Structured Transforms for Low-Overhead Quantization of Language Models

Replacing dense orthogonal matrices with sign-randomized DCTs accelerates Kashin-based LLM quantization to O(N log N) with guaranteed convergence, achieving 4-bit accuracy competitive with OPTQ and QuIP while maintaining numerical stability and native 2-bit hardware compatibility.

Daria Cherniuk, Alexander Rudikov, Boris Kashin, Ivan Oseledets

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026