insight-o3-multimodal

Enable VLMs to perform generalized visual search—locating relational, fuzzy, and conceptual regions from free-form language descriptions. Introduces O3-Bench benchmark with high-density composite charts/maps, uses RL-trained vSearcher for spatial localization, improving frontier models (GPT-5-mini 39%→61.5%) without architecture changes.

Independent GitHub snapshot · Automated classification · Original documentation governs