MMSearch-OPD: Easy Made Hard, Hard Made Learnable for Long-Horizon Multimodal Search Agents
MMSearch-OPD enhances long-horizon multimodal deep search agents by combining controllable data refinement with joint reinforcement learning and on-policy distillation. It uses a 'Refine' pipeline to systematically increase search demands—like expanding reasoning chains and requiring cross-modal evidence—and trains the agent using reward signals modulated by a teacher model, leading to stronger performance on complex investigations.