Test-Time Adaptation of Quantized ViTs via Single-Pass Quantizer-Aligned Recalibration

Oct 6, 2026·
Hyeongheon Cha
Hyeongheon Cha
,
Young D. Kwon
,
Sung-Ju Lee
· 1 min read
QuAR method overview ( ).

Position

First author

Project Goals & Works

  1. Improve the robustness of quantized vision transformers under distribution shift while keeping the model and quantizers frozen.
  2. Recalibrate activations at late-block quantizer inputs and the classifier input using running channel statistics and source calibration statistics.
  3. Adapt within a single forward pass, without backpropagation, parameter updates, or storage of past samples.

Key Results

  1. On ImageNet-C at severity 5, improved ViT-B/16 accuracy over the strongest evaluated backpropagation-free baseline by 4.00 percentage points at W3A3 and 2.28 points at W8A8.
  2. Used about 170 KB of state and reduced latency by at least 46% compared with the evaluated two-forward-pass methods.
  3. Evaluated robustness across continual streams, label shifts, additional distribution shifts, and multiple architectures.