DiG-bench benchmark tests AI's ability to infer unwritten environmental rules through exploration

DiG-bench, a new benchmark comprising 70 self-contained text-based games, measures how well AI systems can discover hidden rules and objectives through interaction, rather than being explicitly fed them. Researchers from institutions including the University of Oxford and MIT found that current frontier models struggle with the benchmark; for instance, Opus 5 and Claude Code achieved only 20% success rate in Tier 7 tasks, the hardest level. The games are handcrafted by human experts, kept private to prevent AI training, and are designed to be difficult for today's models.