MMSearch-OPD: Joint RL and On-Policy Distillation for Long-Horizon Multimodal Search Agents
PekingUniversity · hf · 2026-10-16
Peking University with HKUST(GZ), Tsinghua, HKU, Zhejiang University, and Huawei Cloud propose MMSearch-OPD, tackling training bottlenecks as multimodal deep search shifts from short queries to long-horizon investigations: scarce challenging training questions and insufficient process feedback.
The approach combines controllable data graph composition and visual enrichment (relation collection, image curation, evidence validation) to build harder questions from existing image–question pairs, with joint training of reinforcement learning and on-policy distillation to address common failure modes in long-horizon investigations—redundant verification, premature stopping, and incorrect answers—where terminal rewards alone offer limited signal.
More from coding & agent
- Vibe coder chains ChatGPT sessions into Claude Code CLI via Desktop Commander MCP — MidahBootyQuay · 2026-10-16
- What a single operator can do with many agents is mind blowing — natesiggard · 2026-10-16
- When is a Claude Code session too long? One 3-hour spiral reignites the context debate — Additional-Buy-2562 · 2026-10-16
- Claude Code quirk: team member agent replies never reach the lead agent — Sauers_ · 2026-10-16
- Agent wrote the full answer, but the harness silently threw it away as a 'progress note' — Stunning-Sherbet1853 · 2026-10-16
- Vibe coding gets a bad rap: an engineer draws the line between vibes and real engineering — bendee983 · 2026-10-16