# O03: RouteLLM: Learning to Route LLMs with Preference Data, v4

- Publisher: Ong et al., arXiv / ICLR 2025
- URL: [Primary source](https://arxiv.org/html/2406.18665v4)
- Published: 2024-06-26
- Updated/version date: 2025-02-23
- Accessed: 2026-10-22
- Type: primary research paper, full text read
- Verification: opened source, rather than relying on a search excerpt

## Supported finding

Routing between model tiers can improve tested cost/performance trade-offs. Some out-of-domain evaluations performed poorly before targeted training-data augmentation.

## Limits and transfer

Tested GPT-4/Mixtral pair and academic benchmarks do not establish a current enterprise savings rate.

## Use in the session

Mechanism and decision conditions. Retain the qualifications when translating the finding into slide copy. No source establishes Knauf's savings.