DARC-Sched: A Data-Aware Multi-Agent Large Language Model Framework for Scalable Software Vulnerability Detection and Repair with Optimized Code Data Scheduling

Abstract

Software vulnerability detection and automated repair have become critical for ensuring the security and reliability of large-scale software systems, particularly in modern development environments where massive code repositories and continuous integration pipelines significantly increase the complexity of vulnerability management. However, existing approaches based on Large Language Models (LLMs) often suffer from limited scalability and inefficient code retrieval when operating on large-scale code corpora, leading to degraded detection accuracy and high inference latency. To address these challenges, this paper proposes DARC-Sched (Data-Aware Reasoning and Collaborative Scheduling framework), a Data-Aware Multi-Agent Large Language Model framework that integrates multi-agent collaborative reasoning with optimized code data scheduling mechanisms for scalable vulnerability detection and repair. Specifically, DARC-Sched employs multiple specialized agents to perform vulnerability localization, evidence retrieval, verification, patch generation, and security validation, while a hierarchical data scheduling module dynamically prioritizes and caches vulnerability-relevant code segments based on learned risk-aware access patterns. Experimental results on benchmark datasets demonstrate that DARC-Sched achieves a 2.9% improvement in F1-score over strong retrieval-augmented baselines, reaching 91.9% F1-score, while reducing average retrieval latency by approximately 33% (from 295 ms to 198 ms). These results validate that integrating data-aware scheduling with multi-agent reasoning significantly improves both detection accuracy and system efficiency in large-scale software vulnerability analysis