Rank, Don't Regress: A 279M Text Ranker Beats 2B and 7B VLMs on Referential Urdu GUI Grounding
arXiv cs.CL · 16 Nov 2026, 03:26 AEST
A 279M-parameter text ranker outperforms larger vision-language models on Urdu GUI grounding tasks. The ranker achieved a point-in-box accuracy of 73.4%, significantly higher than its competitors.
Key facts
- UI grounding asks a model to find the screen element that a user instruction refers to.
- We built a frozen test set of 304 instructions from six Urdu news sites; 149 of them (49%) are referential and never quote the full target text.
- The ranker scores 73.4% point-in-box accuracy, against 29.6% for UI-TARS-2B and 23.7% for Qwen2.5-VL-7B, both zero-shot.
- In a controlled comparison at small scale, with training data drawn from the same pages, a coordinate-regression model of similar size scores 1.6%.
- The ranker's errors concentrate in two places: elements that share identical text, and ordinal instructions such as 'the fourth link'.
Read aloud
A 279M-parameter text ranker outperforms larger vision-language models on Urdu GUI grounding tasks. The ranker achieved a point-in-box accuracy of seventy three point four percent, significantly higher than its competitors. UI grounding asks a model to find the screen element that a user instruction refers to. We built a frozen test set of three hundred four instructions from six Urdu news sites; one hundred forty nine of them forty nine percent are referential and never quote the full target text. The ranker scores seventy three point four percent point-in-box accuracy, against twenty nine point six percent for UI-TARS-2B and twenty three point seven percent for Qwen2.five-VL-7B, both zero-shot. In a controlled comparison at small scale, with training data drawn from the same pages, a coordinate-regression model of similar size scores one point six percent. The ranker's errors concentrate in two places: elements that share identical text, and ordinal instructions such as 'the fourth link'.
Sources in this cluster
- Rank, Don't Regress: A 279M Text Ranker Beats 2B and 7B VLMs on Referential Urdu GUI Grounding
arXiv cs.CL · 16 Nov 2026, 03:26 AEST