Jul 2026
Benchmarks Are Not Monolithic: Sample-Level Auditing and Orchestration for LLM Evaluation
A dataset-centric meta-evaluation framework that audits benchmark datasets at the sample level along five latent dimensions, providing a principled methodology for analyzing and re-composing existing benchmarks to better reflect diverse evaluation needs.
P. D. Siedler, Jordan Sassoon
· arXiv.org · 1 citation