CollabFlow: Recursive Self-Improvement of Agent Collaboration

DOI10.48550/arXiv.2609.38662MaRDI QIDQ7360765arXivFDO

Authors Mingda Zhang, Junming Zhang, Yue Dai, Zijia Wang, Xiao-Ying Tang, Xiao Huang, Qiang Huang, Hanwen Zhang, Xiaoying Tang

Publication date 29 September 2026

Published in arXiv

Tldr We propose a recursive self-improvement system of learned agent collaboration where a director constructs teams of agents based on their past performance, evidence-conditioned communication protocols are learned in each round, and a flow-based objective encourages the emergence and retention of multiple effective teams.

Full work available at URL https://doi.org/10.48550/arXiv.2609.38662

arXiv classification

cs.AIcs.MA

Cites work

• An invariant form for the prior probability in estimation problems
• Modeling by shortest data description
• Reaching a Consensus
• The serializability of concurrent database updates
• When Evidence Shapes Collaboration: Knowledge-Conditioned Topology Generation for Multi-Agent Systems

This page was built for publication: CollabFlow: Recursive Self-Improvement of Agent Collaboration

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7360765)