Reproduce “Evaluating Model Capabilities After Benchmark Saturation” 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

Morio Hamada, Jehyeok Yeon, Matteo Migliarini, Justin Olive, Magda Dubois

The authors test whether correct-response transcripts from saturated benchmarks can predict models’ capability on harder items. They compare transcripts pairwise, calibrating judgments with earlier models whose capability is measured using item response theory. Across 16 targets in four domains, pairwise judging reduced mean squared error from 1.37 to 0.26; retrospective results show predictive signal, not validated deployment performance.

Summary by alphaXiv.

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