, Recolora-trained models retained 95% of original performance on prior tasks while achieving competitive results on new ones. The method also showed strong scaling properties, maintaining effectiveness as the number of sequential tasks increased.

This work is particularly relevant for scenarios where LLMs must be continuously updated with new data, such as enterprise knowledge bases, specialized coding assistants, or domain-specific chatbots. Recolora offers a practical solution for keeping models current without requiring full retraining or sacrificing past expertise.