News without the noise
Benchmark
Reported in 19 papers on Zotpaper. The latest, each as the paper states it:
9 results for “WikiText-2”
Research · Efficiency & systems
FIGFLEX: A Mixed-Precision FP-INT GEMM Accelerator Co-Designed with Low-Bit LLM Quantization
A radix-4 bit-parallel datapath and GPTQ-aware quantizer let FIGFLEX match or beat GPU-level accuracy at 2 to 3 bit precision while raising throughput as precision falls.
WikiText-2: 8.77 perplexity (as reported)
University of California, Irvine
Research · Efficiency & systems
Two For The Price Of One: Curvature-Aware Pruning Preserves Influence Functions Too
CACTUS, an ADMM-based optimizer, achieves order-of-magnitude speedups over prior methods while retaining gradient alignment and enabling data selection.
WikiText-2: 14.17 perplexity (as reported)
Massachusetts Institute of Technology, Google Research
Research · Efficiency & systems
Learning Functional Subspaces for Neural Network Compression
Learnable Subspace Projections jointly optimize which activation directions to retain, improving accuracy, decoding speed and memory use after factorization.
WikiText-2: 16.2 perplexity (as reported)
Helmholtz Munich, Technical University of Munich, MCML
Research · Efficiency & systems
When Lower Reconstruction Loss Hurts: Distributionally Robust Refinement for Low-Bit LLM Quantization
DRQ refines integer codes to minimize worst-case loss over activation distribution shifts, improving six PTQ methods without inference overhead.
WikiText-2: 14.63 perplexity (as reported)
University of Science and Technology of China, University College London, Shanghai University of Finance and Economics
Research · Efficiency & systems
Deflating the Hessian: Rank-4 W4A4 Quantization for Multimodal Diffusion Transformers
A unified framework jointly modeling weight and activation quantization error achieves better performance with rank 4 than existing methods with rank 32.
WikiText-2: 10.44 perplexity (as reported)
The University of Hong Kong, Huawei Technologies Co., Ltd.
Research · Efficiency & systems
Few Bits, One Law: Toward W2A4KV2
CanonQ uses fixed transforms and reusable codebooks, then jointly trains models to tolerate quantized weights, activations and attention caches.
WikiText-2: 17.8 perplexity (as reported)
Meta AI
Research · Efficiency & systems
Quantization Error Is Spectrally Flat: A Single Random Probe Is a Calibrated, Data-Free Sensitivity Estimator, with Application to Budget-Targeted Mixed-Precision Quantization
The authors show that spectrally flat rounding errors make Gaussian probes reliable enough to allocate per-tensor bit widths under an exact memory budget.
WikiText-2: 6.59 median perplexity (as reported)
Blacksheep ai.
Research · Efficiency & systems
Hardware-Native Joint Sparse-Quantization for Trillion-Scale Mixture-of-Experts
A hardware-software co-design framework compresses expert weights into low-precision sparse representations, achieving up to 4.35% accuracy improvement and 1.65x kernel speedup on NVIDIA B200 GPUs.
Grouped GEMM on Kimi-K2.5 gate/up layer: 1.65 speedup (as reported)
POSTECH, ISTA
Research · Efficiency & systems
Compression Technique Interactions in Small Language Models: A Factorial Ablation Study
A systematic study of 72 compression configurations shows that INT8 quantization is lossless, while NF4 with pruning causes total failure, and LoRA fine-tuning recovers most quality without a teacher model.
WikiText-2: 24.27 perplexity (as reported)
Amazon Web Services