Why this matters
LLM-based coding agents often fail because they try to patch before they truly understand a repository. This paper shows that explicitly converting implicit knowledge gaps into evidence-grounded QA items before repair reduces factual errors and speeds up knowledge-intensive fixing.
Key Findings
- Two-stage design: a Questioner formulates targeted knowledge queries and an Answerer autonomously explores the repo to produce evidence-grounded answers; a separate Resolver consumes the QA artifacts to generate patches. This decoupling focuses exploration on actual knowledge gaps rather than blind pre-repair crawling.
- Empirical gain: on SWE-bench Verified, ACQUIRE raises Pass@1 by up to 4.4 percentage points compared to representative pre-repair baselines while incurring only modest additional time and cost.
- Practical implication: turning ambiguous repository context into explicit QA records both accelerates downstream repair steps and improves factual reliability of generated fixes.
Who it's for and trade-offs
Great fit if you build or evaluate LLM-driven repair agents that must avoid factually incorrect edits and operate on moderately complex codebases. The approach is especially useful when repository context is large or noisy and when small improvements in correctness justify extra exploration time.
Look elsewhere if low-latency or zero-overhead patching is the primary constraint: QA-driven acquisition adds exploration steps and may scale less efficiently on extremely large or frequently changing repositories unless paired with efficient indexing and caching.
Where it fits
Positions between naive pre-repair crawling and full retrieval-augmented generation: instead of retrieving broad context, ACQUIRE produces concise, evidence-linked QA that directly addresses the agent's uncertainty, making it complementary to RAG pipelines and other context-collection strategies.