Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2609.38662.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

CollabFlow proposes a recursive self-improvement system for LLM-based multi-agent collaboration. A trainable Collab-Director assembles agent teams; edges carry Evidence-Conditioned Communication protocols, and a Collaborative Trajectory Balance objective credits teams across construction orders. Evaluated on 12 datasets, it outperforms baselines and improves across rounds.

Concrete architecture for learned multi-agent collaboration with evidence-gated communication and a flow-matching objective. Code is published; single-paper claim across 12 datasets.

Cited by

No citations on record.