Researchers introduce COMPASS, a self-tuning system that automatically improves how AI explains its decisions by learning what each user understands

A research team published COMPASS (COgnitive Modelling for Prompt Automated SynthesiS), a proof-of-concept tool that automatically refines the instructions (prompts) given to large language models (AI systems trained on text) so they generate better explanations of AI decision-making. Rather than requiring engineers to manually tweak prompts through trial-and-error, COMPASS models what users actually understand and adjusts explanations in real-time.

COMPASS treats prompt engineering as a decision-making problem: it tracks observable clues (how long a user reads, what they click) and models unobservable mental states (whether they're confused, paying attention, or uncertain) using a probabilistic framework called POMDP. This means the system learns which explanation style works best for which person and adjusts automatically, instead of one-size-fits-all explanations.

For business users relying on AI tools to justify automated decisions—loan approvals, hiring recommendations, task scheduling—this means explanations could become personalized and clearer. For engineers building AI systems, it removes the tedious work of manually rewriting prompts for different audiences. For compliance teams needing to defend AI decisions to regulators, better explanations reduce legal risk.