Mixture of Roles composes multiple agent personas into one steering vector for single-turn inference

arXiv·medium signalXBlueskyLinkedInCopy link

Posted 27 August, MoRe attacks the cost side of multi-agent specialization: single-agent personas impose one fixed specialization, while multi-agent systems get dynamic multi-perspective reasoning only through multi-turn interaction that inflates context length and inference cost. MoRe learns a diversified codebook of steering vectors, each encoding a latent role, and a query-aware router fuses the codebook into one composed vector applied to a frozen backbone. Training is a three-stage SFT curriculum plus GRPO post-training with the backbone untouched, so the claim is multi-perspective specialization at single-agent, single-turn cost.