VIPER-R1

Visual-to-symbolic discovery of physical laws from motion plots and observations. Accepted at NeurIPS 2026.

VIPER-R1 connects multimodal perception with scientific reasoning to discover symbolic physical laws from visual observations and trajectory data. It studies how vision-language models can reason about motion plots and phase portraits, formulate physical hypotheses, and refine them against numerical evidence.

Publication: NeurIPS 2026, accepted.

VIPER-R1 overview linking motion visualizations, scientific reasoning, and symbolic equations

The framework combines visual structure induction, reward-guided symbolic calibration, and refinement with an external symbolic regression tool. The accompanying PhysSymbol benchmark supports the evaluation of interpretable physics formula discovery.

Project website · Paper