This paper introduces Arch-Router, a preference-aligned routing framework for large language models (LLMs). It addresses limitations in existing routing approaches by focusing on matching queries to user-defined preferences (domain and action types) rather than solely relying on benchmark performance. The framework includes a 1.5B parameter model, Arch-Router, and a data creation pipeline. Experiments demonstrate state-of-the-art results in matching queries with human preferences and improved adaptability.
This paper proposes a preference-aligned routing framework for LLMs that guides model selection by matching queries to user-defined domains or action types. It introduces Arch-Router, a compact 1.5B model that learns to map queries to domain-action preferences for model routing decisions, outperforming proprietary models in subjective evaluation criteria.
