Know Before Fix: QA-Driven Repository Knowledge Acquisition for Software Issue Resolution
agenticmulti-agentresearch
Summary
This paper introduces ACQUIRE, a QA-driven framework for automated software issue resolution. It decouples repository knowledge acquisition from patch generation using a two-stage, multi-agent process where a 'Questioner' and 'Answerer' agent first collaborate to build a structured understanding of the codebase. A 'Resolver' agent then uses this knowledge to generate a more accurate fix, improving performance on the SWE-bench benchmark by up to 4.4 percentage points in Pass@1.
Problem
LLM-based coding agents often fail at software issue resolution due to factual errors caused by an insufficient understanding of the target repository. Existing methods attempt to mitigate this by exploring the repository before generating a fix, but their fix-driven strategies often yield imprecise context because they don't explicitly identify the agent's specific knowledge gaps.
Method
ACQUIRE employs a two-stage process that mimics how experienced developers first understand unfamiliar code before attempting a fix. In the first stage, a 'Questioner' agent poses targeted questions to identify knowledge gaps, and an 'Answerer' agent autonomously explores the repository to provide evidence-grounded answers, building a structured knowledge base. In the second stage, a 'Resolver' agent leverages this QA-derived knowledge to generate an informed code patch.
Details
ACQUIRE was evaluated on the SWE-bench Verified dataset, where it demonstrated a consistent performance improvement over representative pre-repair methods by explicitly addressing agent knowledge gaps before patch generation.
Metric
Result
Pass@1 on SWE-bench Verified
Up to +4.4 percentage points vs. baseline
The framework transforms implicit knowledge gaps into an explicit, factually reliable understanding of the repository context.
This pre-repair knowledge acquisition stage accelerates knowledge-intensive repair tasks and leads to more accurate resolutions.
The performance gains are achieved with only modest additional computational cost and time overhead.
What's new
The primary innovation is the explicit decoupling of repository knowledge acquisition from patch generation. Unlike prior methods where a single agent explores and fixes concurrently, ACQUIRE introduces a structured, multi-agent QA phase (Questioner and Answerer) dedicated solely to building a reliable understanding of the codebase before a separate Resolver agent attempts the fix.
Conclusion
The QA-driven approach of first identifying and filling an agent's knowledge gaps before attempting a fix is an effective strategy for automated software issue resolution. By transforming implicit uncertainty into explicit, grounded knowledge, ACQUIRE enables LLM agents to generate more accurate patches. This decoupling of comprehension and repair consistently improves performance with manageable overhead, suggesting a promising direction for developing more robust coding agents.