AI Paper Atlas

Research map / Architectures and efficiency / Quantization

Towards Efficient Post-training Quantization of Pre-trained Language Models

Haoli Bai, Lu Hou, Lifeng Shang, Xin Jiang, Irwin King, Michael Lyu

NeurIPS 2022 · Architectures and efficiency › Quantization · approach: Post-training quantization of LLMs · 20 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
Replace layer-wise REM with joint module-wise reconstruction minimizationBERT quantization · MNLI matched, BERT-base, 4-4-8accuracy (%): 81.8_{\pm 0.2} -> 83.5_{\pm 0.1}Improved
Add annealed teacher forcing (ablation)parallel MREM · MNLI matched, BERT-base, 2-2-8, 2,000 stepsaccuracy (%): 82.2_{\pm 0.2} -> 82.7_{\pm 0.2}Improved
Add annealed teacher forcing (ablation)parallel MREM · MNLI matched, BERT-base, 2-2-4, 250 stepsaccuracy (%): 73.9_{\pm 0.5} -> 77.3_{\pm 0.4}Improved
Use parallel MREM with annealed teacher forcing instead of sequential MREMBERT-large quantization · MNLI matched, 4-4-8accuracy (%): 86.1_{\pm 0.1} -> 85.5_{\pm 0.1}Hurt
Add per-channel quantization (ablation)MREM · MNLI matched, BERT-base, 4-4-8accuracy (%): 83.5_{\pm 0.1} -> 83.6_{\pm 0.1}No clear effect
Increase the number of model partitions (ablation)MREM · MNLI, BERT-baseaccuracyHurt
Use more than 4,096 calibration samples (ablation)MREM-S/P · MNLI, BERT-baseaccuracyNo 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