GUI-Failure-Suite

GUI-Failure-Suite is a task suite for analyzing and benchmarking GUI agents through their failure modes rather than only their success rates.
It is designed to support research on failure taxonomies, process-level evaluation, and robustness of agents acting in realistic desktop and web environments.

Motivation

Existing GUI benchmarks mainly report aggregate metrics such as task success or click accuracy, which makes it hard to understand how and why agents fail.
GUI-Failure-Suite provides carefully constructed tasks that target specific error patterns (e.g., mis-grounding, instruction misunderstanding, recovery failures), enabling more fine-grained analysis and comparison across models.

Contents

• Curated tasks grouped by failure category and scenario (desktop, web, forms, multi-step flows).
• Structured annotations describing the intended trajectory, observable failure patterns, and evaluation hooks per task.
• Split configuration for development (analysis, ablations) and test (reporting) runs.

Intended Use

This suite is intended for:

• Researchers building taxonomies or process-level metrics for GUI agents.
• Practitioners stress-testing existing agents on brittle workflows, edge cases, and recovery behavior.
• Benchmark authors who want to plug failure-aware tasks into their own evaluation pipelines.

Basic Structure

A typical task entry includes:

• Environment specification (OS, app/website, window configuration).
• Natural language goal and any auxiliary instructions/prompts.
• Ground-truth trajectory and failure labels (e.g., wrong widget, premature submit, irreversible state change).
• Metadata such as category, difficulty, and tags.

Getting Started

You can clone the repository or use it directly from the Hub as a dataset/asset in your evaluation pipeline.

git lfs install
git clone https://huggingface.co/datasets/MazenBaShammakh/gui-failure-suite
cd gui-failure-suite