ReBug Reproduces Real Web GUI Bugs From Natural-Language Reports at 49.96% Success Across 667 Reports
Reproducing browser bugs from reports is hard mainly because reports omit prerequisites like dependencies and input files, and prior techniques target code units or mobile apps without end-to-end visual execution. ReBug splits the work into a preparation stage that reconstructs missing prerequisites and plans, and an execution stage that drives a real browser with tool-mediated interactions, structured page-state summaries, and validation against report-derived expectations. On 667 real reports from four open-source web apps it hits 49.96% average reproduction success rate, 74.96% task completion, and 86.54% action execution success — with historical replay showing successful procedures often re-expose the original bug on restored buggy versions.
↳ Follow the thread