Research map / Architectures and efficiency / Quantization
Stabilizing Ultra-Low-Bit Quantization of Multimodal LLMs via Global Bit Allocation
Guilin Li, Yuexiao Ma, Yue Zhang, Xinxiong Wu, Jiaqi Zhou, Qingheng Zhang, Yan Zhang, Fei Chao, Xiawu Zheng, Rongrong Ji
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 GloBitQ mixed-precision (3.01-bit) instead of uniform per-channel GPTQ/VLMQ at 3-bit | Qwen2-VL-7B-Instruct · 6-benchmark average, 3-bit | average score: 64.74 -> 81.70 (VLMQ), full precision 83.76 | Improved |
| use GloBitQ instead of VLMQ at 3-bit | Qwen2.5-VL-7B-Instruct · 6-benchmark average, 3-bit | average score: 77.49 -> 78.3 | Mixed |
| use GloBitQ instead of VLMQ at 3-bit | LLaVA-OneVision · 6-benchmark average, 3-bit | average score: 72.17 -> 76.66 | Improved |
| use GloBitQ per-channel 2.01-bit instead of VLMQ group-128 at 2-bit | Qwen2-VL-7B-Instruct · 6-benchmark average, 2-bit | average score: 64.0 -> 66.6 | Improved |
| use GloBitQ per-channel 2.01-bit instead of VLMQ group-128 | Qwen2.5-VL-7B-Instruct · 6-benchmark average, 2-bit | average score: 64.6 -> 66.4 | Improved |
| use GloBitQ per-channel 2.01-bit instead of VLMQ group-128 | LLaVA-OneVision · 6-benchmark average, 2-bit | average score: 61.5 -> 60.8 | No clear effect |
| raise average bit-width from 2.01 to 2.05 (ablation) | GloBitQ · LLaVA-OneVision 2-bit, 6-benchmark average | average score: 60.8 -> 63.1 | Improved |
| add global importance-balanced mixed precision (2.01 bit) over uniform 2-bit FlatQuant/GPTQ (ablation) | Qwen2-VL-7B-Instruct · 6-benchmark average, 2-bit | average score: 53.5 -> 66.6 (GPTQ), 51.1 -> 66.6 (FlatQuant) | Improved |
| use per-channel GPTQ at 2.01-bit instead of GloBitQ | Qwen2-VL, Qwen2.5-VL, LLaVA-OneVision · 2-bit | average score: NAN (collapse) | Hurt |
| set balance factor alpha to 0.5 rather than 0.2-0.3 or 0.7-0.8 (ablation) | GloBitQ · LLaVA-OneVision-Int2.01 | average score | 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.
- GPTAQ — GPTAQ: Efficient Finetuning-Free Quantization for Asymmetric Calibration (ICML 2025) · better on 3 of 3 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
- Q-Palette: Fractional-Bit Quantizers Toward Optimal Bit Allocation for Efficient LLM Deployment (NeurIPS 2025 · similarity 0.91)
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.91)
- ScaleBITS: Scalable Bitwidth Search for Hardware-Aligned Mixed-Precision LLMs (NeurIPS 2026 · similarity 0.91)
- CBQ: Cross-Block Quantization for Large Language Models (ICLR 2025 · similarity 0.91)
- ABQ-LLM: Arbitrary-Bit Quantized Inference Acceleration for Large Language Models (AAAI 2025 · similarity 0.90)
- Norm Tweaking: High-Performance Low-Bit Quantization of Large Language Models (AAAI 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.
