Reproduce “DiffGate: Difficulty-Gated Teacher Guidance for 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
Karn Tiwari, Varnith Chordia, Prathosh A P
DiffGate combines group-relative policy optimization (GRPO) with bounded teacher guidance for training language models on verifiable math and code tasks. A verifier selects failed trajectories for token-level guidance, whose weight increases as fewer sampled answers pass. In evaluations, pass@8 improves over matched GRPO across math and code settings, though math avg@8 slightly declines; agentic and long-horizon tasks remain untested.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.