Reproduce “DiG-bench: Discovery in Games” 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
Ruairidh M. Battleday, Kai Sandbrink, Jimi Cullen-Drohan, Zihan Yan, Timothy Muller, Clare Maguire, Ales Kubicek, Fraser Greenlee-Scott and 8 more
DiG-bench tests whether AI systems can learn unfamiliar game rules through interaction. Its 70 text-based games require players to probe mechanics and use observations to choose later actions, reducing the role of visual perception. Gemini 3.1 Pro won 18 of 70 games without rules and 69 with rule descriptions, suggesting rule discovery is a primary challenge, not the only one.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.