Block-Sphere Vector Quantization
Researchers have introduced Block-Sphere Quantization (BlockQuant), a novel rotation-based algorithm for vector quantization. This new method is designed to better preserve the geometry of rotated embeddings by quantizing blocks on a sphere, outperforming existing techniques like EDEN, RabitQ, and TurboQuant. Experiments on embedding datasets and long-context LLM inference tasks demonstrate practical improvements consistent with theoretical gains. AI
IMPACT Improves efficiency for LLM inference and memory-intensive machine learning tasks.