The Arch team has released Arch-Router, a 1.5B LLM router model that improves prompt routing to various LLMs by using user-defined English policies, aiming to enhance flexibility and effectiveness in multi-turn conversations without needing retraining.
1.5B parameter auto-regressive language model for routing, leveraging natural language policy definitions and a Domain-Action Taxonomy
2025-07-18
- website
Highlight 1
Flexible policy-driven routing allows non-technical users to define routing logic in plain English, simplifying customization and updates.
Highlight 2
Efficient 1.5B parameter model runs on a single GPU or CPU for testing, lowering resource requirements compared to larger models.
Highlight 3
Robust handling of conversational intent drift and multi-turn contexts enhances routing accuracy over simpler intent classification methods.
Improvement 1
Expand integration tools or APIs for easier adoption across diverse LLMs and environments to streamline deployment.
Improvement 2
Improve user interface or visualization for policy management to make policy creation and debugging more intuitive.
Improvement 3
Provide more extensive documentation and examples for complex use cases to lower the learning curve for new users.
Product Functionality
Develop more API integrations and adapters for popular LLM providers to simplify connecting Arch-Router with diverse models. Introduce monitoring and analytics dashboards to track routing performance in real time.
UI & UX
Create a user-friendly web interface for visual policy authoring, testing, and debugging to help users manage complex routing rules without manual config edits. Add interactive tutorials or guided workflows for onboarding.
SEO or Marketing
Enhance the project page with clear value propositions, use cases, and comparison with existing routing methods. Publish case studies and demo videos to showcase practical benefits. Improve discoverability by optimizing metadata and keywords targeting AI developers and product teams.
MultiLanguage Support
Introduce multi-language support for the policy definitions and UI to attract a broader global audience. Allow policies to be written and matched in multiple languages to support international deployments.
- 1
How does Arch-Router decide which model to route a prompt to?
Arch-Router uses a lightweight 1.5B parameter language model to map the user prompt and conversation context to human-readable policies you define in plain English. These policies specify domains and actions, and each policy is mapped to a specific LLM, enabling precise routing.
- 2
Do I need to retrain Arch-Router when I add or change models?
No retraining is needed. You can simply update the routing policies or the policy-to-model mapping configuration with one-line changes to incorporate new models or adjust routing behavior.
- 3
What advantages does Arch-Router have over traditional embedding-based or performance-based routers?
Unlike embedding-based routers that rely on fixed intent classifiers or performance-based routers tied to benchmarks, Arch-Router routes based on user-defined preferences written as policies in natural language. This approach better captures domain-specific nuances, adapts to intent drift, and allows easy customization without opaque or arbitrary quality judgments.