| name | bisco-llm-binary-spherical-coding-extreme-compression |
| description | Codebook-free binary spherical coding for extreme low-bit LLM weight compression. Maps weight chunks onto unit hypersphere and binarizes into sign streams. Residual BSQ stage for reconstruction error. Category-wise recovery distillation. Activation: binary spherical coding, LLM compression, low-bit quantization, lookup-free coding, model deployment. |
| metadata | {"arxiv_id":"2607.08643","published":"2026-07-09","authors":"Yuantian Shao, Peisong Wang, Zhilei Liu, Chuangyi Li, Yuanteng Chen","tags":["binary-spherical-coding","llm-compression","low-bit-quantization","lookup-free-coding","model-deployment"]} |
BiSCo-LLM: Lookup-Free Binary Spherical Coding for Extreme Low-Bit LLM Compression
Overview
LLMs are increasingly constrained by memory capacity, weight bandwidth, and checkpoint storage during deployment. BiSCo-LLM presents a codebook-free binary spherical coding framework for extreme low-bit LLM weight compression, eliminating the need for explicit VQ codebooks and index lookups while maintaining rich representation capacity.
Key Innovations
Binary Spherical Coding
- Local weight chunks mapped onto unit hypersphere and binarized into compact spherical codes
- Main payload is a bit-packed sign stream rather than explicit VQ centroids
- Eliminates codebook storage and lookup overhead
Residual BSQ Stage
- Encodes reconstruction error left by base spherical codec
- Provides explicit rate-distortion path without stored codebooks
- Enables progressive quality improvement
Category-Wise Recovery Distillation
- Performed after replacing each Transformer module category
- Reduces mismatch between local weight reconstruction and assembled model behavior
- Preserves end-to-end model performance
Protected-Channel Path
- Small 8-bit auxiliary stabilization for sensitive channels
- Counted separately from BSQ payload for transparent storage accounting
Methodology
- Spherical Mapping: Map weight chunks to unit hypersphere
- Binary Coding: Binarize into sign streams (main payload)
- Residual BSQ: Encode reconstruction error progressively
- Category-Wise Distillation: Recover model behavior per Transformer module category
- Protected Channels: Stabilize sensitive channels with 8-bit path
Implications
- Codebook-free approach eliminates storage and lookup overhead of VQ methods
- Bit-packed sign streams are highly storage-efficient
- Category-wise distillation ensures model quality preservation
- Transparent storage accounting enables fair comparison with other methods
Pitfalls
- Binary codes may not capture all weight distribution complexity at very low bit-rates
- Residual BSQ adds inference-time computation for decoding
- Category-wise distillation requires multiple training passes
- Protected channels increase storage budget beyond pure binary codes
Activation Keywords
binary spherical coding, LLM compression, low-bit quantization, codebook-free, BSQ, weight compression, model deployment, rate-distortion, category-wise distillation
Paper Reference
arXiv:2607.08643 - "BiSCo-LLM: Lookup-Free Binary Spherical Coding for Extreme Low-Bit Large Language Model Compression" (Jul 2026)
