Reproduce “Inference after coding agents” 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
Florian Gunsilius, David Van Dijcke, Kaspar Wüthrich
The authors introduce a replay-based correction for p-values selected by coding agents. It reruns the same agent on subsamples and measures whether it selects the same analysis more often when that analysis has a favorable p-value. In a zero-effect employment-support simulation, GPT-6-Luna reported p<.05 in 97/99 cases (98.0%), while replay rejected in 5.1%; validity is asymptotic and assumption-dependent.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.