GPT-5 shows promise in VAV control, but reinforcement fine-tuning struggles

Researchers explored the use of a frontier reasoning model, GPT-5, for controlling multi-zone variable-air-volume (VAV) systems, aiming to balance comfort, air quality, and energy use. While GPT-5 showed promise by reducing HVAC electricity consumption by 6.2% without building-specific training, it also reduced the ventilation margin. Further attempts to fine-tune an open-weight model using reinforcement learning (RFT) with a rollout verifier proved less successful, failing to improve energy efficiency or temperature compliance compared to a baseline. AI

IMPACT Demonstrates potential for LLMs in complex control systems, though highlights challenges in fine-tuning for optimal performance and safety.

RANK_REASON Research paper detailing the application of a frontier reasoning model and reinforcement fine-tuning for a specific control task. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• GPT-5

• paper
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →