OccuReward: LLM-Guided Occupant-Centric Reward Shaping for Demographic Equity in Grid-Interactive Buildings

Researchers introduce OccuReward, an LLM-guided framework that shapes reward functions for AI-controlled building energy systems to promote demographic equity in occupant comfort. Testing with four occupant profiles reveals significant disparities in initial AI performance, with elderly female occupants experiencing lowest satisfaction, though targeted refinement achieved dramatic improvements (567% for elderly females) while reducing energy costs by 3.2%.

OccuReward addresses a critical blind spot in AI-driven building automation: the potential for algorithmic systems to systematically disadvantage specific demographic groups while optimizing for aggregate efficiency. The framework leverages large language models not for real-time decision-making but for static reward function design, a pragmatic approach that reduces computational overhead while enabling iterative refinement based on equity metrics. The Comfort Equity Index serves as a novel accountability mechanism, allowing researchers to track disparities across demographic segments rather than treating occupant satisfaction as a monolithic objective.

Building automation represents a rapidly expanding domain where AI optimization directly impacts daily quality of life. Current systems typically prioritize energy efficiency and cost reduction, creating inherent tension with comfort optimization—tensions that manifest unevenly across populations with different thermal preferences and health sensitivities. The ASHRAE database grounding provides empirical validity often absent from theoretical fairness research.

The results demonstrate both promise and persistent challenges. While reward-level intervention achieved substantial improvements (particularly the 53.8% gain for health-sensitive occupants), the authors acknowledge that demographic disparities persisted despite optimization efforts. This finding carries implications for building developers, facility managers, and AI vendors implementing autonomous energy systems. Systems deployed without equity considerations risk creating regulatory exposure and tenant satisfaction issues.

Future work must extend beyond reward shaping to explore whether disparities originate in training data, model architecture, or fundamental trade-offs between efficiency and equity. As smart buildings proliferate in urban environments, establishing fairness baselines becomes increasingly urgent for ensuring equitable resource allocation.

• →LLM-guided reward shaping successfully reduced demographic disparities in building comfort control, with elderly female satisfaction increasing 567% through iterative refinement
• →The Comfort Equity Index provides a quantifiable framework for measuring algorithmic fairness in building automation systems
• →Equity-aware optimization achieved simultaneous improvements in occupant satisfaction across multiple demographics while reducing energy costs by 3.2%
• →Despite significant progress, demographic disparities in AI-controlled building systems persist, indicating that reward-level intervention alone is insufficient for full fairness
• →Building automation represents a critical domain where algorithmic bias directly impacts occupant health and comfort, yet fairness considerations remain underexplored in current deployments