4,760 milestones fed in, predicting new discoveries is still at a coin flip level. No matter how strong the mechanistic reasoning is, when it comes to unknown fields without standard answers, it's still clueless.

Stanford/Oxford/Allen AI Institutes have launched the temporal benchmark CUSP to evaluate AI’s ability to predict scientific progress. Tests of GPT-5.4, Claude Sonnet 4.5, DeepSeek R1, and others show strong performance in mechanistic reasoning about the pathways of existing technologies, but whether it can predict new discoveries is nearly random—and there is a systematic lag in the timing of breakthroughs. CUSP uses a time-based knowledge cutoff, compiling the latest developments from Nature/Science. The benchmark covers 4,760 milestones and 17,429 tasks. The conclusion is that, in scientific exploration without standard answers, existing models cannot provide reliable forward-looking judgments.