Reproduce “Weak-to-Strong Generalization under Distribution Shifts” 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

Myeongho Jeon, Jan Sobotka, Suhwan Choi, Maria Brbić

The EPFL research team developed RAVEN, a framework for robust weak-to-strong generalization that effectively operates under distribution shifts by dynamically combining an ensemble of weak models. RAVEN achieved up to a 55% average improvement in Performance Gap Recovered on out-of-distribution image classification tasks and automatically assigned higher weights to weak models performing better on target data without ground-truth access.

Key insights

  • Naive weak-to-strong generalization methods exhibit critical failures under distribution shifts, often resulting in strong models performing worse than their weak supervisors.
  • Dynamically learning to weight an ensemble of weak models enables the strong model to automatically identify and prioritize more reliable supervision signals, even without access to ground-truth labels for the target data.
  • An easy-sample guided initialization phase is vital for robust weak-to-strong generalization, providing a stable starting point that prevents the strong model from taking "shortcuts" and ensures better overall performance.

Summary by alphaXiv.

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