The authors built Refine, a data pipeline that reuses existing image-question pairs while increasing the amount and type of evidence needed to answer them.
Refined Questions and Teacher Guidance Improve Long-Horizon Multimodal Search
MMSearch-OPD turns simpler image-based questions into harder search tasks, then trains an agent with reinforcement learning and selective teacher supervision.
Chinese Tech
Huanyao Zhang · Jiepeng Zhou · Jiaye Lin · Ontai Tang · Jingtian Hao · Mugeng Liu · +6 more
Peking University · HKUST(GZ) · Tsinghua University · The University of Hong Kong · Zhejiang University
Research Digest··2 min read
Zhang and colleagues present a pipeline for training multimodal agents to investigate questions through multiple rounds of web search, visual retrieval and reasoning.
Why this paper
From Huawei Cloud BU and 5 others
In one line
MMSearch-OPD improves long-horizon multimodal search agents by combining data refinement with joint reinforcement learning and on-policy distillation.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§