BiSCo-LLM: Lookup-Free Binary Spherical Coding for ~2-Bit LLM Compression

arXiv·medium signal

BiSCo-LLM tackles the memory, weight-bandwidth, and checkpoint-storage limits of LLM deployment at extreme low-bit budgets. It positions itself between scalar/group-wise quantization (kernel-friendly but weak near 2 bits per weight) and vector quantization (richer block-level representation but slower), using lookup-free binary spherical coding to get vector-quant-level representation without the lookup-table overhead. Aimed at practitioners pushing weights toward 2-bit for edge and cost-sensitive serving.