Researchers present an enhanced quantization method for large language models that improves on Kashin-decomposition-based approaches by replacing dense random orthogonal matrices with a sign-randomized Discrete Cosine Transform, cutting per-iteration cost from O(N squared) to O(N log N). Their greedy algorithm with alternating updates produces the four-peak distribution needed for stable 2-bit clustering and enables closed-form initialization, removing the k-means bottleneck used in prior methods. Combined with sequential error compensation and preprocessing, the JAX implementation performs competitively with existing quantization methods on Llama-2 and Pythia while showing superior numerical stability in low-bit configurations.
