This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
This paper proposes SE-Agent, a novel framework for optimizing the problem-solving process (interaction paths) of agents based on large-scale language models (LLMs). We highlight the inefficiencies of existing methods, such as MCTS, due to interdependencies and a lack of diverse search space. SE-Agent iteratively optimizes the problem-solving process in a self-evolutionary manner through three operations: modifying, recombining, and improving existing paths. This allows it to explore diverse solution paths and mitigate the impact of inefficient paths, thereby improving performance. Experimental results using SWE-bench Verified demonstrate that our approach achieves state-of-the-art performance, achieving up to 55% performance gains on five robust LLMs.
Takeaways, Limitations
•
Takeaways:
◦
A novel approach to optimizing the problem-solving process of LLM-based agents.
◦
Expanding the search space and improving performance through a self-evolutionary framework.
◦
Efficient learning through reuse of existing routes
◦
Validated practicality and achieved excellent performance (up to 55% performance improvement) by solving actual GitHub issues.
◦
Improving accessibility through open source disclosure
•
Limitations:
◦
Further research is needed to determine the generalizability of the proposed framework.
◦
Since the evaluation results are for a specific domain (GitHub issue), verification of scalability to other domains is required.
◦
Potential increase in computational cost (repeated computation of self-evolution process)
