Reproduce “Knowing When to Teach: Learning from Teachable Prefixes in Multilingual Reasoning Distillation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jaewon Cho, Minseo Yoon, Jinyoung Park, Jinwoo Seo, Jaewon Chu, Alice Oh, Hyunwoo J. Kim
TOPSD selectively distills the privileged teacher’s predictions on student-generated reasoning trajectories in multilingual mathematical reasoning. It selects a prefix using teacher–student likelihood alignment, then verifies whether the teacher can recover the correct answer from that prefix. On PolyMath High, Qwen3-8B reaches 25.12% overall Avg@3, 2.64 percentage points above its strongest baseline. Across ten languages, broader domains and model families remain untested.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.