Study Finds LLM Self-Consistency and Cross-Model Agreement Are Unreliable Proxies for Accuracy

Jul 10, 2026

A large-scale study involving 53 models and 265,000 samples finds that agreement among LLM judges or within a model's own outputs is only a weak predictor of correctness (correlation rho 0.20-0.59). Frontier models were especially over-confident, agreeing on 77% of GPQA cases but being wrong on 48% of those. The authors conclude that self-consistency is a conditional proxy, not a standalone confidence score.

Why it matters: This challenges the common assumption in enterprise AI evaluation that consistency implies correctness, urging caution in using LLM-as-judge ensembles for high-stakes decisions.

Full story at: arXiv AI/ML ↗