Research map / Vision-language and multimodal / Multimodal LLMs
Bi-VLM: Binary Post-Training Quantization for Vision-Language Models
Xijun Wang, Rayyan Abdalla, Junyun Huang, Chengyuan Zhang, Ruiqi Xian, Dinesh Manocha
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Bi-VLM, a binary post-training quantization method for vision-language models aimed at ultra-low-bit weight precision.
Problem
The computational demands of vision-language models clash with the need for ultra-low-bit weight precision (bitwidth-level compression) for deployment.
Approach
Binary post-training quantization of VLM weights, targeting ultra-low-bit (near 1-bit) weight precision without retraining. Details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add hybrid precision quantization (2-bit salient weights) to fixed binary (ablation) | Bi-VLM · ScienceQA-IMG ablation (Llama 3.2-Vision-11B) | accuracy: 6.33 -> 13.74 | Improved |
| add adaptive saliency search (ablation) | Bi-VLM · ScienceQA-IMG ablation (Llama 3.2-Vision-11B) | accuracy: 13.74 -> 49.51 | Improved |
| add multiple unsalient search (ablation) | Bi-VLM · ScienceQA-IMG ablation (Llama 3.2-Vision-11B) | accuracy: 49.51 -> 58.35 | Improved |
| quantize only the language model vs only the vision model (ablation) | Bi-VLM · ScienceQA-IMG, component quantized (Llama 3.2-Vision-11B) | accuracy: 85.23 -> 58.75 | Hurt |
| quantize adaptor/projector only (ablation) | Bi-VLM · ScienceQA-IMG, component quantized (Llama 3.2-Vision-11B) | accuracy: 85.82 -> 85.86 | No clear effect |
| use Bi-VLM instead of BiLLM for whole-VLM binarization | Llama 3.2-Vision-11B · ScienceQA-IMG (Llama 3.2-Vision-11B) | accuracy: 17.4 -> 58.35 | Improved |
| use Bi-VLM instead of BiLLM for quantization procedure | Llama 3.2-Vision-11B · ScienceQA-IMG (Llama 3.2-Vision-11B) | quantization time: 139.8 Mins -> 9.6 Mins | Improved |
| use Bi-VLM instead of BiLLM on MME Cognition | Llama 3.2-Vision-11B · MME Cognition (whole VLM quantized, 1-1.1 bit) | score: 222.85 -> 196.07 | Hurt |
| use Bi-VLM instead of BiLLM on MME Perception | Qwen2.5-VL-7B-Instruct · MME Perception (whole VLM quantized, 1-1.1 bit) | score: 828.54 -> 1458.70 | Improved |
| use Bi-VLM instead of BiLLM on MMMU | LLaVA-OneVision-7B · MMMU (whole VLM quantized, 1-1.1 bit) | score: 28.78 -> 39.12 | Improved |
Closest papers by meaning
- VLM-PTQ: Efficient Post-Training Quantization for Large Vision-Language Models (CVPR 2026 · similarity 0.91)
- Fine-Grained Post-Training Quantization for Large Vision Language Models with Quantization-Aware Integrated Gradients (CVPR 2026 · similarity 0.91)
- SPHQuant: Efficient extreme low bit weight quantization for Vision-Language Models (NeurIPS 2026 · similarity 0.91)
- Q-VLM: Post-training Quantization for Large Vision-Language Models (NeurIPS 2024 · similarity 0.90)
- MBQ: Modality-Balanced Quantization for Large Vision-Language Models (CVPR 2025 · similarity 0.90)
- Breaking Modality Heterogeneity in Low-Bit Quantization for Large Vision-Language Models (NeurIPS 2026 · similarity 0.88)
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.
