MMSearch-OPD for Long-Horizon Multimodal Search Agents
October 14, 2026
MMSearch-OPD uses a refinement pipeline to expand reasoning chains from simple multimodal questions into complex, long-horizon search tasks. It combines joint reinforcement learning with on-policy distillation to train agents on difficult, evidence-heavy investigations.
HOW THIS AFFECTS YOU
●
builderYou can develop agents capable of complex, multi-step visual and textual information retrieval.
●
researcherYou can train multimodal agents on more challenging, long-horizon reasoning data.