LLM-Driven Generative Agents for Simulating Occupant Feedback in Built Environments

Occupant feedback plays a vital role in optimizing indoor environmental conditions in intelligent buildings. Yet in operational settings, such feedback is often sparse, delayed, or difficult to integrate into real-time control. Existing occupant-centric control (OCC) approaches often rely on static rules or infrequent user input, limiting their ability to reflect diversity of occupant preferences, particularly in shared or dynamic spaces, such as classrooms and offices. Furthermore, most existing simulations of occupant behavior rely on single-user models or predefined assumptions, lacking contextual awareness and adaptability. To address these gaps, this paper proposes a large language model (LLM)-driven generative agent that simulates context-aware occupant feedback grounded in indoor environmental sensor data with occupant persona profiles. Each generative agent is defined by a unique combination of demographic background, environmental history, lifestyle patterns, and personality traits, allowing differentiated responses to the same environmental stimuli. The proposed LLM agent further incorporates retrieval-augmented generation, where each agent maintains an evolving memory of past experiences using a Facebook AI similarity search index (FAISS). Retrieved memories are integrated into the agent’s prompt, allowing feedback to reflect prior context. The proposed agent is evaluated and validated using real indoor environmental sensor data collected from the Virginia Tech Blacksburg campus, alongside real occupant feedback collected through a web-based feedback interface. The results highlight the potential of LLM-based generative agents in OCC and adaptive surrogates for sparse human feedback.