Multi-agent reasoning systems distribute information across agents via retrieval and tool use, making communication essential. But not every message is useful, and excessive exchange can inflate token costs and introduce noise that degrades performance.
The new paper, posted on arXiv, introduces a neuro-inspired approach called inhibitory message passing. Borrowing from the way biological neural circuits suppress weak or irrelevant signals, the method selectively dampens unhelpful agent-to-agent messages.
The authors frame this as a way to achieve both higher accuracy and lower token consumption. While the abstract is truncated, the title and opening suggest a practical direction for making multi-agent systems more efficient without sacrificing reasoning quality.