CPR for LLMs: Critical-Point Routing against Catastrophic Forgetting in Domain Adaptation
2026-08-31 • Computation and Language
Computation and LanguageArtificial Intelligence
AI summaryⓘ
The authors address a problem with adapting large language models to specific topics, which usually harms their overall knowledge. Instead of mixing everything in one model, they keep the original model for general use and call a specialized expert model only when needed. They created a system called CPR that decides for each word whether to use the expert or the base model, improving topic-specific accuracy without losing general understanding. This method works better than previous ones while keeping the extra work low.
Large Language ModelsSupervised Fine-TuningCatastrophic ForgettingDomain AdaptationToken-level RoutingHierarchical RouterInference ProcedureExpert ModelDomain-General Trade-off
Authors
Kwangmin Ki, Yunhun Nam, Jongheon Jeong, Jaehyung Kim
Abstract
Supervised fine-tuning (SFT) is the de facto standard for adapting large language models (LLMs) to target domains, but it often degrades the model's general capabilities, a phenomenon known as catastrophic forgetting. Existing approaches typically modify the SFT loss to mitigate forgetting, but they inevitably operate along a domain-generality trade-off. In this work, we step outside this trade-off by decoupling the two capabilities at the model level: we keep the original base model for general capability, and selectively invoke the SFT expert only when domain-specific knowledge is required. Specifically, we propose CPR (Critical-Point Routing), a token-level routing framework between a base model and its expert derivative, based on critical tokens where the base model fails but the expert succeeds. We train a lightweight hierarchical router that estimates the expert-call probability per token, and pair it with a tailored inference procedure that combines momentum smoothing and threshold gating. Across diverse model-domain configurations, CPR achieves state-of-the-art across all settings, surpassing SFT expert by 1.4-5.5% in domain performance while recovering its general-capability drop from 3.4-14.5% to at most 0.5%, with minimal overhead from invoking the expert on only one-third of tokens.