PACT (LLM quantization): superseded — cited as a baseline and beaten by newer methods. 1 paper(s) critique it, 2 beat it on benchmarks — #24 of 80 most-superseded. Sub-problem: cluster led by PACT. Newer alternatives in the same sub-problem include GABFusion, MEC-Quant.

LLM quantization

PACT

PACT: Parameterized Clipping Activation for Quantized Neural Networks

Superseded baseline#24 of 80 most-superseded · first seen May 16, 2018

Superseded — cited as a baseline and beaten by newer methods

1 papers critique it · 2 beat it on benchmarks

What papers say

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

For instance, the state-of-the-art PACT choi2018pact entails a minimum of $10^{8.12}$ iterations with $1.2M$ training samples to converge on ImageNet.

— Reclaiming Residual Knowledge: A Novel Paradigm to Low-Bit Quantization

Beaten on benchmarks

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

MEC-Qaunt (Setting (B)) beats PACT
62.99 vs 3.25
MBV1 · [W/A 2/2]
MEC-Quant: Maximum Entropy Coding for Extremely Low Bit Quantization-Aware Training
MEC-Qaunt (Setting (A)) beats PACT
81.20 vs 67.18
MBV2 · [W/A 2/4]
MEC-Quant: Maximum Entropy Coding for Extremely Low Bit Quantization-Aware Training
+ Ours beats PACT
66.6 vs 62.9
mAP · [YOLOv5s W3A3]
GABFusion: Rethinking Feature Fusion for Low-Bit Quantization of Multi-Task Networks

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Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.