Mixture of Roles matches multi-agent performance in one turn at 20x lower token cost.

MoRe learns a codebook of steering vectors, each encoding a latent role, then uses a query-aware router to fuse them into a single composed vector for single-turn inference. The backbone stays frozen; training is a three-stage SFT curriculum plus GRPO. Across reasoning and personality benchmarks it beats single-agent baselines by 2.2% on average and matches multi-agent systems at a twentieth of the tokens. If your reason for running a multi-agent pipeline is role diversity rather than parallelism or tool isolation, this is a direct argument you're paying 20x for something a steering vector does. (arXiv)