Making LLMs Cheaper and Better via Performance-Efficiency Optimized Routing
2 days ago
- #LLM
- #Machine Learning
- #AI Optimization
- Avengers-Pro is a test-time routing framework for LLMs that optimizes performance and efficiency.
- It dynamically routes queries to the most suitable model based on a performance-efficiency score.
- Avengers-Pro achieves state-of-the-art results, surpassing GPT-5-medium by +7% in average accuracy.
- It matches the strongest single model's accuracy at 27% lower cost and reaches ~90% performance at 63% lower cost.
- The framework establishes a Pareto frontier, offering the best accuracy for any given cost and vice versa.
- Code for Avengers-Pro is publicly available.