When Does Better AI Actually Save More Energy? Decision-Aware Adaptive Fidelity for Building Control

Presented at the Fifth Annual UW–Madison Sustainability Symposium, hosted by the Office of Sustainability and the Nelson Institute for Environmental Studies, on October 28, 2026. This poster examines how numerical precision affects downstream decisions in neural networks and considers the implications for energy-aware adaptive computation. Using a controlled four-stage visual model, the study evaluates precision placement, prediction changes, decision quality, and cross-stage interactions. The findings motivate selective computational fidelity in applications such as intelligent building control, where improvements in decision quality must be considered alongside computational and operational energy costs. The University of Wisconsin–Madison institutional logo is used for identification purposes. University trademarks and third-party materials are excluded from the Creative Commons license and remain subject to their respective rights and usage policies.