Authors Kun Zeng, Yu Huo, Siyu Zhang, Siyue Chen, Xiao-Ying Tang, Yuecheng Zhuo, Yuquan Lu, Haoyue Liu, Xiaoying Tang
Publication date 18 June 2026
Tldr This paper introduces SIGMA where agents are constructed from task-conditioned bundles of reusable skills within graph-based systems rather than fixed nodes to improve generalization across tasks. It predicts skill incidence matrices to compose embeddings and decode topologies, achieving best benchmark performance while demonstrating robustness against unseen libraries compared to existing baselines.
This page was built for publication: SIGMA: Skill-Incidence Graphs for Compositional Multi-Agent Design
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7269890)
