# RouteLLM: Learning to Route LLMs with Preference Data

**Authors:** Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E. Gonzalez, M Waleed Kadous, Ion Stoica  
**arXiv:** [2406.18665](https://arxiv.org/abs/2406.18665)  
**Date:** 2024-06-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2406.18665v1/

## Key points

- Authors address the challenge of balancing performance and cost when choosing large language models (LLMs) for various tasks.
- Their novel approach uses efficient router models to dynamically select between a stronger and weaker LLM during inference to optimize the balance between cost and response quality.
- They develop a training framework for these router models that leverages human preference data and data augmentation techniques to enhance performance.
- Evaluation on widely-recognized benchmarks shows that their approach significantly reduces costs by over 2 times in certain cases without compromising response quality.
- Router models exhibit significant transfer learning capabilities, maintaining performance even when strong and weak models are changed at test time.
- Research highlights the potential of efficient router models to provide a cost-effective yet high-performance solution for deploying LLMs across various tasks.

## Summary

In their paper titled "RouteLLM: Learning to Route LLMs with Preference Data," authors Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E. Gonzalez, M Waleed Kadous, and Ion Stoica address the challenge of balancing performance and cost when choosing large language models (LLMs) for various tasks. <br>
Their novel approach using efficient router models dynamically selects between a stronger and weaker LLM during inference to optimize the balance between cost and response quality. <br>
The authors develop a training framework for these router models that leverages human preference data and data augmentation techniques to enhance performance. <br>
Evaluation on widely-recognized benchmarks demonstrates that their approach significantly reduces costs by over 2 times in certain cases without compromising response quality. <br>
Interestingly, the router models also exhibit significant transfer learning capabilities, maintaining their performance even when the strong and weak models are changed at test time. <br>
Overall, this research highlights the potential of efficient router models to provide a cost-effective yet high-performance solution for deploying LLMs across various tasks.