Researchers introduced ExplorationBench, a benchmark for evaluating whether AI systems can discover and apply unfamiliar rules through exploration in verifiable alien worlds. The benchmark addresses two difficulties in assessing scientific exploration: checking whether a hypothesis is genuinely new and distinguishing discovery through exploration from recall of related knowledge in pre-training data. Its alien-world rules are executable, allowing exact checking, and conflict with familiar knowledge so recall alone cannot solve the tasks.

ExplorationBench contains two sandboxes: Alien Code, with 31 discovery targets and 70 tasks, and Alien Logic, with 24 discovery targets and 70 tasks. Each sandbox provides a flawed manual, task-specific environmental feedback, and a dedicated tool-call schema. Systems use these resources to explore, then solve held-out tasks. The researchers evaluated 10 AI systems and found that the strongest could acquire and apply unfamiliar rules, while performance varied substantially across exploration paths; continued exploration could stall or reverse earlier gains. The authors describe the benchmark as a step toward systems that acquire and apply genuinely new knowledge by exploring unknown environments. The paper was published on September 24 and submitted on September 25, 2026; the page lists Tencent Hunyuan and 9 upvotes.

The project frames scientific discovery as forming hypotheses, designing experiments, and iterating on results. Its test setup makes the worlds' rules checkable while withholding the task solutions, letting researchers assess what systems learn from exploration rather than what they may already recall.

Researchers introduced ExplorationBench, a benchmark for evaluating whether AI systems can discover and apply unfamiliar rules through exploration in verifiable alien worlds. The benchmark addresses two difficulties in assessing scientific exploration: checking whether a hypothesis is genuinely new and distinguishing discovery through exploration from recall of related knowledge in pre-training data. Its alien-world rules are executable, allowing exact checking, and conflict with familiar knowledge so recall alone cannot solve the tasks.

ExplorationBench contains two sandboxes: Alien Code, with 31 discovery targets and 70 tasks, and Alien Logic, with 24 discovery targets and 70 tasks. Each sandbox provides a flawed manual, task-specific environmental feedback, and a dedicated tool-call schema. Systems use these resources to explore, then solve held-out tasks. The researchers evaluated 10 AI systems and found that the strongest could acquire and apply unfamiliar rules, while performance varied substantially across exploration paths; continued exploration could stall or reverse earlier gains. The authors describe the benchmark as a step toward systems that acquire and apply genuinely new knowledge by exploring unknown environments. The paper was published on September 24 and submitted on September 25, 2026; the page lists Tencent Hunyuan and 9 upvotes.

The project frames scientific discovery as forming hypotheses, designing experiments, and iterating on results. Its test setup makes the worlds' rules checkable while withholding the task solutions, letting researchers assess what systems learn from exploration rather than what they may already recall.