[1Yin-A] Poster 1

[1Yin-A-43] Topic Maintenance in Large Language Models Using Control Barrier Functions

〇NAOYA SENOO¹, Masaki Inoue¹, Yuya Miyaoka¹ (1. Keio University)

Keywords:

Large Language Model,Control Barrier Function,Topic Maintenance

The objective of this study is to achieve topic maintenance in Large Language Models (LLMs) through a training-free approach. In practical scenarios such as specialized consultation systems and educational chatbots, where stable text generation within a specific topical scope is required, existing training-based methods suffer from high computational costs and difficulties in adapting to new constraints. In this research, we applied CBF-LLM—a training-free alignment method utilizing Control Barrier Functions (CBF) from control engineering—to the task of topic maintenance. Experiments were conducted using Llama 3 (8B), and statistical evaluations across multiple prompts confirmed the effectiveness of topic maintenance under both neutral and disruptive (disturbance-inducing) conditions. The results demonstrate that CBF-LLM is effective for topic maintenance across a diverse range of conditions.