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.
Agent KB: Leveraging Cross-Domain Experience for Agentic Problem Solving
Created by
Haebom
Author
Xiangru Tang, Tianrui Qin, Tianhao Peng, Ziyang Zhou, Daniel Shao, Tingting Du, Xinming Wei, Peng Xia, Fang Wu, He Zhu, Ge Zhang, Jiaheng Liu, Xingyao Wang, Sirui Hong, Chenglin Wu, Hao Cheng, Chi Wang, Wangchunshu Zhou
Outline
This paper points out the limitations of current AI agents, which cannot effectively learn from their mutual problem-solving experiences or leverage past success experiences to self-reflect and correct errors in new tasks. To address this, we propose Agent KB, a shared knowledge base that captures both high-level problem-solving strategies and detailed execution experiences. Agent KB implements a novel teacher-student dual-stage search mechanism, where the learning agent searches workflow-level patterns for strategic guidance, and the teacher agent identifies detailed execution-level patterns to improve them. This hierarchical approach allows the agent to integrate diverse strategies from external sources to break free from limited inference paths. Evaluation results on the GAIA benchmark and SWE-bench code repair tasks show that Agent KB improves the success rate by up to 6.06% in pass@1, and achieves an 8.67% improvement (from 23% to 31.67%) in pass@1 for o3-mini in SWE-bench.
Takeaways, Limitations
•
Takeaways:
◦
We present Agent KB, an effective mechanism for knowledge transfer between AI agents.
◦
Improving problem-solving strategies and practices through hierarchical teacher-student learning.
◦
Empirically demonstrating performance improvements on GAIA and SWE-bench benchmarks.
◦
Integrating different problem-solving strategies to improve the agent's reasoning ability.
•
Limitations:
◦
Further research is needed on the scalability and generalization ability of Agent KB.
◦
Need to verify the applicability of Agent KB to various domains and tasks.
◦
Since these are evaluation results for specific benchmarks, it is difficult to guarantee generalized performance.
◦
Lack of clear explanation of the reliability and selection criteria of teacher agents.
