Skip a Layer or Loop It? Learning Program-of-Layers in LLMs

In three linesLLMs typically execute all layers in fixed order. This research reveals that dynamic executions (PoLar) can skip or loop layers per input, reducing depth while maintaining or improving accuracy. A lightweight prediction network generates customized execution programs, improving results on mathematical reasoning benchmarks.

Summary generated by Claude — human-verified