Quantum Random Access Quantization: A New Approach to One-Bit Post-Training Quantization

A recent study published on arXiv (2608.05240) presents Quantum Random Access Quantization (QRAQ), a novel framework that tackles the shared-sign limitation in one-bit post-training quantization (PTQ). In this approach, each weight is defined solely by its sign, requiring a uniform binary weight matrix across all deployment scenarios, despite differing activation statistics. QRAQ utilizes a quantum random-access code to encode context-specific signs, which are retrieved through context-aligned Pauli measurements. Utilizing a fresh-copy logical readout model, QRAQ generates an unbiased, context-relevant binary surrogate while managing a manageable shot-noise penalty. The study demonstrates a row-wise distinction from shared-sign one-bit PTQ with signed per-row scales, achieving a lower ideal reconstruction risk when optimal context signs conflict. Additionally, the authors outline conditions for finite-shot and calibrated noise that preserve this separation. This theoretical research holds potential for enhancing the deployment efficiency of neural networks across diverse contexts.