BANDMAS: Causality-Inspired Semantic Packet Scheduling for Bandwidth-Efficient Multi-Agent Collaboration
3.40T1 sourcearXiv cs.MA
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2608.00458.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryBANDMAS is a multi-agent collaboration framework that reduces communication overhead by decomposing messages into semantic packets and scheduling transmission via causality-inspired replay valuation, only sending packets whose predicted contribution exceeds resource cost. Tests on Qwen3-4B across SciFact, HotpotQA, and FanOutQA cut application-layer bytes 53.2–77.3%.
Why it mattersConcrete method for cutting bandwidth in LLM multi-agent systems at the packet-semantic level rather than agent or message pruning, with measured numbers across three QA workloads.
Cited by
No citations on record.