AI Paper Atlas

Research map / Architectures and efficiency / Quantization

TWEO: Transformers Without Extreme Outliers Enables FP8 Training And Quantization For Dummies

Guang Liang, Jie Shao, Ningyuan Tang, Xinyao Liu, Jianxin Wu

CVPR 2026 · Architectures and efficiency › Quantization · approach: Post-training quantization of ViTs · 7 citations (OpenAlex)

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 TWEO loss to Swin-TSwin-T · ImageNet-1Klargest outlier (peak): 1556 -> 22Improved
add TWEO loss to Swin-TSwin-T · ImageNet-1Ktop-1: 81.2 -> 81.4No clear effect
add TWEO loss to Swin-SSwin-S · ImageNet-1Klargest outlier (peak): 6402 -> 35Improved
add TWEO loss to GPT2-XL under per-tensor FP8GPT2-xl · OpenWebText FP8validation perplexity: 93.28 -> 12.58Improved
add TWEO loss to GPT2 124M under FP8GPT2 124M · OpenWebText FP8validation perplexity: 169.81 -> 19.26Improved
add TWEO loss to GPT2 7B under FP8GPT2 7B · OpenWebText FP8validation perplexity: 67.12 -> 12.02Improved
add TWEO loss, BF16 trainingGPT2-xl · OpenWebText BF16validation perplexity: 13.84 -> 12.39Improved
add TWEO loss to Medium before W8A8 per-tensor static quantizationGPT-2 Medium · OpenWebText W8(T)A8(T)perplexity: 1491.11 -> 16.50Improved
quantize residual stream with SmoothQuant on default modelGPT-2 XL default · OpenWebText W8(C)A8(T)perplexity: 14.81 -> 1876.70Hurt
apply SmoothQuant on top of TWEO model instead of AbsMax (ablation)GPT-2 XL TWEO · OpenWebText W8(C)A8(T), residual quantizedperplexity: 13.06 -> 12.89No clear effect
use invasive attention bias instead of TWEO (ablation)GPT2 124M · OpenWebText FP8validation perplexity: 19.26 (TWEO) vs 24.93 (Attn bias)Hurt
random inputs to pretrained model (ablation)Qwen2.5-0.5B · extreme outliersoutlier magnitudeNo clear effect

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 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