arxivcs.AIcs.CL2026-07-02
Coresets Before Score Sets: Evaluation-Unsupervised Prompt Subset Selection for LLM Benchmarks
Jihan Yao, Gantavya Bhatt, Arnav Das, Peter Jin, Ke Bao, Qiaolin Yu, et al.
We study LLM benchmark coreset selection: selecting a small subset of prompts over multiple benchmarks whose induced model scores and rankings approximate those obtained from the full benchmark suite. In evaluation-unsupervised benchmark coreset selection (our approach), the sele…