Research map / Architectures and efficiency / Quantization
Improving Block-Wise LLM Quantization by 4-bit Block-Wise Optimal Float (BOF4): Analysis and Variations
Patrick Blumenberg, Thomas Graave, Tim Fingscheidt
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 |
|---|---|---|---|
| use signed absmax normalization (BOF4-S) instead of absolute (ablation) | BOF4 (MSE) · WikiText-2 Llama-3.1 8B | perplexity: 8.51 -> 8.46 | Improved |
| add outlier-preserving quantization (OPQ) (ablation) | BOF4-S (MSE) · WikiText-2 Llama-3.1 8B | perplexity: 8.46 -> 8.43 | Improved |
| use EM-optimized BOF4 codebook instead of NF4 | block-wise absmax quantization · WikiText-2 Llama-3.1 8B | perplexity: 8.53 -> 8.51 | Improved |
| optimize for MSE instead of MAE (ablation) | BOF4-S · WikiText-2 | perplexity: 8.49 -> 8.46 (Llama-3.1 8B) | Improved |
| fix reconstruction levels 0, -1 and 1 (ablation) | BOF4 · WikiText-2 Llama-3.1 8B | perplexity: 8.81 -> 8.51 | Improved |
| fix reconstruction levels 0, -1 and 1 (ablation) | BOF4 · Weight MSE Llama-3.1 8B | MSE (1e-6): 1.506 -> 1.566 | Hurt |
| use BOF4-S + OPQ instead of NF4 | QLoRA Llama-3.2 3B · HumanEval+/MBPP+ average | accuracy: 29.3 -> 32.0 | Improved |
| use BOF4-S + OPQ instead of BOF4 (MSE) alone | QLoRA Llama-3.2 3B · IFEval average | accuracy: 31.6 -> 30.0 | Hurt |
| use BOF4-S + OPQ instead of NF4 on Qwen-2.5 3B | inference quantization · NAV ACC | normalized accuracy: 38.2 -> 39.7 | Mixed |
Closest papers by meaning
- BlockDialect: Block-wise Fine-grained Mixed Format Quantization for Energy-Efficient LLM Inference (ICML 2025 · similarity 0.92)
- Norm Tweaking: High-Performance Low-Bit Quantization of Large Language Models (AAAI 2024 · similarity 0.90)
- Q-Palette: Fractional-Bit Quantizers Toward Optimal Bit Allocation for Efficient LLM Deployment (NeurIPS 2025 · similarity 0.90)
- Optimizing Large Language Model Training Using FP4 Quantization (ICML 2025 · similarity 0.90)
- CBQ: Cross-Block Quantization for Large Language Models (ICLR 2025 · similarity 0.90)
- DuQuant: Distributing Outliers via Dual Transformation Makes Stronger Quantized LLMs (NeurIPS 2024 · similarity 0.90)
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.
