GeoArbiter Improves Remote-Sensing LLM Accuracy and Reduces Hallucinations

Multimodal Large Language Models (MLLMs) used for remote sensing often make factual claims that cannot be directly confirmed by satellite imagery alone, such as the specific identity or function of a facility. While geographic retrieval can provide this missing context, simply adding all retrieved records can lead to new problems, as some records might contradict clear visual evidence. GeoArbiter addresses this by introducing a "cross-modal verifiability" principle. It intelligently filters geographic information, only injecting facts that the image cannot verify and withholding information that disputes visually verifiable attributes. This content-level filtering significantly reduces claim-level hallucinations and improves robustness to conflicting data, leading to better accuracy in land-use classification without requiring model retraining.