ICML 2025 study of 350+ models found highly correlated errors across providers

deliprao · x · 2026-09-26

Delip Rao notes the correlated-errors problem is "obvious" in hindsight, linking it to prior ICML 2025 findings: Kim et al. examined more than 350 models and found larger, more accurate models make highly correlated errors, even across providers — background for his critique of Jev-like judge models.

Related event: Experiments Show LLM Judge Errors Are Highly Correlated, Limiting Cascades(3 posts)→

Original post →

More from Research

Research channel →