Avengers-Pro: Performance-Efficiency Optimized Routing Framework for Large Language Models

Summary

Researchers present Avengers-Pro, a test-time routing framework that dynamically assigns queries to different LLMs based on performance-efficiency optimization. The system clusters incoming queries and routes them to the most suitable model (either efficient or high-capacity) to achieve optimal balance between accuracy and cost. Across 6 benchmarks and 8 leading models including GPT-5-medium and Gemini-2.5-pro, Avengers-Pro achieves state-of-the-art results: +7% higher accuracy than the strongest single model, 27% lower cost at equivalent accuracy, and 63% lower cost at 90% performance.