Research map / Architectures and efficiency / Quantization
LittleBit-2: Maximizing the Spectral Energy Gain in Sub-1-Bit LLMs via Latent Geometry Alignment
Banseok Lee, Youngmin Kim
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Joint-ITQ rotation to LittleBit initialization (ablation) | LittleBit · WikiText-2, 1.0 bpp | perplexity: 16.30 -> 11.53 | Improved |
| add random internal rotation only (ablation) | LittleBit · WikiText-2, 1.0 bpp | perplexity: 16.30 -> 12.63 | Improved |
| replace random rotation with Joint-ITQ rotation (ablation) | LittleBit + random rotation · WikiText-2, 1.0 bpp | perplexity: 12.63 -> 11.53 | Improved |
| add random rotation only (ablation) | LittleBit · WikiText-2, 0.1 bpp | perplexity: 26.11 -> 24.16 | Improved |
| replace random rotation with Joint-ITQ (ablation) | LittleBit + random rotation · WikiText-2, 0.1 bpp | perplexity: 24.16 -> 23.74 | Improved |
| low-rank binary instead of tiny-rank FP16 (ablation) | Llama-3 8B at 0.1 bpp · WikiText-2 | perplexity: 59.44 -> 26.11 | Improved |
| apply Joint-ITQ geometric alignment | LittleBit · Llama-3 8B zero-shot avg, 0.1 bpp | accuracy: 41.80 -> 41.64 | No clear effect |
| apply Joint-ITQ geometric alignment | LittleBit · Llama-2 13B zero-shot avg, 1.0 bpp | accuracy: 51.49 -> 60.53 | Improved |
| apply Joint-ITQ geometric alignment | LittleBit vs OneBit · Llama-2 7B, 1.0 bpp | perplexity: 9.08 -> 8.27 | Improved |
| apply Joint-ITQ | LittleBit · Gemma-3 27B, 0.1 bpp | perplexity: 17.72 -> 16.38 | Improved |
| apply internal rotation (ablation) | LittleBit · Llama-2 7B synthetic spectrum break-even gamma | break-even gamma: 0.36 -> 0.41 | Improved |
| apply Joint-ITQ on top of rotation (ablation) | LittleBit · synthetic spectrum break-even gamma | break-even gamma: 0.41 -> 0.51 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- OneBit — OneBit: Towards Extremely Low-bit Large Language Models (NeurIPS 2024) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- LittleBit: Ultra Low-Bit Quantization via Latent Factorization (NeurIPS 2025 · similarity 0.90)
- ScaleBITS: Scalable Bitwidth Search for Hardware-Aligned Mixed-Precision LLMs (NeurIPS 2026 · similarity 0.87)
- STBLLM: Breaking the 1-Bit Barrier with Structured Binary LLMs (ICLR 2025 · similarity 0.87)
- QBB: Quantization with Binary Bases for LLMs (NeurIPS 2024 · similarity 0.86)
- SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression (ICLR 2024 · similarity 0.86)
- CodeGEMM: A Codebook-Centric Approach to Efficient GEMM in Quantized LLMs (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
