Researchers built a geolocation benchmark to test whether AI agents can combine visual clues and web search to pinpoint locations

Researchers created GeoBrowse, a test suite that forces AI systems to combine ambiguous visual evidence with multi-step web searches to identify locations — a harder problem than existing benchmarks require. The system tests whether AI can actually reason across fragmented clues rather than just pattern-matching, which matters because most AI agents today either ignore images or search without visual reasoning.

This is a measurement problem dressed as a benchmark. Right now, there's no standard way to test whether AI agents can do the actual cognitive work of combining visual and textual evidence — most existing tests either use text alone or treat images as separate from reasoning. GeoBrowse makes that visible. The paper shows that an AI agent using the right mix of tools (visual tools plus search tools) beats agents that use only text or only images, which means the structure of the problem matters more than raw capability. This hints at something uncomfortable: AI agents might look capable until you force them to integrate evidence the way humans actually do.