AI Paper Atlas

Research map / Architectures and efficiency / Quantization

ECO: Quantized Training without Full-Precision Master Weights

Mahdi Nikdan, Amir Zandieh, Dan Alistarh, Vahab Mirrokni

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 ECO momentum error injection (SR) (ablation)FP8 w/o master weights · C4 pretraining 30Mvalidation loss: 3.4008 -> 3.3317Improved
add ECO momentum error injection (SR) (ablation)FP8 w/o master weights · C4 pretraining 430Mvalidation loss: 2.8378 -> 2.6544Improved
add ECO momentum error injection (SR) (ablation)FP8 w/o master weights · C4 pretraining 800Mvalidation loss: 2.9471 -> 2.5399Improved
add ECO to RTN weight removal (naive RTN diverges) (ablation)FP8 w/o master weights + RTN · C4 pretraining 30M-800Mvalidation loss: dvg -> 3.3640Improved
use stochastic rounding instead of RTN with ECO (ablation)ECO · C4 pretraining 800Mvalidation loss: 2.6046 -> 2.5399Improved
use stochastic rounding instead of RTN with ECO (ablation)ECO · C4 pretraining 30Mvalidation loss: 3.3640 -> 3.3317Improved
remove FP32 master weights with ECO (SR) versus FP8 w/ MW + SRFP8 training · C4 pretraining 430Mvalidation loss: 2.6500 -> 2.6544No clear effect
remove master weights with ECO (RTN)INT4 QAT of DeepSeek-MoE-16B · MMLUaccuracy: 37.87 -> 38.63No clear effect
remove master weights with ECO (SR)INT4 QAT of DeepSeek-MoE-16B · GSM8Kaccuracy: 16.15 -> 16.00No clear effect
remove master weights naivelyINT4 QAT of DeepSeek-MoE-16B · fine-tuning training losstraining lossHurt
remove FP32 master weights to FP8 with ECOSMoE 2.1B · peak memorybytes per parameter: 12 -> 9Improved

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