Licheng Jiao

Reasoning segmentation models stumble by cramming semantic interpretation and spatial localization into a single prompt token; decoupling "what" from "where" hits state-of-the-art grounding while fine-tuning less than 0.4% of the base MLLM.

Achieving a top rank in the MeViS-Audio challenge, this system showcases the power of agreement-based segmentation in accurately identifying objects described by spoken expressions.