ExplorationBench comes from Tencent Hunyuan with Fudan and Tsinghua universities. Rules are executable, so an answer is checked rather than judged.
Benchmark
ExplorationBench
Whether an AI system can discover rules it has never seen by running its own experiments. The tasks live in two constructed worlds whose rules deliberately conflict with familiar knowledge, so recalling training data does not help. AlienCode is a small calculation language with 31 discovery targets and 70 tasks; AlienLogic has 24 inference rules and 70 tasks.
- Score
- %
- Last checked
- October 1, 2026
Leaderboard
No verified results recorded yet.
Methodology
Every rule is executable, so an answer can be checked exactly rather than judged. A system probes the world, observes the result, forms a hypothesis and is scored on tasks that require the discovered rule. The maintainers also run a control in which a system replays probes chosen by another system, which separates the value of the information from the value of having chosen to seek it.
Important limitations
The worlds are synthetic and small, and success on an invented rule system is not direct evidence about scientific discovery in the physical world. The benchmark is published by one of the labs whose models it evaluates, and the discovery targets are fixed, so scores will become less meaningful once the tasks are in training data.