Closed-form Estimator for Codex pass@k

sirbayes · x · 2026-07-15

The text introduces a closed-form estimator for Codex pass@k that computes the exact expectation across all \(C(N,k)\) subsets, resulting in zero subsampling variance.

The author adds that you can apply a bootstrap layer over this estimator to aggregate results across different problems, yielding a pass@k curve with confidence intervals.

Original post →

More from Research

Research channel →