BranchOPD from Tsinghua, PKU and SenseTime: branching on-policy distillation lifts math accuracy 18.55 points
BranchOPD · hf · 2026-10-13
Researchers from Tsinghua University, Peking University and SenseTime propose BranchOPD (Branching On-Policy Distillation), addressing the limited supervision coverage of standard on-policy distillation.
- Method: Informative decision points are selected via student uncertainty and teacher-student disagreement; continuations are resampled from shared prefixes with token-level teacher supervision applied to both main trajectories and branches.
- Cost: Shared-prefix reuse cuts student-generated token costs by 19% on ALFWorld and 31% on WebShop vs seven independent trajectories.
- PBR: Protocol-boundary repair applies targeted supervision to mitigate repetition and wasted generation.
- Results: Outperforms Vanilla OPD and Multi-Rollout OPD (n=7) across all benchmarks — +18.55 points mean math accuracy (avg@16) and +13.8 points ALFWorld ID/OOD success. Code and models are open-sourced.
More from Research
- XPeng open-sources GroundingPI, a 4B grounding model beating larger models across 34 benchmarks — _akhaliq · 2026-10-14
- DOE puts $159M into Genesis projects, expanding OSPREY agent that runs particle accelerators — omnivaughn · 2026-10-14
- Real2Sim gets easy with coding models, but sim-to-real physics and contact solvers lag — jonstephens85 · 2026-10-14
- MIT CSAIL's InstructMesh makes genAI 3D models you can actually 3D print — MIT_CSAIL · 2026-10-14
- Swing Choreo launches with a self-built motion benchmark: 61/100 vs 38 for the next best — andrew_n_carr · 2026-10-13
- No public benchmark for motion generation, so Swing built one: Choreo scores 61 vs 38 — andrew_n_carr · 2026-10-13