Get Started
Home
Topics
Search
Library
Research questionHow can extreme low-bit LLM quantization adapt its basis across layers, calibration data, and quantizers?Activation outliers and anisotropic weight curvature make extreme quantization fragile. Fixed Hadamard transforms can improve robustness but cannot adjust their rotated basis to different layers, calibration distributions, or quantizers.
AI
Inference Optimization
LLM Pretraining & Post-training
Machine Learning
Research Paper
Technology
Latest papersRecent research connected to this question, newest first.HARP: Hadamard-Preconditioned Adaptive Rotation Processor for Extreme LLM QuantizationThe evidence covers HARP, a learnable structured two-sided orthogonal processor fitted only on calibration data. It reports 2–4-bit results for Llama models from 1B to 70B and a 2-bit transfer experiment on Qwen3-8B, plus throughput measurements for Llama 2 7B at 2 bits.research paper · Sep 2, 2026
Related questions
How can complete transformer layers be removed from LLMs without sacrificing quality or standard serving compatibility?How can multi-shell Leech-lattice 2-bit weights be decoded efficiently for batch-one GPU serving?How can large language models cut training and inference costs without materially harming accuracy?How can we robustly compress LLM KV caches across open-domain inputs without input-specific budget thresholds?