Large-scale language models (LLMs) are vulnerable to hallucinations due to mismatch in self-awareness when processing queries that cross knowledge boundaries. Existing mitigation strategies use uncertainty estimation or query rejection mechanisms, but they suffer from poor computational efficiency and usability. In this paper, we propose an explicit knowledge boundary modeling (EKBM) framework that integrates fast and slow inference systems to balance reliability and usability. The framework first uses a fast model to generate confidence-indicated responses, allowing high-confidence outputs to be utilized immediately. On the other hand, uncertain predictions activate a slow refinement model to improve accuracy. To adapt the model’s behavior to the proposed goal, we propose a hybrid learning pipeline to enhance self-awareness without compromising task performance. Evaluation results on the conversation state tracking task show that EKBM achieves better model reliability than the uncertainty-based baseline model. Further analysis shows that the refinement significantly improves accuracy while maintaining low computational overhead. The framework effectively balances accuracy and practicality by establishing a scalable paradigm for deploying reliable LLMs in error-sensitive applications.
