Reproduce “Eliciting Weak-to-Strong Generalization with On-Policy Reverse 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

Youngrok Park, Sangmin Bae, Hojung Jung, Jongwoo Ko, Yunseon Choi, Young Jin Kim, Pashmina Cameron, Aaron Courville and 1 more

OPRD guides a stronger model with a smaller model’s post-training shift to modify verifier-driven updates, rather than distilling toward its outputs. In Reasoning Gym, Qwen3-8B-Base scores 58.77 Pass@1 averaged across four tasks and five checkpoints, versus Mix-RL’s 47.68. It exceeds every specialist; this is not the final-update score or a guarantee of faster convergence.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.