Saved in:
Bibliographic Details
Main Authors: Wang, Yangyue, Sikka, Harshvardhan, Mathur, Yash, Zhou, Tony, Nyachhyon, Jinu, Guruprasad, Pranav
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2604.14262
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914476761022464
author Wang, Yangyue
Sikka, Harshvardhan
Mathur, Yash
Zhou, Tony
Nyachhyon, Jinu
Guruprasad, Pranav
author_facet Wang, Yangyue
Sikka, Harshvardhan
Mathur, Yash
Zhou, Tony
Nyachhyon, Jinu
Guruprasad, Pranav
contents GUI grounding models report over 85% accuracy on standard benchmarks, yet drop 27-56 percentage points when instructions require spatial reasoning rather than direct element naming. Current benchmarks miss this because they evaluate each screenshot once with a single fixed instruction. We introduce GUI-Perturbed, a controlled perturbation framework that independently varies visual scenes and instructions to measure grounding robustness. Evaluating three 7B models from the same architecture lineage, we find that relational instructions cause systematic accuracy collapse across all models, a 70% browser zoom produces statistically significant degradation, and rank-8 LoRA fine-tuning with augmented data degrades performance rather than improving it. By perturbing along independent axes, GUI-Perturbed isolates which specific capability axes are affected-spatial reasoning, visual robustness, reasoning calibration-providing diagnostic signal that aggregate benchmarks cannot. We release the dataset, augmentation pipeline, and a fine-tuned model.
format Preprint
id arxiv_https___arxiv_org_abs_2604_14262
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle GUI-Perturbed: Domain Randomization Reveals Systematic Brittleness in GUI Grounding Models
Wang, Yangyue
Sikka, Harshvardhan
Mathur, Yash
Zhou, Tony
Nyachhyon, Jinu
Guruprasad, Pranav
Machine Learning
Artificial Intelligence
GUI grounding models report over 85% accuracy on standard benchmarks, yet drop 27-56 percentage points when instructions require spatial reasoning rather than direct element naming. Current benchmarks miss this because they evaluate each screenshot once with a single fixed instruction. We introduce GUI-Perturbed, a controlled perturbation framework that independently varies visual scenes and instructions to measure grounding robustness. Evaluating three 7B models from the same architecture lineage, we find that relational instructions cause systematic accuracy collapse across all models, a 70% browser zoom produces statistically significant degradation, and rank-8 LoRA fine-tuning with augmented data degrades performance rather than improving it. By perturbing along independent axes, GUI-Perturbed isolates which specific capability axes are affected-spatial reasoning, visual robustness, reasoning calibration-providing diagnostic signal that aggregate benchmarks cannot. We release the dataset, augmentation pipeline, and a fine-tuned model.
title GUI-Perturbed: Domain Randomization Reveals Systematic Brittleness in GUI Grounding Models
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2604.14262