SE-Agent

SE-Agent is a self-evolution framework designed for LLM-based code agents that enables information exchange between multiple reasoning paths to solve complex software engineering tasks. The framework implements three core operations—Revision, Recombination, and Refinement—that work together to expand search space and escape local optima. Revision analyzes failed trajectories and generates architecturally different problem-solving approaches, Recombination intelligently merges strengths from multiple solution paths, and Refinement optimizes promising trajectories by eliminating redundancies. SE-Agent achieves state-of-the-art performance on SWE-bench Verified with 80% success rate, demonstrating significant gains across various open-source and closed-source LLMs. This framework is ideal for developers building autonomous agents that handle multi-step reasoning and complex software engineering tasks.

[{"step":"Define the task and constraints clearly. Use the prompt template to specify the [TASK_DESCRIPTION], [SPECIFIC_ISSUE], [ALTERNATIVE_SOLUTIONS], and [PERFORMANCE_METRICS] to guide the SE-Agent's evolution process.","tip":"Be specific about the constraints (e.g., time complexity, space complexity, or correctness metrics) to ensure the evolved solution meets your requirements."},{"step":"Provide the initial solution or code snippet to the SE-Agent. This serves as the starting point for the Revision phase, where the agent identifies issues or inefficiencies.","tip":"Include comments or explanations in the code to highlight areas of concern or potential improvements."},{"step":"Specify the alternative solutions or approaches to recombine. These can be high-level ideas (e.g., iterative vs. recursive) or specific snippets from other sources.","tip":"Use the SE-Agent to explore diverse approaches, even if they seem unconventional, to expand the search space and avoid local optima."},{"step":"Run the SE-Agent process iteratively. Review the evolved solution, identify further issues, and repeat the process until the solution meets all constraints and metrics.","tip":"Track the performance metrics (e.g., time, space, correctness) at each iteration to measure progress and guide refinements."},{"step":"Validate the final evolved solution. Test it against edge cases, large inputs, and real-world scenarios to ensure robustness and correctness.","tip":"Use tools like pytest, timeit, or memory profilers to verify the solution's performance and correctness."}]