CollabFlow is a system for in LLM-based multi-agent systems. It uses a trainable Collab-Director to construct teams of agents, a frozen executor to run them, and outcomes from each round retrain the director. Two key components are Evidence-Conditioned Communication, where a receiver adopts a differing answer only if the sender's evidence is stronger by a margin, and Collaborative Trajectory Balance (CTB), a flow-based objective that credits each team across its construction orders and targets a reward-proportional distribution over teams, keeping multiple good teams in play.
On twelve datasets, CollabFlow outperformed all baselines and continued improving across rounds. The system also bounds how far its self-generated target moves between rounds, a bound that shrinks as records accumulate.
This work is interesting because it addresses limitations of existing , such as pre-defined topologies and error propagation, by enabling the system to learn both who communicates and how, and to maintain diverse effective teams through recursive self-improvement.