Reward-guided graph generation cuts multi-agent communication tokens 20.5% at equal accuracy

arXiv·medium signalXBlueskyLinkedInCopy link

RGA-Designer (arXiv 2608.20099, August 20) trains a reward model that scores both task correctness and structural compactness, then fine-tunes a pretrained graph generator against it to design communication topologies for LLM multi-agent systems. It holds task accuracy level with the prior ARG-Designer while cutting token consumption by an average of 20.5%. For anyone running fan-out agent teams where inter-agent chatter dominates the bill, topology is a cost lever that is usually left on the default.