Toward Multimodal RAG Agents for Subsurface Characterization: Introducing GeoMM-Bench and Baseline Results

GeoMM-Bench is the first benchmark for lithofacies interpretation from well-log display images, the visual form in which much subsurface data, including legacy scanned archives, actually exists. This preprint introduces the benchmark and reports baseline results for off-the-shelf vision-language models across two tracks, real operator displays from the Vilkyciai-22 well and rendered displays at scale from the public FORCE 2020 dataset. The central finding is a separation between learnability and off-the-shelf visual extraction. The lithofacies signal is readily recovered from the underlying curves by a simple numeric classifier, yet off-the-shelf vision encoders do not read it from the display image. The benchmark, the evaluation harness, and a datasheet are released to support reproducible assessment of future models. Source code: https://github.com/synkalabs/geomm-bench