Reproduce “SWE-Replay: Efficient Test-Time Scaling for Software Engineering Agents” 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
Yifeng Ding, Lingming Zhang
Researchers at the University of Illinois Urbana-Champaign developed SWE-Replay, an algorithm that enhances the efficiency and performance of test-time scaling for LLM-based software engineering agents. It achieves this by recycling previously sampled trajectories and intelligently branching from them, leading to reduced computational costs by up to 17.4% and improved bug-fixing capabilities, with a 22.6% performance gain on the SWE-Bench Multilingual benchmark.
Key insights
- Recycling and intelligently branching from prior agent experience can significantly reduce computational costs and improve solution quality without relying on external, potentially noisy LLM-as-a-Judge mechanisms.
- Abstracting agent states by the set of repository files explored before a step effectively groups similar states, promoting diverse exploration of the codebase, particularly less-visited 'long tail' files.
- Prioritizing branching from reasoning-intensive steps, identified by the structural proxy of paragraph count in reasoning content, allows agents to re-evaluate critical decision points and escape local optima.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.