Teaching AI to read a map

Google Research · 7 months ago · 6

Google researchers created MapTrace, a synthetic data generation pipeline that teaches multimodal language models to trace valid paths on maps by generating 2M annotated training examples using Gemini and Imagen models. Fine-tuning models on 23,000 paths from the dataset reduced path-tracing error (NDTW metric) from 1.29 to 0.87 for Gemini 2.5 Flash and improved success rates by up to 6.4 percentage points. The approach demonstrates that spatial reasoning abilities can be taught explicitly to AI models through targeted synthetic supervision rather than manually annotating real-world maps.