Research map / Vision-language and multimodal / Cross-modal learning
MASQuant: Modality-Aware Smoothing Quantization for Multimodal Large Language Models
Lulu Hu, Wenhu Xiao, Xin Chen, Xinhua Xu, Bowen Xu, Kun Li, Yongliang Tao
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 |
|---|---|---|---|
| add modality-aware smoothing (MAS) instead of uniform smoothing (ablation) | MASQuant W4A8 · Librispeech (Omni-3B) | WER: 77.4 -> 3.8 | Improved |
| add modality-aware smoothing with learned optimization (ablation) | MASQuant W4A8 · OmniBench (Omni-3B) | accuracy: 33.3 -> 47.7 | Improved |
| add modality-aware smoothing with learned optimization (ablation) | MASQuant W4A8 · MMMU (VL-3B) | accuracy: 33.3 -> 46.7 | Improved |
| use equal modality loss weights (1.0/1.0) vs lowering vision weight to 0.5 (ablation) | MASQuant W4A8 · MMMU (VL-3B) | accuracy: 30.0 -> 37.8 | Improved |
| use equal modality loss weights vs lambda_v=0.1 (ablation) | MASQuant W4A8 · PPL (VL-3B) | PPL: 33.7 -> 17.2 | Improved |
| train smoothing factors 2 epochs instead of 1 (ablation) | MASQuant W4A8 · average accuracy (VL-3B) | average accuracy: 56.9 -> 61.2 | Improved |
| train more than 2 epochs (5 or 10) (ablation) | MASQuant W4A8 · average accuracy (VL-3B) | average accuracy: 61.2 -> 59.2 (5 epochs), 60.8 (10 epochs) | No clear effect |
| apply MASQuant vs MBQ at W4A8 on VL-3B | Qwen2.5-VL-3B · multimodal average | average accuracy: 64.6 -> 64.7 | No clear effect |
| apply MASQuant vs MBQ at W4A8 on VL-7B | Qwen2.5-VL-7B · multimodal average | average accuracy: 68.5 -> 69.0 | Improved |
| apply MASQuant vs MBQ at W8A8 on VL-3B | Qwen2.5-VL-3B · multimodal average | average accuracy: 70.1 -> 70.9 | Improved |
| apply MASQuant vs MBQ on omni audio at W4A8 | Qwen2.5-Omni-3B · Librispeech | WER: 9.5 -> 3.6 | Improved |
| apply MASQuant vs MBQ on Wenetspeech at W4A8 | Qwen2.5-Omni-3B · Wenetspeech | WER: 8.5 -> 8.7 | Hurt |
Closest papers by meaning
- Strong Helps Weak: Directional Cross-Modal Alignment Transfer in Multi-modal LLMs (NeurIPS 2026 · similarity 0.89)
- MBQ: Modality-Balanced Quantization for Large Vision-Language Models (CVPR 2025 · similarity 0.88)
- Diagnosing and Mitigating Modality Interference in Multimodal Large Language Models (NeurIPS 2026 · similarity 0.87)
- Explore How to Inject Beneficial Noise in MLLMs (AAAI 2026 · similarity 0.87)
- OSTQuant: Refining Large Language Model Quantization with Orthogonal and Scaling Transformations for Better Distribution Fitting (ICLR 2025 · similarity 0.87)
- Modality-Aware SAM: Sharpness-Aware-Minimization Driven Gradient Modulation for Harmonized Multimodal Learning (NeurIPS 2025 · similarity 0.87)
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.
