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

Researchers present Geo-Strat-RL, a synthetic environment that trains vision-language models to reason about geological histories through reinforcement learning with verifiable rewards. The system demonstrates that geological reasoning learned from stratigraphic diagrams can transfer to seismic data without domain-specific training, suggesting AI models can learn generalizable geological principles across different observation formats.

Geo-Strat-RL addresses a fundamental challenge in training AI systems for scientific reasoning: how to evaluate and improve model performance when ground-truth answers are ambiguous or unavailable. The research constructs a verifiable geological reasoning task using a synthetic environment that generates stratigraphic observations paired with executable verifiers, enabling reinforcement learning with meaningful reward signals. This approach bridges a gap between pattern recognition and causal reasoning—teaching models to understand temporal relationships and structural principles rather than simply identifying visual features.

The work emerges from broader efforts to enhance vision-language models' capabilities in specialized domains requiring domain expertise. Previous approaches struggled with geological reconstruction because valid interpretations of rock layers and deformation patterns can vary, and no single "correct" answer exists when data is incomplete or ambiguous. The Geo-Strat-RL framework solves this by encoding geological principles directly into verifiable tasks that score chronology, event identity, deposition, and structural consistency.

The most significant finding concerns transfer learning across observation domains. When the model trained on stratigraphic diagrams was tested on synthetic seismic data—a completely different representation derived from acoustic properties—it maintained reasoning accuracy without seismic-specific examples. This demonstrates that reinforcement learning with verifiable rewards teaches reusable conceptual understanding rather than domain-specific pattern matching, a critical capability for AI applications in science and engineering.

The implications extend beyond geology. This methodology offers a template for training AI systems in other domains where ground truth is complex, reasoning must follow domain principles, and observations span multiple data modalities. Future work may apply similar verifiable-reward frameworks to planetary science, climate modeling, and other fields where causal reasoning and principle-guided inference are essential.

• →Geo-Strat-RL uses reinforcement learning with verifiable rewards to improve geological reasoning in vision-language models beyond pattern recognition.
• →The synthetic environment includes executable verifiers that score geological accuracy based on chronology, event identity, and structural relationships.
• →Models trained on stratigraphic diagrams transfer geological reasoning to synthetic seismic data without seismic-specific training.
• →The approach demonstrates that verifiable task design enables AI to learn generalizable domain principles across different observation formats.
• →This methodology provides a replicable framework for training scientific reasoning in AI across domains with complex or ambiguous ground truth.