Work

Towards Generative Occupants: LLM Agents in Physics-Based Building Simulation

Year: 2026
Type: conference-paper
Abstract: Current occupant representations in building simulation often capture schedules, rules, or stochastic variation, but provide limited support for modelling context-dependent and language-mediated decision processes. This paper presents a work-in-progress framework coupling a built-from-scratch physics-based building simulation environment with LLM-driven occupant agents. The simulation core draws on established building simulation approaches and is checked against ASHRAE 140-2023 thermal test cases. The agent architecture uses planning, memory retrieval, action, reflection, and memory update to interact with the simulated environment. An exploratory pilot case study compares fixed-schedule, deterministic rule-based, and LLM-driven occupant representations under identical boundary conditions. The results show that LLM-driven occupants can be coupled to the physics-based simulation loop and can influence simulated energy demand, but should not be interpreted as validation of occupant realism. The paper concludes by outlining an evaluation roadmap for behavioural validation and occupant-centric control applications.
Language: English
Open Access status: hybrid
Award: G663249300
OpenAlex ID: W7165367252
FWCI: 9.482
Cites: 20
Cited by: 1
Related to: 10