ICKE : inference-controlled knowledge editing for stable large language models

Abstract

Knowledge editing aims to update the internal knowledge of large language models (LLMs) without full retraining. Existing approaches typically rely on parameter modification, which leads to high knowledge costs and cumulative interference, particularly for small models. In addition, focusing only on the success of target editing often masks instability and knowledge drift. This paper introduces a hybrid knowledge editing framework that combines a parameter-edited GPT-2 language model and Qwen2.5-1.5B model with query-dependent external knowledge retrieved via BM25 to guide inference-time generation, while keeping model parameters fixed during inference that reformulates knowledge editing as a constrained conditional generation task. Experiments conducted on GPT-2 Large and Qwen2.5-1.5B show that the proposed method maintains a high editing efficiency while achieving superior stability and consistency compared to baseline methods, offering a reliable solution for knowledge editing in small language models.