Scaling evaluation of large models using human and AI audit strategies
From Self-Confidence to Statistical Accuracy: Scaling Human-in-the-Loop Evaluation with Direct and LLM-Assisted Auditing
Computation and Language
Summary
Large language models often guess their own confidence incorrectly, which makes checking their accuracy tricky. The authors found ways to still use these confidence estimates to help guide human checks more efficiently. They also studied using AI models as automated judges to reduce human effort but found this only works well under certain conditions. Their analysis explains when each approach saves time and cost while still giving good accuracy estimates.
What this means in practice
- •For machine learning engineers: Use confidence-based sampling to audit model accuracy more efficiently without needing perfect confidence calibration.
- •For automated testing teams: Incorporate AI judges to assist human auditing when model calibration and cost conditions make it beneficial.
Authors
Harine Choi, Chenchen Kuai, Zihao Li, Jiwan Jiang, Yunlong Zhang, Zhiwen Fan, Zhengzhong Tu, Yang Zhou
Abstract
For large language models (LLMs) and multimodal large language models (MLLMs), model-reported self-confidence is often miscalibrated, yet it may still be useful for efficient evaluation. At the same time, LLM-as-a-judge offers a scalable alternative to direct human auditing, but its practical benefit depends on calibration quality and inference cost. This paper develops a unified statistical framework for both evaluation routes under alternative sampling strategies. For direct auditing, confidence-stratified sampling is used to recover population accuracy from partially verified outputs without requiring calibrated confidence. Finite-sample uncertainty is characterized, and the analysis shows how highly uneven confidence strata affect audit efficiency and label allocation. For judge-assisted evaluation, the judge is modeled through sensitivity and false-positive rate, yielding an error-corrected estimator with calibration uncertainty. The two routes are then compared through human-label and total-cost efficiency boundaries. Experiments on MMLU-Pro and MMMU-Pro show that miscalibrated self-confidence can still support statistically valid stratified auditing. Under highly imbalanced confidence distributions, proportional allocation substantially reduces estimation error relative to equal-per-stratum allocation. For judge-assisted evaluation, only a small subset of target--verifier pairs satisfies the idealized human-label efficiency condition, and automated inference costs can further limit potential savings. These findings clarify the statistical and economic conditions under which confidence-guided auditing and automated verification are useful.