DiG-bench: Discovery in Games

Ruairidh M. Battleday, Kai Sandbrink, Jimi Cullen-Drohan, Zihan Yan, Timothy Muller, Clare Maguire, Ales Kubicek, Fraser Greenlee-Scott, Sukrit Sumant, Tri Dao, Jürgen Schmidhuber, Michal Valko, Joshua Tenenbaum, Thomas L. Griffiths, Zeb Kurth-Nelson, James C.R. Whittington

Preprint · Technical report (2026)

Abstract

Discovery, formulating novel generalizations, is a central part of the scientific process. DiG-bench is a benchmark for scientific discovery through interaction and experimentation in controlled environments with unknown objectives. It contains 70 independent text-based games, each with unknown transformation rules that agents must uncover, across seven difficulty tiers. Twenty-one games are publicly released and the remaining games are held private for secure evaluation; every game was solved by at least one human on a first attempt.

PDF · Project website · GitHub · arXiv · Semantic Scholar · bibtex