Research Paper Introduces Zoom Consistency as Free Confidence Metric for Visual Grounding Pipelines

A recent research paper presents "zoom consistency" as an innovative confidence indicator for multi-step visual grounding processes. This geometric metric quantifies the distance between a model's second-step prediction and the crop center within a shared coordinate framework. Unlike conventional confidence metrics such as log-probabilities or token-level uncertainty, zoom consistency operates without the need for calibration and is applicable across various vision-language architectures. The study reveals a correlation between zoom consistency and prediction accuracy in two VLMs: KV-Ground-8B and Qwen3.5-27B. For KV-Ground-8B, the correlation yields AUC = 0.60 with Spearman rho = -0.14 and p < 10^{-6}, while for Qwen3.5-27B, Spearman rho = -0.11 with p = 0.0003. The findings indicate that this metric can effectively estimate step-1 spatial error under ideal conditions. Although multi-step zoom-in pipelines are frequently utilized for GUI grounding, intermediate predictions are often overlooked after coordinate remapping. This paper highlights that these intermediate outputs possess valuable confidence data that can be utilized without incurring extra computational expenses. The correlation remains stable across various models, application types, and operational scenarios, despite its modest magnitude. This research was published on arXiv under the identifier 2604.15376v1 and was introduced as a cross-type abstract.