AI Paper Atlas

Research map / Architectures and efficiency / Quantization

Attn-QAT: 4-Bit Attention With Quantization-Aware Training

Peiyuan Zhang, Matthew Noto, Wenxuan Tan, Chengquan Jiang, Will Lin, Wei Zhou, Hao Zhang

ICML 2026 · Architectures and efficiency › Quantization · approach: Ultra-low-bit LLM quantization · 0 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 high-precision output O' for the backward pass (ablation)Attn-QAT · VBench (Wan 2.1 1.3B)Overall Quality: 0.7185 -> 0.8252Improved
add fake quantization of P in backward recomputation (ablation)Attn-QAT · VBench (Wan 2.1 1.3B)Overall Quality: 0.8254 -> 0.8252No clear effect
add K smoothing to Attn-QAT (ablation)Attn-QAT · VBench (Wan 2.1 1.3B)Overall Quality: 0.8252 -> 0.8232No clear effect
add two-level quantization of P to Attn-QAT (ablation)Attn-QAT · VBench (Wan 2.1 1.3B)Overall Quality: 0.8252 -> 0.8257No clear effect
apply Attn-QAT training to FP4 attention instead of training-free FP4Wan 2.1 1.3B · VBench (Wan 2.1 1.3B)Overall Quality: 0.7785 -> 0.8252Improved
apply Attn-QAT training instead of SageAttention3Wan 2.1 14B · VBench (Wan 2.1 14B)Overall Quality: 0.8203 -> 0.8279Improved
apply Attn-QAT instead of SageAttention3 on Wan 1.3BWan 2.1 1.3B · VBench (Wan 2.1 1.3B)Imaging Quality: 0.5507 -> 0.6775Improved
add Attn-QAT to FP4 attention (continued pretraining)Qwen3-14B · WinoGrandeaccuracy: 0.7214 -> 0.7585Improved
add Attn-QAT to FP4 attention (continued pretraining)Llama 3.1-70B · MMLUaccuracy: 0.7577 -> 0.7773Improved
add Attn-QAT to FP4 attention (continued pretraining)Llama 3.1-70B · WikiTextperplexity: 0.3275 -> 0.3076Improved
use Attn-QAT instead of BF16 attention in SFTQwen3-14B · MMLU-Redux (SFT)accuracy: 0.8316 -> 0.8392No clear effect
use Attn-QAT instead of BF16 attention in SFTLlama 3.1-70B · GSM8K (SFT)accuracy: 0.8840 -> 0.8673Hurt

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