BiSCo-LLM: Lookup-Free Binary Spherical Coding for Extreme Low-Bit LLM Compression

This paper presents BiSCo-LLM, a codebook-free binary spherical coding framework for extreme low-bit weight compression of LLMs. The core pipeline consists of three components: First, local weight chunks are mapped onto a unit hypersphere and binarized into compact spherical codes, where the main payload is a bit-packed sign stream without explicit VQ codebooks or index lookups. Second, a residual BSQ stage encodes the reconstruction error from the base spherical codec, providing a rate-distortion path without stored codebooks. Third, category-wise recovery distillation is performed after replacing each Transformer module category, reducing the mismatch between local weight reconstruction and assembled model behavior. A small 8-bit protected-channel path is used as an auxiliary stabilization mechanism for sensitive channels, counted separately from the BSQ payload. The reported storage budget includes binary codes, neural decoders, protected-channel payloads, LoRA adapters, and metadata. Experiments demonstrate that BiSCo-LLM significantly outperforms existing scalar/group-wise quantization and vector quantization methods at extremely low bitrates (e.g., sub-2-bit), offering an efficient and compact solution for LLM storage and deployment.