Try O3-Bench on MapTrace-tuned Models

Hi @google-research and MapTrace contributors,

Thanks for the great work on **MapTrace**! I’m Kaican, part of the team behind **O3-Bench**, a multimodal benchmark aimed at evaluating whether models can truly “think with images” (including challenging map navigation/interpretation problems).

Given that MapTrace focuses on fine-grained spatial and route-tracing capabilities on map images, we thought it’d be interesting to see whether models fine-tuned on MapTrace can generalize to the map reasoning subset of O3-Bench.

Would you be open to giving O3-Bench a try with models fine-tuned on MapTrace, and sharing any insights/observations? This could also serve as an additional external data point on MapTrace’s transfer to broader map reasoning.

Here’s the O3-Bench dataset:
[https://huggingface.co/datasets/m-Just/O3-Bench](https://huggingface.co/datasets/m-Just/O3-Bench)

Happy to help with evaluation scripts or answers to any questions! :)

Best,
Kaican