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 address the problems of existing methods, such as MCTS, which ignore the interdependencies between interaction paths and insufficient search space diversity, resulting in inefficient inference. SE-Agent iteratively optimizes the inference process through three operations: modifying, recombining, and improving existing paths. This allows it to explore diverse solution paths beyond local optima and mitigate the impact of inefficient paths, thereby improving performance. Experimental results using the SWE-bench Verified dataset 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.
◦
Solving the problems of ignoring interdependencies and lacking search space diversity in existing methods (__T325856_____)
◦
Efficient performance improvement and mitigation of inefficient path impact through modification, recombining, and improvement operations.
◦
Achieve cutting-edge performance on SWE-bench Verified (up to 55% performance improvement)
◦
Open source release for improved accessibility
•
Limitations:
◦
The performance improvements of SE-Agent may be limited to a specific dataset (SWE-bench Verified).
◦
Further research is needed on generalization performance to other types of problems or environments.
◦
Further research is needed on parameter tuning and optimization of the three operations (modification, recombination, and improvement).
