Route Anything proposes a general AI model and policy routing framework

The authors of the Route Anything paper propose a unified framework for studying model and policy routing in general AI systems, and establish an optimal routing theory under the quality-cost tradeoff. The authors constructed an all-modality trajectory matrix covering 25 benchmarks, 452 candidate configurations, and 374,401 scoring runs across language understanding, image, video and speech generation, coding and office agents, and robotic manipulation. The authors state that learned routers outperform the framework by 0.3 points on 16 benchmarks, by 2.1 points on video generation, and by 1.1 points on robotic manipulation, while reducing policy latency by 24%.

AI 根据报道生成 · 7 days ago更新