COMPASS: Self-Adaptive LLM Prompt Engineering for Task Plan Explanations

A new arXiv preprint (2604.21092) introduces COMPASS, a proof-of-concept system for self-adaptive generation of task plan explanations via large language models. The approach formalizes prompt engineering as a cognitive and probabilistic decision-making process, modeling users' latent cognitive states such as attention and comprehension as a partially observable Markov decision process (POMDP). COMPASS synthesizes policies to adaptively generate explanations and refine prompts, addressing the challenge of diverse stakeholder groups formulating prompts without systematic understanding. The system aims to automate the generation of human-understandable explanations of opaque AI processes like automated task planning. Evaluation was conducted using two unspecified datasets. The work highlights the dependency of LLM explanation quality on effective prompt engineering and proposes a framework to automate this process.