Structured Transforms for Low-Overhead Quantization of Language Models

Excerpt

arXiv:2609.11687v1 Announce Type: new Abstract: We revisit Kashin-decomposition-based weight quantization for large language models and propose an improved algorithm with stronger convergence properties and structured, efficient orthogonal transforms. The method retains the core factorization of each weight into two components -- one with bounded infinity norm and the other with bounded infinity norm after an orthogonal transformation -- but replaces the dense random orthogonal matrix with a sig