The ProKDA method addresses performance degradation in multimodal models by decoupling explanation generation from label prediction. It uses an agentic pipeline to construct external background knowledge, preventing task objective interference during the training of explainable detectors.
HOW THIS AFFECTS YOU
●
researcherYou can improve multimodal detection accuracy by aligning knowledge to decisions progressively.
●
policyThis contributes to safer content moderation through more reliable, explainable detection results.