发表机构
Korea University; Yonsei University（高丽大学; 延世大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
本研究提出CPR框架，通过解耦基础模型与SFT专家模型，基于临界token进行路由，在领域自适应中缓解LLMs的灾难性遗忘，实现最优性能且开销极小。
AI 中文摘要
监督微调（SFT）是将大语言模型（LLMs）适配到目标领域的事实上的标准方法，但它常会降低模型的通用能力，这一现象被称为灾难性遗忘。现有方法通常会修改SFT损失来缓解遗忘，但不可避免地会面临领域-通用性的权衡问题。在本研究中，我们通过在模型层面解耦这两种能力，跳出了该权衡：我们保留原始基础模型以维持通用能力，仅在需要领域特定知识时选择性调用SFT专家模型。具体而言，我们提出了CPR（Critical-Point Routing，临界路径路由），这是一种基于临界token的基础模型与其专家衍生模型之间的token级路由框架，临界token是指基础模型失效但专家模型成功的token。我们训练了一个轻量级分层路由器，用于估计每个token的专家调用概率，并将其与结合了动量平滑和阈值门控的定制推理过程配对。在多种模型-领域配置下，CPR在所有设置中均达到了最优性能，在领域性能上较SFT专家模型提升了1.4-5.5%，同时将其通用能力的下降幅度从3.4-14.5%恢复至最多0.5%，且仅在三分之一的token上调用专家模型，开销极小。
英文摘要
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.
