arXiv:2605.26339v1 Announce Type: new 
Abstract: Scalar post-training quantizers discard pairwise coordinate structure within weight rows. We introduce QAM-W (Quadrature Amplitude Modulation for Weights), a codec that recovers this structure: each row is L2-normalized, block-Hadamard rotated, paired into 2D coordinates, and quantized against a single Lloyd-Max codebook trained on the unit circular Gaussian, with activation-aware per-channel scaling. In a cross-model study spanning five LLMs from four families (1.1B--13B parameters) and eight quantized configurations, the activation-aware variant at $\approx 5.5$ bpw stays within $\pm 0.4\%$ of BF16 WikiText-2 perplexity on every model, matching the SmoothQuant W8A8 quality envelope at $32\%$ fewer weight bits. Joint 2D coding outperforms polar (amplitude $\times$ phase) coding by 2--15~pp $\Delta$PPL at equal bitrate, and paired KL against BF16 tracks $\Delta$PPL\% at Spearman $\rho = 0.99$ across 37 (method, model) rows, consistent with a monotone composite bound from codec distortion to KL divergence. A 3.5~bpw variant is competitive on quantization-tolerant architectures. At strict 4~bpw, the rotated-codebook frontier method QTIP outperforms QAM-W; the contribution is the quality-preserving 5--6~bpw band.

يقدم المقال بعنوان 'QAM-W: Joint 2D Codebook Quantization for LLM Weights via Hadamard Rotation and Activation-Aware Scaling' كوديك جديد، QAM-W، الذي يحسن من كميّة أوزان نماذج اللغة الكبيرة (LLM) من خلال الحفاظ على بنية الإحداثيات داخل صفوف الأوزان. تستخدم هذه الطريقة تطبيع L2، وتدوير هادامارد الكتلي، وتعديل الوعي بالتفعيل، مما يحقق دقة عالية عبر نماذج LLM متعددة مع تقليل عدد بتات الوزن المطلوبة.

El artículo titulado 'QAM-W: Joint 2D Codebook Quantization for LLM Weights via Hadamard Rotation and Activation-Aware Scaling' presenta un nuevo códec, QAM-W, que mejora la cuantificación de los pesos de los modelos de lenguaje de gran tamaño (LLM) al preservar la estructura de coordenadas dentro de las filas de pesos. Este método emplea normalización L2, rotación de Hadamard en bloques y escalado consciente de la activación, logrando alta precisión en varios LLM mientras reduce la cantidad de bits de peso requeridos.

L'article intitulé 'QAM-W: Joint 2D Codebook Quantization for LLM Weights via Hadamard Rotation and Activation-Aware Scaling' présente un nouveau codec, QAM-W, qui améliore la quantification des poids des modèles de langage de grande taille (LLM) en préservant la structure des coordonnées au sein des lignes de poids. Cette méthode utilise la normalisation L2, la rotation de Hadamard par blocs et l'échelle consciente de l'activation, atteignant une grande précision sur divers LLM tout en réduisant le nombre de bits de poids requis.

The paper titled 'QAM-W: Joint 2D Codebook Quantization for LLM Weights via Hadamard Rotation and Activation-Aware Scaling' introduces a novel codec, QAM-W, which enhances the quantization of large language model (LLM) weights by preserving the coordinate structure within weight rows. This method employs L2 normalization, block-Hadamard rotation, and activation-aware scaling, achieving high accuracy across various LLMs while reducing the number of weight bits required.