A new arXiv paper introduces DiG-bench, a benchmark for testing whether AI systems can discover new rules through experimentation. The benchmark contains 70 independent games, each encoded as a short string with hidden transformation rules.
The goal is different from answering known questions. Systems must interact with the environment, form hypotheses and use later levels to show whether they have discovered the underlying rules. That makes the benchmark closer to controlled scientific discovery than ordinary question answering.
The authors argue that current AI benchmarks do not directly measure this capability often enough, even though discovery is central to science and problem solving. A model can be strong at pattern matching or explanation without being good at designing experiments that reveal unknown rules.
DiG-bench is a controlled game setting, so it does not prove real-world scientific ability. Its value is that it gives researchers a clearer testbed for separating discovery behavior from memorized knowledge or post-hoc explanation.