LayerRoute is a parameter-efficient method that decides per input whether a small language model can skip selected transformer layers. Applied to Qwen2.5-0.5B-Instruct, it made nine middle layers—blocks 8 through 16—eligible for skipping.
Each of the model’s 24 blocks receives a lightweight router, totaling about 21,500 routing parameters. The system jointly trains rank-8 LoRA adapters with roughly 1.08 million parameters so the model can adapt while learning hard skip-or-run decisions.
Across 10 independently seeded training runs, the same nine-layer structure emerged every time. Measured wall-clock speedups ranged from 1.02 to 1.06 times, averaging 1.04 times. Perplexity improved relative to the unchanged backbone in every tested configuration, and routing decisions changed actual execution for 87% to 100% of held-out samples, indicating that it was not merely fixed pruning.
Training took under seven minutes on one A100 GPU, according to the authors. The measured gain is modest, and the experiment uses a 0.5B model, so larger architectures and production-serving stacks may behave differently. The new arXiv preprint is useful partly because it reports real timing rather than inferring speed solely from skipped computation.