GLiDE, ranked 7th in judgment reliability despite leading in System One performance

GLiDE led all System One models in overall performance on the SO-Arena benchmark but ranked only seventh of 18 models on judgment reliability, according to a paper submitted to arXiv on June 11. The study compared 21 models across 2,919 tasks spanning 11 domains, and two System Two-based models plus one model that automatically decides whether to apply additional reasoning were included in the evaluated set. System One models produce structured, machine-readable outputs such as per-option scores or probabilities rather than generating long-form answers. The gap matters because it shows top aggregate scores do not guarantee dependable judgments, a key requirement for models whose outputs feed directly into software pipelines. What to watch next is whether GLiDE and rival System One models can close the reliability gap without surrendering their performance lead.