Dynamic Routing as a New Dimension for Test-time Versatility of LLMs

Dynamic routing programs provide a test-time adaptation axis complementary to chain-of-thought by executing subsets or iterations of model layers. Selecting these programs using label probabilities from 3 to 10 demonstrations improves performance across 49 MMLU tasks without gradient updates or task-specific training.