Replay to Remember: Retaining Domain Knowledge in Streaming Language Models

2025-04-24

AI Summary

To address severe catastrophic forgetting and stringent computational/data constraints in continual domain adaptation of large language models (LLMs) under streaming settings, this paper proposes a lightweight and efficient method that synergistically optimizes minimal-sample replay—requiring only a handful of historical examples—with low-rank adaptation (LoRA). It is the first work to validate this approach under realistic streaming conditions. Comprehensive evaluation—including perplexity, semantic similarity, and GPT-assisted human assessment—demonstrates its effectiveness in significantly mitigating forgetting and partially recovering domain-specific knowledge. Experiments across three specialized domains—healthcare, genomics, and law—show that the method achieves both stability and adaptability with negligible computational overhead and storage cost. This work establishes a practical, resource-efficient paradigm for real-time, continual domain adaptation of LLMs in data- and compute-constrained environments.

Technology Category

• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Natural Language Processing: (Large) Language Models
• Cognitive Modeling & Cognitive Systems: Adaptive Behavior

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Continual learning in large language models (LLMs) typically encounters the critical challenge of catastrophic forgetting, where previously acquired knowledge deteriorates upon exposure to new data. While techniques like replay buffers and parameter-efficient tuning (e.g., Low-Rank Adaptation or LoRA) have been proposed, few studies investigate real-time domain adaptation under strict computational and data-stream constraints. In this paper, we demonstrate a lightweight method combining LoRA and a minimal replay mechanism in a realistic streaming setting across three diverse knowledge domains: medical question answering, genetics, and law. Using perplexity, semantic similarity, and GPT-based human-like evaluation metrics, we quantify the model's adaptation, forgetting, and recovery over time. Our experiments reveal that while catastrophic forgetting naturally occurs, even minimal replay significantly stabilizes and partially restores domain-specific knowledge. This study contributes practical insights for deploying adaptable LLMs in resource-constrained, real-world scenarios.

Problem

• Addressing catastrophic forgetting in streaming language models
• Enabling real-time domain adaptation with computational constraints
• Evaluating knowledge retention across diverse domains using replay

Innovation

• Combines LoRA with minimal replay mechanism
• Lightweight method for real-time domain adaptation
• Quantifies adaptation and forgetting using multiple metrics