Published
Elo Router
Elo Rating based routing — competitive, dynamic routing that adapts to changing LLM performance.
Overview
The Elo Router borrows the Elo rating system from competitive chess to maintain dynamic ratings for LLMs, routing queries to models with the highest predicted performance.
How It Works
Each LLM starts with a base Elo rating. After each query, the router observes the outcome (success, quality score, etc.) and updates the ratings accordingly. Higher-rated models are selected more frequently, but the system continuously adapts as performance changes.
Strategy
Maintains Elo ratings for LLMs and routes based on predicted performance.
API Endpoint
autoroute:elorouter
Use Cases
- Dynamic environments where LLM performance changes over time
- Competitive routing scenarios
- When you want automatic adaptation to model drift
Best Practices
Rating Volatility
Adjust the K-factor to control how quickly ratings respond to new data. Higher K values (32+) make the system more responsive but noisier. Lower values (8-16) provide stability at the cost of slower adaptation.
Related Models
- MF Router — For static historical performance data
- AutoMix — For combining multiple model outputs