Geo-Strat-RL: Learning Geological Event Reasoning from Verifiable Tasks

Can vision-language models (VLMs) move beyond simple visual description to perform genuine geological reasoning? Specifically, the paper investigates whether models can reconstruct a latent geological event history—such as deposition, tilting, faulting, and erosion—from static, ambiguous visual observations like stratigraphic cross-sections or seismic-style images.

To make this reasoning measurable, the author introduces Geo-Strat-RL, a synthetic environment that generates stratigraphic diagrams paired with ground-truth event histories. The environment includes an executable verifier that scores model outputs based on chronological accuracy, event identity, and structural relationships. The author uses Reinforcement Learning with Verifiable Rewards (RLVR) to train LoRA adapters on open-source VLMs (Qwen series). By providing a structured, verifiable reward signal, the model learns to output a compact JSON representation of the geological history without requiring human-labeled preference data.

• Improved Reconstruction: RLVR training significantly improves the ability of VLMs to correctly identify and order geological events in held-out stratigraphic diagrams.
• Cross-Domain Transfer: Geological reasoning learned in the stratigraphic diagram domain transfers effectively to synthetic seismic-style observations. Even without training on seismic data, models trained on diagrams successfully reconstructed the latent history of seismic images, suggesting the model learned transferable geological concepts rather than just visual pattern matching.
• Verifiable Feedback: The use of an executable verifier allows for precise, component-wise evaluation of model performance, isolating failures in chronology, event identification, and structural logic.