LLM explanations are meant to make model behavior legible and help users make better decisions on downstream tasks. But default training approaches for explanations typically target model-centric objectives, which may not align with how humans actually use the information. This paper from arXiv addresses that gap by proposing a calibration method based on reinforcement learning.
The authors suggest training explanations with rewards tied to human decision quality, rather than to static explanation properties. That means the model learns to produce explanations that are not just internally consistent, but genuinely useful when a person acts on them. The approach reframes explanation quality as a function of human outcomes, not just model outputs.
Because the abstract is brief, the article does not include experimental details or quantitative results. Still, the idea is significant: it points toward a more user-centered evaluation of explainability, where the success of an explanation is measured by the decisions it enables.