Research
Hugging Face paper proposes BestRand for low-cost, stable benchmark compression
Researchers derive an error model for estimating full benchmark scores from subsets and introduce BestRand, which reportedly cuts high-error risk by 50.2% and needs 51.3% of the construction cost. Validated on 70+ text, multimodal and agent benchmarks, with code publicly available.
Every outlet that ran it 1 article
Hugging Face PapersResearchLead
Random Is Almost Enough? Stable and Low-Cost Benchmark Compression via BestRand