Reproduce “Beyond Local Code Optimization: Multi-Agent Reasoning for Software System Optimization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Huiyun Peng, Parth Vinod Patil, Antonio Zhong Qiu, George K. Thiruvathukal, James C. Davis
An architecture-aware multi-agent framework uses static program representations and LLM-based reasoning to identify and implement performance improvements that span services, calls, synchronization, and resource usage. In a TeaStore microservice evaluation, the generated changes increased throughput by 36.58% and reduced average response time by 27.81% while passing the application’s JUnit tests, though broader comparisons and repeated-trial analysis remain open.
Key insights
- Static representations serve as an intermediate layer that gives LLM agents system context beyond the local code surrounding a suspected hotspot.
- The prototype identified optimizations involving shared HTTP clients, lock-contention reduction, and reuse of Jackson ObjectMapper instances—changes whose effects can accumulate along service request paths.
- The current evidence supports feasibility rather than generality: the evaluation used one Java benchmark, lacked ablations and baseline-system comparisons, and did not fully integrate validation into the automated loop.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.