Reproduce “Domain-Routed Delta Optimization for Multi-Teacher On-Policy 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
Lin Yao
DRD-MOPD consolidates domain experts into one dense student by composing expert-minus-base components and routing each sample’s training updates according to its domain. All components contribute to every forward pass. In one OWN step-20 checkpoint, their projections remain near their original directions, while the student’s six-benchmark macro score is 32.72%, versus 32.44% at initialization; Math and Code scores decline.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.