Discovering Governing Laws with World Models
Published in Monash University, 2026
Discovering closed-form mathematical equations from observational data is a cornerstone of scientific discovery and engineering practice, yet most symbolic regression schemes largely rely on dated combinatorial search heuristics that scale poorly with system complexity. In this work, we propose decoding the latent space of large pretrained vision and multi-physics foundation models into closed-form equations with large language models. To tap into the latent space of decoder-only foundation models, we introduce decoder compression, a mechanism that bottlenecks intermediate residual stream activations into compact embedding tokens. Evaluated across both analytic and challenging real-world benchmarks spanning fluid dynamics, reaction-diffusion and magnetohydrodynamics, our approach is able to consistently outperform classical genetic algorithms and end-to-end neural baselines by up to an order of magnitude in fit accuracy. Furthermore, our framework is roughly 2 to 48 times faster than most baselines, demonstrating a scalable paradigm for AI-driven scientific discovery.
Recommended citation: Novomestsky, M., Harandi, M., Ali, R. (2026). Discovering Governing Laws with World Models. Monash University.
Download Paper
