RotDroid recorded more true bugs than either of two comparison methods when each received the same 30-minute testing budget per app. Against DLD, it recorded 300 true bugs with 84.75% precision, compared with 167 true bugs and 76.26% precision for DLD. Against DOC, a DOC-based same-orientation method, it recorded 268 true bugs with 84.28% precision, compared with 139 true bugs and 86.34% precision for DOC. Precision is the share of flagged cases that were counted as true bugs.
The result was split on precision: RotDroid's true-bug count was higher in both comparisons, but DOC's reported precision was slightly higher. These figures compare detection counts and reported precision within a fixed testing window.
The test is built around state equivalence
RotDroid generates and mutates state-preserving action sequences to construct semantically equivalent cross-orientation GUI states and uses RotVL to detect inconsistencies between them. In practical terms, it checks whether the same interface state remains equivalent when an interaction sequence is tested across orientations.
Building RotBench from paired app states
The study's empirical screening started with 6,031 unique closed GitHub reports and reduced them to 898 confirmed rotation-induced GUI issues. Those numbers describe the report-screening stage.
For RotBench, the non-defective-pair collection yielded 1,818 valid pairs from 623 applications. Data augmentation produced 11,233 paired samples.
The final task sets contained 3,620 bug-detection samples from 623 apps and 2,960 classification and localization samples from 530 apps. The project-wise split was 8:1:1 for training, validation and test data, and the detection set used a 1:1 positive-to-negative ratio.
The model comparison
On RotBench's bug-detection task, RotVL-8B reported an F1 score of 85.29%, compared with 70.84% for GPT-5.2. F1 combines precision and recall into one score. On classification, RotVL-8B scored 62.68%, compared with 49.95% for GPT-5.2.
RotVL-8B also reported a higher orientation-localization F1, 79.97% versus 68.13% for GPT-5.2. Orientation localization is the task of indicating where the problem appears on the interface. Its center-point distance was 198 pixels, compared with 276 pixels for GPT-5.2, so the reported location was closer by that measure. RotVL-8B's reported area ratio was 3.17, and the paper reports p < 0.001 for improvements across all evaluated dimensions.
On a natural bug set, RotVL-8B was the best of the listed models, with 70.00% accuracy, 74.81% precision, 70.00% recall and 72.32% F1.
Findings from open and closed apps
In the open-source evaluation, RotDroid identified 78 issues. It submitted 72, while six had already been fixed in newer releases. The paper's RQ4 answer reports 41 confirmed among the 72 submitted.
The paper also gives an overall status breakdown of 47 confirmed, 30 pending and one rejected. That is a different reporting breakdown from the 41 confirmed figure tied specifically to submitted issues.
On closed-source apps, RotDroid identified 22 GUI rotation issues in 20 representative commercial applications, with six confirmed. At the abstract level, the open- and closed-source studies are summarized as 94 previously unknown bugs, with 47 confirmed or fixed by developers.
The materials are available for replication
The supplied record is arXiv version 1 dated 26 Aug 2026. The paper provides a replication package containing code, model and issues. The work is supported by the State Key Laboratory of Complex & Critical Software Environment, or SKLCCSE.
Paper data and sources
Original title: RotDroid: Cross-Orientation State Equivalence Testing for Detecting GUI Rotation Bugs in Android Apps
Authors: Mengdi Qin, Bo Jiang
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text