VeriBRT: Plan-Guided, Evidence-Based Automated Bug Reproduction

Status: Manuscript under review · PDF: Download

Authors

Yeheng Chen*, Runguo Li*, Yuling Shi, Zimu Wang, Xiaodong Gu, Yanfu Yan.

* Equal contribution.

Abstract

Bug-reproducing tests provide executable evidence for localizing, understanding, and validating software defects. VeriBRT represents issue-report behavior as a persistent Two-Axis Plan: a Checklist specifies where a generated test should reach, while an Intentlist specifies what behavior it should verify. It combines this representation with a reliability-aware evidence strategy and incorporates localization coverage into the generation loop as a deterministic validity criterion. On SWT-Bench Verified, VeriBRT reproduces 341 of 433 issues under the reported evaluation setting, outperforming the strongest same-model baseline by 9.7 percentage points.

Full text: Download the PDF

Recommended citation: Yeheng Chen*, Runguo Li*, Yuling Shi, Zimu Wang, Xiaodong Gu, and Yanfu Yan. (2026). "VeriBRT: Plan-Guided, Evidence-Based Automated Bug Reproduction." Manuscript under review. (*equal contribution)
Download Paper | Download Bibtex