Research map / Architectures and efficiency / Quantization
NeUQI: Near-Optimal Uniform Quantization Parameter Initialization for Low-Bit LLMs
Li Lin, Xinyu Hu, Xiaojun Wan
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 |
|---|---|---|---|
| replace Min-Max-style integer-zero-point initialization with NeUQI (float zero-point, optimized scale) | GPTQ-style PTQ on LLaMA 2 7B · 2-bit channel-wise, Wiki2 | perplexity: 6953 -> 17.14 (GPTQ vs NeUQI) | Improved |
| relax the integer zero-point constraint (NeUQI vs Int-Search with the same loss) (ablation) | NeUQI · LLaMA 2 7B 2-bit channel-wise, Wiki2 | perplexity: 26.26 -> 17.14 | Improved |
| relax the integer zero-point constraint (NeUQI vs Int-Search) (ablation) | NeUQI · LLaMA 2 70B 2-bit channel-wise, average accuracy | accuracy: 61.61 -> 65.13 | Improved |
| use NeUQI instead of other init methods at same group-wise bit budget (group 128 vs integer-constrained group 64) (ablation) | LLaMA 2 7B 2-bit · average accuracy | accuracy: 49.73 -> 51.33 (best group-64 baseline GPTAQ vs NeUQI) | Improved |
| add NeUQI initialization to lightweight distillation | 2-bit g128 LLaMA 2 7B · Wiki2 | perplexity: 14.95 -> 8.38 (GPTQ vs NeUQI distillation) | Improved |
| replace EfficientQAT initialization with NeUQI | EfficientQAT on LLaMA 2 7B · Wiki2 | perplexity: 7.19 -> 6.81 | Improved |
| only allow floating zero-point in EfficientQAT without NeUQI initialization (ablation) | EfficientQAT on LLaMA 2 7B · average accuracy | accuracy: 59.50 -> 59.37 | No clear effect |
| freeze scale in first stage of EfficientQAT with NeUQI init (ablation) | EfficientQAT+NeUQI on LLaMA 3 8B · Wiki2 | perplexity: 9.14 -> 8.87 | Improved |
| add coarse-to-fine scale search (ablation) | NeUQI · LLaMA 2 7B 2-bit, relative runtime | relative runtime: 0.548 -> 0.027 for Q projection, loss 1.00197 -> 1.00193 | Improved |
| add transition-point reduction (ablation) | NeUQI zero-point search · LLaMA 2 7B 2-bit | runtime: nearly 50% runtime reduction | Improved |
| add NeUQI 3-bit to larger model versus smaller BF16 model at same memory | Qwen 2.5 14B · average accuracy vs 3B BF16 | accuracy: 64.99 -> 69.12 | Improved |
| combine NeUQI with Hadamard transform at W4A4 | LLaMA 2 7B · Wiki2 | perplexity: 6.02 -> 5.99 (MagR vs NeUQI) | No clear effect |
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.
- PV-tuning — PV-Tuning: Beyond Straight-Through Estimation for Extreme LLM Compression (NeurIPS 2024) · better on 2 of 2 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
- ABQ-LLM: Arbitrary-Bit Quantized Inference Acceleration for Large Language Models (AAAI 2025 · similarity 0.92)
- GANQ: GPU-Adaptive Non-Uniform Quantization for Large Language Models (ICML 2025 · similarity 0.92)
- BiLLM: Pushing the Limit of Post-Training Quantization for LLMs (ICML 2024 · similarity 0.91)
- Exploring Post-training Quantization in LLMs from Comprehensive Study to Low Rank Compensation (AAAI 2024 · similarity 0.91)
- SqueezeLLM: Dense-and-Sparse Quantization (ICML 2024 · similarity 0.91)
- PV-Tuning: Beyond Straight-Through Estimation for Extreme LLM Compression (NeurIPS 2024 · similarity 0.91)
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.
