Research map / Architectures and efficiency / Quantization
ProQuant: Progressive Quantization-aware Training for Edge MLLMs
Yufei Xue, Yushi Huang, Jiawei Shao, Pingcheng Dong, Yonghao Tan, Shiyao Li, Kwang-Ting Cheng, Jun Zhang
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 ProQuant, a progressive quantization-aware training approach for compressing multi-modal LLMs for edge deployment.
Problem
Multi-modal LLMs have heavy compute and memory requirements that hinder deployment on resource-constrained edge devices; post-training quantization is one route, but the abstract is truncated before its limitations are stated.
Approach
Uses progressive quantization-aware training for MLLMs, as the title indicates; the specific mechanism is not described in the available abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add self-relabeling (SRL) to naive QAT (ablation) | naive QAT · Qwen3-VL-2B W4A8 / W4A4 | average accuracy: +1.00 (W4A8), +3.36 (W4A4) | Improved |
| add progressive block activation (PBA) to naive QAT (ablation) | naive QAT · Qwen3-VL-2B W4A8 / W4A4 | average accuracy: +0.67 (W4A8), +3.02 (W4A4) | Improved |
| combine SRL and PBA (ablation) | naive QAT · Qwen3-VL-2B W4A8 / W4A4 | average accuracy: +1.44 (W4A8), +6.07 (W4A4) | Improved |
| prioritize vision-sensitive blocks (high tau) instead of low tau (ablation) | ProQuant · Qwen3-VL-2B W4A8 | average accuracy: 72.03 -> 72.54 | Improved |
| use tau instead of weight gradient norm as activation metric (ablation) | ProQuant · Qwen3-VL-2B W4A8 | average accuracy: 72.19 -> 72.54 | Improved |
| use logarithmic instead of linear scheduler (ablation) | ProQuant · Qwen3-VL-2B W4A8 | average accuracy: 71.00 -> 72.54 | Improved |
| treat each decoding layer as unit vs two layers (ablation) | ProQuant · Qwen3-VL-2B W4A8 | average accuracy: 70.28 -> 72.54 | Improved |
| split decoding layer into attn and ffn units (finer granularity) (ablation) | ProQuant · Qwen3-VL-2B W4A8 | average accuracy: 72.54 -> 72.15 | Hurt |
| use self labels instead of GPT-4 labels (ablation) | ProQuant · Qwen3-VL-2B W4A4 | average accuracy: 56.78 -> 59.83 | Improved |
| use self labels instead of 8B homogeneous labels (ablation) | ProQuant · Qwen3-VL-2B W4A4 | average accuracy: 59.04 -> 59.83 | Improved |
| keep down_proj activations at 8-bit | W4A4 ProQuant · three Qwen VL backbones | average accuracy: +1.44, +9.39, +3.21 | Improved |
| use ProQuant QAT instead of MBQ PTQ | W8A8 quantization · Qwen3-VL-2B W8A8 | average accuracy: 73.95 (MBQ) vs 72.88 (ProQuant) | 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.
- SmoothQuant — SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models (ICML 2023) · 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
- Fine-Grained Post-Training Quantization for Large Vision Language Models with Quantization-Aware Integrated Gradients (CVPR 2026 · similarity 0.89)
- VLM-PTQ: Efficient Post-Training Quantization for Large Vision-Language Models (CVPR 2026 · similarity 0.89)
- GuidedQuant: Large Language Model Quantization via Exploiting End Loss Guidance (ICML 2025 · similarity 0.89)
- LeanQuant: Accurate and Scalable Large Language Model Quantization with Loss-error-aware Grid (ICLR 2025 · similarity 0.88)
- MBQ: Modality-Balanced Quantization for Large Vision-Language Models (CVPR 2025 · similarity 0.88)
- Exploring Post-training Quantization in LLMs from Comprehensive Study to Low Rank Compensation (AAAI 2024 · 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.
