Source note
Know Before Fix: QA-Driven Repository Knowledge Acquisition for Software Issue Resolution
Software Foundation ModelCode IntelligenceAutomated Software ProductionMulti Agent Software EngineeringRepository Knowledge
Summary
ACQUIRE improves automated software issue resolution by acquiring repository knowledge through targeted question-answering before patch generation. On SWE-bench Verified, it raises Pass@1 by up to 4.4 percentage points over Mini-SWE-Agent while adding modest cost and time.
Problem
- Coding agents often lack repository-specific knowledge about cross-module dependencies, API contracts, data flow, and external protocols, which leads to incorrect patches.
- Existing pre-repair methods focus on locating suspicious code or generating broad summaries without first identifying the agent's specific knowledge gaps.
- These failed, knowledge-deficient attempts consume over four times the token cost and nearly twice the execution steps of successful resolutions.
Approach
- ACQUIRE separates repository understanding from repair into two stages: knowledge acquisition followed by patch generation.
- A Questioner generates targeted, non-redundant questions across four categories: Mechanism & Behavior, Design & Usage, Locating & Structure, and Ecosystem & Standards.
- Independent read-only Answerer agents explore the repository in parallel and produce evidence-grounded answers that cite files, functions, and code behavior.
- A Resolver receives the issue description and the collected QA pairs before starting its normal navigation, editing, and testing loop.
Results
- On 500 SWE-bench Verified issues, ACQUIRE reached 62.2% Pass@1 with GPT-5-mini versus 58.4% for Mini-SWE-Agent, a gain of 3.8 percentage points; with DeepSeek-V3.2, it reached 70.8% versus 66.4%, a gain of 4.4 points.
- ACQUIRE cost an average of $0.054 and 302 seconds per GPT-5-mini instance, and $0.073 and 1,042 seconds per DeepSeek-V3.2 instance. Richer baselines such as SWE-Debate cost $0.738 and $0.382, with times of 1,552 and 2,517 seconds respectively.
- Human review found 230 of 232 QA pairs supported by repository evidence, or 99.1%; only 2 pairs contained ungrounded central claims.
- QA injection reduced mean repair rounds by 7.1% across all 500 issues and by 17.1% on the 44 instances that changed from failure to success.
- In an oracle experiment on 116 previously failed SWE-bench Lite instances, a single privileged QA pair recovered 26 cases, showing that targeted repository knowledge can unlock fixes that the base agent could not produce.