ExplorationBench Evaluates AI Discovery in Verifiable Alien Worlds

ExplorationBench uses executable, rule-based 'Alien Worlds' to measure an AI's ability to hypothesize and iterate without relying on training data recall. The benchmark includes AlienCode and AlienLogic sandboxes to verify genuine scientific discovery through exact rule-checking.