<2-3 sentences summarising what the project addresses and solves, the target audience, and the benefit.> SE-Agent is a self-evolution framework for LLM code agents, enhancing multi-step reasoning by enabling trajectory-level information exchange via Revision, Recombination, and Refinement. This approach expands the search space and escapes local optima, making it ideal for researchers and engineers tackling complex autonomous tasks, particularly in software engineering, where it achieves SOTA performance on SWE-bench Verified.

How It Works

SE-Agent enhances LLM agent problem-solving through three core self-evolution operations. Revision uses failure-driven reflection to generate architecturally orthogonal strategies, addressing fundamental approach limitations. Recombination synthesizes knowledge by merging high-performing segments from different trajectories, creating synergistic effects. Refinement optimizes promising trajectories by removing redundancies and enhancing efficiency, guided by collective exploration history to mitigate blind spots. This iterative process allows agents to transcend individual trajectory limitations.

Quick Start & Requirements

• Primary install / run command: Clone the repository (git clone https://github.com/JARVIS-Xs/SE-Agent.git), navigate into the directory, and run pip install -e ..
• Non-default prerequisites and dependencies: Python 3.12 is recommended. API keys for DeepSeek, OpenAI, or Anthropic must be configured via a .env file (e.g., echo "DEEPSEEK_API_KEY=your_key_here" > .env).
• Estimated setup time or resource footprint: Not explicitly stated, but a demo mode is available for testing without API calls (python SE/basic_run.py --mode demo).
• Links: Detailed setup and configuration instructions are available in instruction.md.

Highlighted Details

• Achieves State-of-the-Art (SOTA) performance with 80% Top1 accuracy on the SWE-bench Verified benchmark.
• Papers accepted to NeurIPS 2025: RepoMaster as a Spotlight presentation and SE-Agent as a Poster.
• Part of an ecosystem including RepoMaster and GitTaskBench, focused on complex real-world tasks.
• Features intelligent trajectory processing and compression, reducing .tra file sizes by up to 80%.

Maintenance & Community

• Developed by the QuantaAlpha Team Research Team.
• Community support and issue tracking are available via GitHub Issues and Discussions.
• Contact: quantaalpha.ai@gmail.com.
• Homepage: https://quantaalpha.github.io.

Licensing & Compatibility

• License type: MIT License.
• Compatibility notes: The MIT license is permissive and generally suitable for commercial use and integration into closed-source projects.

Limitations & Caveats

The provided README does not explicitly detail any limitations, alpha status, known bugs, or unsupported platforms.