Geo-Strat-RL Teaches AI Geological Reasoning with Verifiable Rewards

Evaluating the geological reasoning capabilities of vision-language models (VLMs) is challenging because ground-truth geological histories are often ambiguous or unavailable. This research introduces Geo-Strat-RL, a novel synthetic environment designed to address this by generating stratigraphic observations with known underlying process histories. The environment combines a geological generator with an executable verifier that scores various aspects of geological reconstruction, including chronology and structural relationships. The core of Geo-Strat-RL is its use of reinforcement learning with verifiable rewards (RLVR). This approach allows models to learn geological event reasoning by optimizing for verifiable tasks, ensuring consistency with both observed evidence and established geological principles. This method moves beyond simple pattern recognition to foster a deeper understanding of temporal and structural relationships. Experiments demonstrate that RLVR significantly enhances geological reconstruction in VLMs, leading to higher geological content scores on unseen stratigraphic diagrams. Crucially, the reasoning learned from stratigraphic diagrams also transfers to synthetic seismic observation domains without specific seismic training, suggesting that RLVR can teach reusable geological concepts across different data formats.