PCST: A Systematic Study of Extreme Low-Bit LLaMA-7B Compression Without Retraining
Year: 2026
Type: preprint
Abstract: This preprint presents a systematic experimental study of extremelow-bit compression of LLaMA-7B without retraining, distillation,pruning, or architectural changes. More than 60 documented compression methods and variants were evaluated,including Product Quantization, selective low-rank residuals,activation- and Hessian-aware approaches, function-preserving gaugetransformations, additive quantization, sparse corrections, andnetwork-level output calibration. The final PCST v5 artifact occupies 2.05 GiB. On a document-resetWikiText-2 evaluation with 6,688 prediction positions, it achievesperplexity 15.56 and 59.59% Top-10 overlap with the Q8 reference.Although the result remains behind standard Q3_K_M quantization inquality and speed, the study identifies reproducible positive results,documents negative results, and demonstrates that local weight or blockreconstruction error often fails to predict end-to-end language-modelbehavior.
Author: Luzhetsky, Ilia
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7208787767
Cites:
Cited by:
Related to: