REVERSE: Reinforcing Evidence Verification and Search for Agentic Image Geo-localization

Researchers introduced REVERSE, a reinforcement-learning framework for agentic image geo-localization.

It trains a model to search for evidence, verify clues, and refine location hypotheses over multiple turns.

REVERSE uses tool-grounded trajectories, process rewards, and an offline search cache for training.

The approach outperformed stronger retrieval-based baselines on photo location estimation with a smaller model.