Programs-of-Layers Enables Dynamic LLM Inference Routing

Replicating the PoLar system, this work demonstrates that treating transformer layers as a functional library allows for dynamic routing. Results show that skipping or repeating layer blocks based on input difficulty outperforms standard fixed-depth forward passes.