BestRand proposes stable, low-cost benchmark compression method, extended to VLA evaluation

BestRand is a stable and low-cost benchmark compression method that reduces the long-tail risk of high errors in random sampling. It was validated on more than 70 text, multimodal, and agent benchmarks and extended to VLA evaluation. Compared with competing methods, it reduces high-error risk by 50.2% and estimation error by 16.4% on average, while construction cost is only 51.3%. The authors also derived an analytical approximation of the error distribution when estimating full-benchmark scores from a subset, noting that random sampling yields small errors in most cases but has rare large errors.