GTArena

GTArena evaluates multimodal models on three decisions of testing a mobile app. Each decision has its own question pool with fixed inputs.

Pool	Items	Apps	The model is given	The model answers
Test intention	114	29	the app's name and one screenshot	five checks
Task execution	792	16	the screen, its controls, the task goal and the earlier steps	the next action
Defect judgment	1,858	105	one screen, or a few frames around an action	defect or clean

Item ids are those of the first release. Ids of items removed since then are not reused, so the task-execution and defect-judgment ids have gaps.

Files

test_intention/items.jsonl: id, app, package, image, screenshot_source (issue_report or reproduction), issue_url, and defect with the expected behavior, the actual behavior and the symptom. The judges compare the checks with defect; the model never sees it. In actual, S is the release with the defect and S′ the release that fixed it.

task_execution/items.jsonl: id, package, image, task, history (the earlier steps), controls, prompt (as shown to the model), exact_match_scored, and answers, the accepted next actions. An answer names its action and, depending on the action, its control_id, text, key, direction or app.

defect_judgment/items.jsonl: id, package, images, image_labels (how the prompt names each image, in order), action_context, prompt, label (defect or clean), defect_type, category (display or interaction), source and issue_url. A defect's source is issue_report, fault_injection or ui_mutation. A clean item's source is normal_use, issue_screenshot, mutation_original, injection_original or fixed_build.

prompts/: test_intention.txt, filled with app_name and, for screenshot_availability, "One decision-point screenshot is attached." (or "No screenshot is supplied." without the image); test_intention_judge.txt, filled with the item's defect and the model's checks; task_execution.txt; defect_judgment.txt.

Scoring

• Test intention: an item is covered when two of the three judges (GPT-5.5, Gemini-3.1-Pro and DeepSeek-V4-Pro) say that the same check would expose its defect. A model's coverage is the share of the 114 items it covers.
• Task execution: an answer matches when its action type and its target are both correct. Unanswered and malformed answers count as wrong. Exact match is computed over the 788 questions with exact_match_scored.
• Defect judgment: accuracy, defect recall and specificity over all 1,858 items.

License

CC BY 4.0. Screenshots taken from issue reports keep the link to their issue in issue_url.