StarVLM ReRank: Better UI Grounding via Enhanced Visual Input and Element Position Perception

Résumé

UI grounding is a fundamental task for enterprise workflow automation. This task maps natural language instructions to precise pixel coordinates in graphical interfaces. Current vision-language models fail to achieve the pixel-perfect accuracy required for production deployment, limiting their adoption in business-critical applications. We identify three architectural limitations: query-agnostic visual processing, imbalanced spatial position encodings, and implicit coordinate regression. To address these, we propose a unified framework called StarVLM ReRank, combining: (1) Query-Conditioned REOrder (QC-REOrder) that dynamically reorders image patches based on textual intent, (2) Interleaved MRoPE (I-MRoPE) that provides balanced frequency coverage across spatial dimensions, and (3) RULER tokens that transform coordinate prediction from regression to retrieval. Our approach reframes UI grounding as a spatial reasoning task rather than standard visual grounding. Extensive evaluation on ScreenSpot, Multimodal-Mind2Web, AndroidControl, and OmniACT benchmarks demonstrates consistent improvements across web, mobile, and desktop platforms. By treating pixel-level precision as a first-class architectural concern, our method enables reliable UI automation for enterprise customers demanding robust, resolution-agnostic performance.