UnpredictaBench: A Benchmark for Evaluating Distributional Randomness in LLMs

In three linesUnpredictaBench evaluates LLMs' ability to capture true underlying distributions across 448 problems (statistical distributions, stochastic programs, natural scenarios). The KS@N metric uses the Kolmogorov-Smirnov test. No model exceeds 40% at KS@100, revealing a major gap in distributional sampling capability.

Summary generated by Claude — human-verified