ExplorationBench tests whether AI systems can discover new rules rather than recall prior knowledge. Systems receive flawed manuals for two deterministic environments, AlienCode and AlienLogic, choose experiments over four rounds, and then solve 70 held-out tasks without tools after each round.
Ten systems were evaluated across three trajectories per sandbox. The strongest learned and applied unfamiliar rules, but outcomes varied substantially between trajectories, and additional exploration sometimes stalled or reduced earlier gains. Exact execution and proof checking make the useful for separating rule discovery from successful application.