Variability Management for Large Language Model Tasks: Practical Insights from an Industrial Application
Kentaro Yoshimura, Haruki Oishi · 2024
This paper presents an industrial experience of incorporating the Variability Management approach into a variability-rich Large Language Model (LLM) application for connected vehicle data management. The development of a driving image caption application using LLM requires the creation of numerous prompt variants for diverse traffic scenes and the application of multiple LLMs for different tasks. To manage the prompt variants and LLM models, we utilize the Variability Management approach, specifically focusing on variability modeling. Our proposed method analyzes the commonality and variability of the driving image captions and develops a feature model for the LLM application. We implement the proposed approach in a traffic scene image captioning application and demonstrate its effectiveness in generating distinct captions for each traffic scene using three different LLM tasks.