AI Paper Atlas

Research map / Architectures and efficiency / Quantization

STaMP: Sequence Transformation and Mixed Precision for Low-Precision Activation Quantization

Marco Federici, Riccardo Del Chiaro, Boris van Breugel, Paul N. Whatmough, Markus Nagel

ICLR 2026 · Architectures and efficiency › Quantization · approach: Post-training quantization of ViTs

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add STaMP sequence transform with 64 8-bit tokensFlatQuant · WikiText-2 W4A4KV4perplexity: Llama 3 8B: 6.89 -> 6.77Improved
add STaMP sequence transformQuaRot · WikiText-2 W4A4KV4, Qwen 2.5 3B itperplexity: 94.86 -> 71.13Improved
add STaMP sequence transformRTN · WikiText-2 W4A4KV4, Llama 3 8Bperplexity: 668 -> 95.3Improved
add STaMP to RTNRTN on PixArt-Sigma · COCO W4A4Image Reward: 0.38 -> 0.80Improved
add STaMP to SVDQuantSVDQuant on PixArt-Sigma · COCO W4A4Image SQNR: 8.78 -> 9.72Improved
add STaMP to ViDiT-QViDiT-Q on PixArt-Sigma · COCO W4A4Image SQNR: 7.82 -> 6.37Hurt
add STaMP to ViDiT-QViDiT-Q on SANA · COCO W4A4Image Reward: 0.89 -> 0.86Hurt
add STaMP to QuaRot feature transform (ablation)QuaRot on PixArt-Sigma ffn.down_proj · A4 single-activation SQNRImage SQNR: 8.42 -> 14.91Improved
apply STaMP to attn2.to_out activations (ablation)PixArt-Sigma · A4 single-activation SQNRImage SQNR: 15.48 -> 15.49No clear effect
use DWT instead of Hadamard for the sequence transform (ablation)STaMP · PixArt-Sigma denoising stepCUDA latency overhead (%): 56.8 -> 4.8Improved
use DCT, WHT or DWT as sequence transform (ablation)STaMP · PixArt-Sigma and Llama v3 8B, A4qualityNo clear effect
increase number of 8-bit high-precision tokens (ablation)STaMP · PixArt-Sigma, activation-only quantizationSQNRImproved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy