# Dynamic Depth Early Exit

Two pretrained time-series foundation models skip whole Transformer blocks at inference, leaving mean forecasting quality nearly unchanged across 1,088 held-out windows while cutting median A100 batch-1 latency by 11.2% and 13.4% at moderate sparsity.

Two pretrained time-series foundation models skip whole Transformer blocks at inference, leaving mean forecasting quality nearly unchanged across 1,088 held-out windows while cutting median A100 batch-1 latency by 11.2% and 13.4% at moderate sparsity.

The work supervises every intermediate depth of one causal speech enhancement model and fine-tunes its output heads so deeper outputs are never worse than shallower ones, yielding a family of static models that are more Pareto-efficient than equivalently-sized counterparts trained from scratch on the same budget (up to 0.11 higher PESQ at equivalent compute, matching the best PESQ at 30% less compute); after int8 quantization on an STM32N6 microcontroller, the dynamic enhancer lies on the same latency-quality frontier as the static models, with the policy running in 26 microseconds per frame on the companion Cortex-M55 and splitting the enhancer into separate NPU graphs adding 2.2% latency overhead.

The work supervises every intermediate depth of one causal speech enhancement model and fine-tunes its output heads so deeper outputs are never worse than shallower ones, yielding a family of static models that are more Pareto-efficient than equivalently-sized counterparts trained from scratch on the same budget (up to 0.11 higher PESQ at equivalent compute, matching the best PESQ at 30% less compute); after int8 quantization on an STM32N6 microcontroller, the dynamic enhancer lies on the same latency-quality frontier as the static models, with the policy running in 26 microseconds per frame on the companion Cortex-M55 and splitting the enhancer into separate NPU graphs adding 2.2% latency overhead.