MMSearch-OPD: Easy Made Hard, Hard Made Learnable for Long-Horizon Multimodal Search Agents
MMSearch-OPD reports stronger multimodal search by expanding simple training questions and gradually reducing teacher guidance.
Its Refine pipeline reuses existing image–question pairs to build longer reasoning chains, with controls for search depth and cross-modal evidence requirements. Training combines reinforcement learning with token-level teacher guidance on imperfect student attempts. That guidance responds to the reward gap and decays as training proceeds. The authors report gains from refined data and further improvements from joint training after supervised fine-tuning; the abstract gives no numerical results.
HF Daily Papers' note
Its Refine pipeline reuses existing image–question pairs to build longer reasoning chains, with controls for search depth and cross-modal evidence requirements. Training combines reinforcement learning with token-level teacher guidance on imperfect student attempts. That guidance responds to the reward gap and decays as training proceeds. The authors report gains from refined data and further improvements from joint training after supervised fine-tuning; the abstract gives no numerical results.
HF Daily Papers' note
score 5