As teams integrate multiple LLMs - each with different strengths, styles, or cost/latency profiles — routing the right prompt to the right model becomes a critical part of the application design. But it's still an open problem. Most routing systems fall into two camps:
- Embedding-based routers use intent classifiers — label a prompt as “support,” “SQL,” or “math,” then route to a matching model. This works for simple tasks but breaks down in real conversations. Users shift topics mid-conversation, task boundaries blur, and product changes require retraining classifiers.
- Performance-based routers pick models based on benchmarks like MMLU or MT-Bench, or based on latency or cost curves. But benchmarks often can't capture what matters in production: domain-specific quality or subjective evaluation criteria. These routers are often opaque, difficult to debug, and their quality judgments can feel arbitrary, failing to capture the subjective nuance of what a “good” response actually means for a specific user’s intent.
Arch-Router takes a different approach: route to LLMs based on preferences written as policies in plain ol English.
You write policies like “contract clauses → GPT-4o” or “quick travel tips → Gemini Flash.” The router maps the prompt (and the full conversation context) to those policies using a lightweight 1.5B auto-regressive model. The model is capable to handle intent drift, supports multi-turn conversations, and lets you swap in or out models with a one-line change to the routing policy. To read more about the strength of our model, check out our research paper here: https://arxiv.org/abs/2506.16655
Essentially, Arch-Router splits the routing process into two distinct parts:
Route Selection: This is the what. The system defines a set of human-readable routing policies using a “Domain-Action Taxonomy.” Think of it as a clear API contract written in plain English. A policy isn’t just intent_123; it’s a descriptive label like Domain: ‘finance’, Action: ‘analyze earnings report’. The router’s only job is to match the user’s query to the best-fit policy description.
Model Assignment: This is the how. A separate, simple mapping configuration connects each policy to a specific LLM. The finance/"analyze earnings report" policy might map to a powerful model like GPT-4o, while a simpler general/"greeting" policy maps to a faster, cheaper model.
- 1.5B params — runs on a single GPU (or CPU for testing)
- No retraining needed — point it at any mix of LLMs
- Outperforms larger closed models on our conversational routing benchmarks (details in the paper)
Links:
[1] Arch Proxy: https://github.com/katanemo/archgw