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
· 0 min read
Abstract
Quantizer-Aligned Recalibration (QuAR) improves the robustness of quantized vision transformers by correcting shifted activations before they enter frozen quantizers. It aligns running channel statistics with source calibration using a single forward pass, without backpropagation or model parameter changes. Experiments span low-bit inference, corruption benchmarks, continual streams, label shifts, and multiple architectures. The results connect improved predictions to restored quantizer code distributions while keeping latency and additional memory small.
Type
Publication
arXiv preprint