Nouveau Recherche PDF, HTML, DOCX et bien d'autres formats.

Prépublication · 2026

Inference after coding agents

Florian Gunsilius, David Van Dijcke et al. — Monde

Source : arXiv

Résumé

Empirical research is changing rapidly as AI coding agents are becoming increasingly proficient in every step of it, from writing code to manipulating data. We argue that while the potential of coding agents for science is vast, their use in hypothesis testing poses a substantial threat to classical inference, as these agents enable a supercharged form of selective inference. In particular, we argue that coding agents carrying out empirical research introduce a new type of post-selection inference that combines human-like, open-ended search over an implicit space of estimators with programmati…

Actions

Citation