APQ and MAQEE: New Quantization Methods for Early-Exit Vision Transformers

A recent preprint on arXiv (2605.07317) presents Amortized-Precision Quantization (APQ) and Mutual Adaptive Quantization with Early Exiting (MAQEE) to tackle the challenges of instability in low-precision early-exit Vision Transformers (ViTs). Current quantization techniques rely on a static full-depth execution model, leading to inaccuracies when exit decisions are influenced by quantization noise. APQ offers a utilization-aware approach that examines layer-wise stochastic exposure to this noise, highlighting trade-offs between depth and precision. Meanwhile, MAQEE employs a bi-level strategy that optimizes both exit thresholds and bit-widths with explicit risk management, enhancing inference stability. This method achieves a more favorable Pareto frontier in the accuracy-efficiency balance, cutting BOPs by as much as 95% without sacrificing accuracy, surpassing robust baselines.