DiG-bench: New AI Benchmark for Scientific Discovery in Games

Researchers have released DiG-bench (Discovery in Games), a new benchmark designed to evaluate AI agents' capacity for scientific discovery through experimentation in controlled environments. The benchmark consists of 70 independent games, each encoded as a short string with unique transformation rules that must be discovered through interaction. The games are structured across seven tiers of difficulty, with the lowest tier solvable by multiple models and the highest tier challenging the best agentic harnesses. The win conditions for each level are unknown, requiring agents to formulate novel generalizations—a central aspect of the scientific process. This benchmark addresses a gap in the AI landscape, as few existing benchmarks directly probe the ability to discover new knowledge when the objective is unknown. The announcement was made via arXiv preprint 2608.12593.