ExplorationBench finds AI exploration gains can reverse

Even the strongest AI systems can pick up unfamiliar rules by exploring a controlled environment, but their gains are unstable, and continued exploration can undo earlier progress. That is the mixed reading of ExplorationBench, a new benchmark and arXiv paper that aims to separate genuine scientific discovery from pretraining recall.

The benchmark's central trick is what the authors call verifiable Alien Worlds: environments whose 'rules are executable, so every answer can be checked exactly, and they conflict with familiar knowledge, so recall alone cannot solve the tasks.' Two sandboxes ship with the paper. AlienCode has 31 discovery targets across 70 tasks; AlienLogic has 24 discovery targets across 70 tasks. Each hands a system a flawed manual, task-specific environmental feedback, and a dedicated tool-call schema, then asks it to explore first and solve held-out tasks second.

Across ten AI systems evaluated, the authors write that 'the strongest systems can acquire and apply unfamiliar rules,' before immediately qualifying the claim: 'performance varies substantially across trajectories and continued exploration can stall or reverse earlier gains.' The paper's abstract lists no per-model scores and does not name which ten systems were tested.

The authors close by framing the benchmark as 'a step towards AI systems that can acquire and apply genuinely new knowledge through exploration in unknown environments.'