GPTAQ (LLM quantization): superseded — cited as a baseline and beaten by newer methods. 1 paper(s) critique it, 3 beat it on benchmarks — #17 of 80 most-superseded. Sub-problem: cluster led by GPTAQ. Newer alternatives in the same sub-problem include SplitQ, MARR, Quantization-aware Integrated Gradients (QIG), SPEED-Q, Quant-dLLM.

LLM quantization

GPTAQ

GPTAQ: Efficient Finetuning-Free Quantization for Asymmetric Calibration

Superseded baseline#17 of 80 most-superseded · first seen Apr 3, 2025

Superseded — cited as a baseline and beaten by newer methods

1 papers critique it · 3 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites GPTAQ as a baseline.

This cross-layer residual is beneficial for reducing accumulated quantization errors; however, it may also introduce additional Hessian-approximation (HA) bias.

— MARR: Module-Adaptive Residual Reconstruction for Low-Bit Post-Training Quantization

Beaten on benchmarks

Head-to-head results where a newer method reports beating GPTAQ. Values are copied from the source paper's tables — verify against the cited paper.

Quant-dLLM beats GPTAQ
47.99 vs 31.30
Average · [2-bit weight quantization on Dream-7B-Instruct]
Quant-dLLM: Post-Training Extreme Low-Bit Quantization for Diffusion Large Language Models
GPTAQ-MARR beats GPTAQ
9.56 vs 11.24
Wiki2 · [Llama2-7b W2A4]
MARR: Module-Adaptive Residual Reconstruction for Low-Bit Post-Training Quantization
VLMQ beats GPTAQ
74.40 vs 73.68
Avg · [Qwen2-VL-7B-Instruct-INT3g128]
VLMQ: Efficient Post-Training Quantization for Large Vision-Language Models via Hessian Augmentation

What to use instead

Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.