AI Paper Atlas

Research map / Architectures and efficiency / Quantization

Hyper-Compression: Model Compression via Hyperfunction

Fenglei Fan, Juntong Fan, Dayang Wang, Jingbo Zhang, Zelin Dong, Shijun Zhang, Ge Wang, Tieyong Zeng

TPAMI 2026 · Architectures and efficiency › Quantization · approach: KV cache quantization · 4 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
apply Hyper-Compression to LLaMA2-7B (2.60x)LLaMA2-7B · 8 downstream tasks averageaccuracy (%): 65.88 -> 64.89Hurt
apply Hyper-Compression on top of pruned Sheared-LLaMA-1.3BSheared-LLaMA-1.3B · 8 downstream tasks averageaccuracy (%): 50.31 -> 49.76Hurt
apply Hyper-Compression to Qwen2-72B-InstructQwen2-72B-Instruct · 8 downstream tasks averageaccuracy (%): 78.27 -> 78.31No clear effect
apply Hyper-Compression to LLaMA2-7BLLaMA2-7B · WikiText-2perplexity: 5.47 -> 5.82Hurt
use Hyper-Compression instead of int4 asymmetric quantizationLLaMA2-7B · WikiText-2perplexity: 26160.34 -> 5.82Improved
use GPTQ int4 instead of Hyper-CompressionLLaMA2-7B · WikiText-2perplexity: 5.82 -> 5.69Improved
apply Hyper-Compression to LiteLLaMALiteLLaMA · WikiText-2perplexity: 31.82 -> 37.85Hurt
apply Hyper-Compression to UNet (7.87x)UNet · UNet segmentationDice (%): 99.86 -> 99.71No clear effect
use Hyper-Compression instead of pruning on UNetUNet · UNet segmentationDice (%): 96.34 -> 99.71Improved
use Hyper-Compression instead of pruning on MobileNetV3MobileNetV3 · MobileNetV3 classificationtop-1 accuracy (%): 69.32 -> 73.93Improved
use Hyper-Compression instead of SQuant-INT4 on ResNet50ResNet50 · ResNet50 classificationtop-1 accuracy (%): 70.80 -> 76.40Improved
use Hyper-Compression on Qwen1.5-7B-Chat (2.67x)Qwen1.5-7B-Chat · 16 tasks averageaccuracy (%): 65.23 -> 64.75Hurt

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