Instance-level Randomization Improves Stability of LLM Evaluations

Instance-level randomization (ILR) averages multiple runs per test case, cutting variance and using less than half the compute of fixed-setting benchmarks. Read more: getnews.me/instance-level-randomiza... #instancelerandomization #llmevaluation