Scientists discover laws of nature by running experiments and forming hypotheses. AI agents are now being used to speed up this process, but current tests often confuse programming ability with true scientific reasoning. They also don't check if an agent can handle situations where the established rules are changed.
To address this, the researchers built SciGame, a that takes real scientific laws and creates both the original version and a modified version that respects the same physical dimensions. AI agents must either choose the correct law from a list or discover it from scratch. The benchmark also compares learning from experiments versus relying solely on prior knowledge.
The key result: AI agents performed well on familiar laws but failed on the modified ones. Even when they provided convincing explanations, their underlying equations were often wrong. This shows that current AI systems are not yet ready to reliably discover new scientific laws, especially when prior knowledge might mislead them.