HiFiRepoQA: DAG-Structured Multi-Agent Framework for Repository-Level QA

ACL ARR 2026 January Submission9631 Authors

Keywords: Large Language Model, Multi-Agent System, Repository-level Question Answering

Abstract:

Understanding and using large software repositories is increasingly difficult as modern projects grow in scale, span multiple programming languages, and evolve rapidly. Developers frequently raise repository-level questions that require integrating information scattered across code files, documentations, and configurations. Existing approaches struggle with challenges such as linear reasoning strategies that cannot model partially ordered knowledge dependencies and weak handling of noisy context. To address these challenges, we propose HiFiRepoQA, a repository-level question answering framework based on structured task decomposition and controlled information acquisition. It models repository-level questions in a directed acyclic graph (DAG), enabling parallel execution of independent goals while isolating irrelevant context. To facilitate realistic evaluation, we further construct HiFiRepoQA Bench, a high-fidelity benchmark with 526 real user questions collected from GitHub repositories. Experimental results show its advantages over open-source and commercial baselines.
We further evaluate its usefulness by submitting generated answers to GitHub, where 9 questions receive positive feedback or are marked as closed due to our response.

Paper Type: Long

Research Area: AI/LLM Agents

Research Area Keywords: LLM agents, multi-agent systems, planning in agents

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 9631