CMU × Meta's HANDRAISER Cuts Multi-Agent Communication Cost 32.2% by Learning to Interrupt

lileics · x · 2026-09-23

A CMU × Meta FAIR paper accepted at CoLM 2026 flips the script on multi-agent communication efficiency: instead of compressing speaker messages, let listeners interrupt. Naive interruption makes LLMs overconfident and cut in too early, so HANDRAISER learns to predict the right moment based on estimated future reward and communication cost. On 2-agent pictionary, 3-agent scheduling, and 3-agent debate tasks, it cuts communication cost 32.2% on average (24.3–48.9% with Llama-3.1-8B listeners) with equal or better task performance, generalizing to unseen GPT-4o speakers without fine-tuning.

Related event: CMU and Meta Teach Listeners to Interrupt, Cutting Multi-agent Communication Cost 32%(2 posts)→

Original post →

More from coding & agent

coding & agent channel →