One Model, Many Minds: Unlocking Multi-Agent Synergy in a Single Agent via Mixture of Roles

Noosaga Placements

• Natural Language Processingsubfield90%
  The work focuses on LLM behavior, role prompting, and reasoning benchmarks, all core NLP topics.
  Specializing Large Language Models (LLMs) toward distinct abilitiesExperiments across reasoning and personality benchmarks
• Attention Mechanisms and Transformersframework80%
  The method relies on transformer-based LLMs and their hidden representations, which are central to attention mechanisms and transformers.
  steering vector into the hidden layersbackbone LLM remains frozen
• Multiagent Systemssubfield70%
  The paper compares against multi-agent systems and aims to replicate their multi-perspective synergy in a single agent, but the core method is not a multi-agent system itself.
  MAS achieves dynamic multi-perspective problem solving by orchestrating agents with distinct text-based rolesMoRe enables multi-perspective specialization in a single-agent, single-turn inference process
• Deep Learningframework70%
  The training of the router and codebook uses deep learning techniques (SFT and GRPO) on neural networks.
  trained via a three-stage SFT curriculum and GRPO post-traininglightweight query-aware composer
• Machine Learningsubfield60%
  The method involves training a router and codebook using SFT and GRPO, which are machine learning techniques, but the focus is on LLM specialization rather than general ML.
  MoRe can be efficiently trained via a three-stage SFT curriculum and GRPO post-trainingOnly the codebook of candidate steering vectors and the lightweight query-aware composer are optimized
• Ensemble Methodsframework60%
  The method combines multiple steering vectors (like an ensemble) to improve performance, though it is not a traditional ensemble method.
  dynamically fuses the codebook into a steering vector that encompasses multiple rolesquery-dependent selection and composition of complementary candidate steering vectors
• Deep Learningsubfield60%
  The method manipulates hidden representations of transformer-based LLMs, which falls under deep learning, but the contribution is not a new deep learning architecture.
  activation steering directly injects a pre-computed steering vector into the hidden layerssteering produces persistent and candidate-specific effects that accumulate toward the model output
• Bayesian Machine Learningframework50%
  The GRPO training involves group-relative advantages and soft rewards, which are probabilistic in nature, but the paper does not emphasize Bayesian methods.
  GRPO to post train the composergroup-relative advantages of a group of rollouts

Abstract

Specializing Large Language Models (LLMs) toward distinct abilities underpins successes ranging from personalized assistants to multi-agent systems (MAS). Single-agent paradigms rely on pre-defined personas or steering vectors to induce specialization, yet they impose a single fixed specialization that fails to adapt to diverse queries. Conversely, MAS achieves dynamic multi-perspective problem solving by orchestrating agents with distinct text-based roles, but fusing these specializations requires multi-turn interactions that inflate context length and inference cost. To address these limitations, we propose Mixture of Roles (MoRe), which adaptively composes multiple specializations into a single steering vector for single-turn inference. Specifically, MoRe learns a diversified codeboox of steering vectors, each of which encodes a latent role. A query-aware router dynamically fuses the codebook into a steering vector that encompasses multiple roles. By steering the backbone LLM with the composed vector, MoRe enables multi-perspective specialization in a single-agent, single-turn inference process. The proposed MoRe can be efficiently trained via a three-stage SFT curriculum and GRPO post-training, while the backbone LLM remains frozen. Experiments across reasoning and personality benchmarks show that MoRe outperforms single-agent baselines by 2.2% on average, and achieves performance on par with MAS while reducing token cost by 20x.

Paper Context