TLDR: A new framework called RECALL addresses catastrophic forgetting in large language models by aligning their internal representations. This data-free method uses hierarchical model merging based on layer-wise similarities, allowing LLMs to continually learn new tasks without losing previously acquired knowledge, demonstrating superior performance in knowledge retention and generalization across various NLP tasks.
Large Language Models (LLMs) have become incredibly powerful, driving advancements in areas like chatbots, text generation, and complex reasoning. These models are typically trained on vast amounts of data and then fine-tuned for specific tasks. However, a significant challenge they face is ‘catastrophic forgetting.’ This occurs when an LLM learns a new task, and in doing so, overwrites or forgets knowledge it previously acquired from older tasks. This problem is particularly critical as LLMs are increasingly used in dynamic, multi-domain environments where they need to continually adapt without losing past expertise.
Traditionally, approaches to combat catastrophic forgetting fall into two main categories: data-based and model-based methods. Data-based methods involve revisiting old data samples while learning new tasks, which helps retain information. However, this can be impractical due to storage limitations or privacy concerns. Model-based methods try to constrain how the model updates its parameters or isolate task-specific knowledge, but they often struggle to maintain performance across diverse tasks and can increase model complexity.
A new research paper introduces a novel framework called RECALL: REpresentation-aligned Catastrophic-forgetting ALLeviation via Hierarchical Model Merging. This innovative approach tackles catastrophic forgetting without needing access to historical data. The core idea behind RECALL is that the internal ‘representations’ within an LLM – essentially how the model processes and encodes information at different layers – can serve as reliable indicators of its learned knowledge. By understanding and aligning these representations, RECALL can effectively merge knowledge from different models.
The researchers observed that different layers of an LLM specialize in different types of information. Lower layers tend to handle more general features, while deeper layers become more specialized for specific tasks. When models are fine-tuned on different tasks, their deeper layer representations tend to diverge significantly. RECALL leverages this insight by computing similarities between these layer-wise hidden representations across different models. It then uses these similarities to guide an adaptive, hierarchical parameter fusion process.
Here’s a simplified breakdown of how RECALL works: First, it identifies a small set of ‘typical samples’ from the new task’s dataset using a clustering algorithm. These samples act as representatives of the new knowledge. Then, it extracts the internal representations of these typical samples from the newly fine-tuned model and all other expert models (models trained on previous tasks). Based on these representations, RECALL calculates a similarity score between models for each layer. These similarity scores are then used to determine adaptive weights for merging the parameters of different models, layer by layer. This hierarchical merging ensures that domain-general features in shallow layers are preserved, while allowing for task-specific adaptation in deeper layers.
The benefits of RECALL are substantial. It offers a data-free solution, addressing privacy and storage concerns. It’s also task-agnostic, meaning it doesn’t require explicit task labels or boundaries. This design enables seamless fusion of knowledge from multiple domains and provides strong resistance to catastrophic forgetting. The framework can merge multiple expert models fine-tuned on different domains, enhancing multi-domain capabilities through weighted representation alignment.
Extensive experiments were conducted across five diverse NLP tasks, including text classification, question answering, and machine translation, and in various continual learning scenarios. RECALL consistently outperformed existing baselines in both retaining previously learned knowledge and generalizing to new tasks. For instance, in scenarios merging a single fine-tuned model with a base model, RECALL achieved the highest average performance and significantly better generalization to unseen tasks. In more complex scenarios involving the merging of multiple task-specific expert models, RECALL again demonstrated superior overall performance, even without the inclusion of the base model.
Furthermore, in sequential fine-tuning experiments, where models learn tasks one after another, RECALL maintained remarkably stable performance on earlier tasks, effectively mitigating the dramatic performance decline seen in other methods due to catastrophic forgetting. This highlights RECALL’s suitability for dynamic learning environments where models need to continuously evolve.
Also Read:
- Enhancing LLM Privacy: A New Approach to Federated Unlearning
- Smart Adaptation: How CKA-RL Improves AI’s Continuous Learning
While RECALL presents a powerful solution, the researchers acknowledge some limitations. It assumes access to multiple fine-tuned models and relies on the quality of selected typical samples. Currently, it is tailored for models with identical architectures and aligned tokenizers, and further research is needed for scaling to a larger number of tasks or integrating with training-time regularization techniques for lifelong learning. Nevertheless, RECALL represents a significant step forward in enabling Large Language Models to learn continually and adapt without forgetting their past knowledge. You can read the full research paper here.
