AI Paper Atlas

Research map / Architectures and efficiency / Quantization

Metis: Training LLMs with FP4 Quantization

Hengjie Cao, Mengyi Chen, Yifeng Yang, Fang Dong, Ruijun Huang, Jixian Zhou, Anrui Chen, Mingzhi Dong, Yujiang Wang, Jinlong Hou, Yuan Cheng, FAN WU and 4 more

ICLR 2026 · Architectures and efficiency › Quantization · approach: Post-training quantization of LLMs · 9 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
Apply Metis spectral decomposition instead of direct NVFP4LLaMA-3 8B FP4 training · DCLM 100B tokenstest loss: 2.53 -> 2.45Improved
Apply Metis spectral decomposition instead of direct NVFP4GPT-2 130M FP4 training · DCLM 100B tokenstest loss: 3.32 -> 3.20Improved
Apply Metis spectral decomposition instead of direct NVFP4GPT-2 1.1B FP4 training · DCLM 100B tokenstest loss: 3.22 -> 3.01Improved
Apply Metis instead of direct NVFP4LLaMA-3 8B FP4 training · Downstream average (7 tasks)average accuracy: 50.0 -> 51.0Improved
Apply Metis instead of direct NVFP4GPT-2 1.1B FP4 training · Downstream average (7 tasks)average accuracy: 47.5 -> 48.6Improved
Add spectral decomposition of gradients (ablation)Metis on LLaMA-3 8B · DCLM 100B tokenstest loss: 2.51 -> 2.45Improved
Add spectral decomposition of gradients (ablation)Metis on LLaMA-3 8B · Downstream averageaverage accuracy: 50.1 -> 51.1Improved
Add spectral decomposition of weights (ablation)Metis on LLaMA-3 8B · DCLM 100B tokenstest loss: 2.47 -> 2.45Improved
Add spectral decomposition of activations (ablation)Metis on LLaMA-3 8B · DCLM 100B tokenstest loss: 2.46 -> 2.45Improved
Add spectral decomposition of weights (ablation)Metis on LLaMA-3 8B · Downstream averageaverage accuracy: 50.7 -> 51.1Improved
Add spectral decomposition of activations (ablation)Metis on LLaMA-3 8B · Downstream averageaverage accuracy: 50.9 -> 51.1Improved
Use sparse random sampling instead of full-batch decomposition (ablation)Metis on LLaMA-3 8B · DCLM 100B tokenstest loss: 2.44 -> 2.45No clear effect

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