One Model, Many Minds: Unlocking Multi-Agent Synergy in a Single Agent via Mixture of Roles
3.40T1 sourcearXiv cs.MA
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2608.27338.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryPaper proposing Mixture of Roles (MoRe), which composes multiple specialized steering vectors into one query-adapted vector for single-LLM inference. Reports 2.2% average gain over single-agent baselines and 20x token cost reduction versus multi-agent systems, trained via a three-stage SFT curriculum plus GRPO with a frozen backbone.
Why it mattersReported 20x token reduction against MAS at comparable performance gives practitioners a concrete data point when choosing between single-agent and multi-agent designs.
Cited by
No citations on record.