sparc-vl-perception-reasoning

Improve vision-language model reasoning efficiency by decoupling perception (identifying task-relevant image regions) from reasoning (generating explanations), enabling asymmetric compute allocation. Reduces token overhead while improving accuracy through separate optimization of visual grounding and semantic reasoning stages.

Independent GitHub snapshot · Automated classification · Original documentation governs