ThermoForce: A Physics-Structured Interventional World Model for Building HVAC Control
arXiv:2607.03942
2026
Architecture
1 ideas extracted · analyzed Aug 30, 2026
What the math gives to ML
The paper isolates passive forecasting from causal actuation response: a frozen foundation model predicts what the system would do without intervention, while a small structured operator predicts the incremental effect of control. The transferable asset is the additive counterfactual decomposition and hard monotonicity constraint, which prevents a high-accuracy observational predictor from assigning the wrong sign to an intervention. This suggests a general neural world-model architecture in which a large pretrained dynamics backbone remains frozen and a low-dimensional, sign-constrained residual adapter is learned from scarce intervention data.
Ideas from this paper
△ Mechanism confirmed, baseline not beaten
2026
Separate a pretrained sequence model's passive prediction from the causal effect of an action, and learn only the latter with a compact monotone adapter. The adapter receives the current latent state and an action deviation, but its action-to-output Jacobian is constrained to have the physically correct sign, preventing intervention predictions that move opposite to the applied control.
Useful7/10
Difficulty4/10
Novelty7/10