Where-OPD Uses Spatially Guided Self-Distillation for MLLMs
September 30, 2026
Where-OPD implements on-policy self-distillation for multimodal models by providing the teacher with textual, spatially grounded guidance. This allows MLLMs to improve fine-grained perception without needing human-annotated grounding data or external teacher models.
HOW THIS AFFECTS YOU
●
builderYou can improve the fine-grained perception of your MLLMs using self-distillation and synthetic spatial guidance.
●
researcherThis offers a way to perform on-policy distillation for multimodal tasks without expensive external grounding labels.