DiG-bench Benchmark for Rule Discovery in Controlled Games

DiG-bench is a new benchmark consisting of 70 independent games designed to test an agent's ability to discover unknown transformation rules through experimentation. The benchmark features seven difficulty tiers, ranging from tasks solvable by current models to those challenging state-of-the-art agentic harnesses.