Reproduce “Who Teaches Which Token? Verifier-Gated Multi-Expert On-Policy Distillation for Scientific Reasoning” 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
Xun Xu, Zaixi Zhang
VG-OPD assigns specialist teachers to selected tokens in scientific answers. It verifies each expert on the same answer criterion as the student, then uses the expert’s disagreement to localize a token-level teacher signal. On Qwen3-8B, it reaches 70.9% accuracy on ChemBench versus 69.3% for rubric-GRPO. Results cover seven text-only benchmarks with Qwen3-4B and Qwen3-8B; broader generality and scaling remain unverified.
Key insights
- Teacher selection, token selection, and supervision strength are treated as separate decisions, making expert guidance conditional on verified usefulness rather than prompt-level domain labels.
- Ablations show that disagreement localization and expert eligibility are central: random masking, random routing, or removing the gate substantially reduces performance.
- The approach improves targeted positions rather than simply correcting high-surprisal tokens, while capability routing remains mostly aligned with the prompt’s main domain but crosses domains in about 12% of licensed cases.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.