Research map / Architectures and efficiency / Quantization
SPEED-Q: Staged Processing with Enhanced Distillation Towards Efficient Low-Bit On-Device VLM Quantization
Tianyu Guo, Shanwei Zhao, Shiai Zhu, Chenguang Ma
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 distillation loss to task loss (ablation) | SPEED-Q (InternVL3-1B, 4-bit) · MMStar | accuracy: 49.33 -> 51.07 | Improved |
| add distillation loss to task loss (ablation) | SPEED-Q (InternVL3-1B, 4-bit) · AI2D | accuracy: 66.13 -> 68.81 | Improved |
| add distillation loss to task loss (ablation) | SPEED-Q (InternVL3-1B, 4-bit) · OCRBench | accuracy: 75.5 -> 79.7 | Improved |
| add task loss to distillation loss (ablation) | SPEED-Q (InternVL3-1B, 4-bit) · MMStar | accuracy: 50.40 -> 51.07 | Improved |
| add task loss to distillation loss (ablation) | SPEED-Q (InternVL3-1B, 4-bit) · AI2D | accuracy: 68.46 -> 68.81 | Improved |
| add task loss to distillation loss (ablation) | SPEED-Q (InternVL3-1B, 4-bit) · OCRBench | accuracy: 77.9 -> 79.7 | Improved |
| use staged instead of joint quantization (ablation) | SPEED-Q (InternVL3-1B, 4-bit) · MMStar | accuracy: 49.67 -> 51.07 | Improved |
| use staged instead of joint quantization (ablation) | SPEED-Q (InternVL3-1B, 4-bit) · AI2D | accuracy: 68.46 -> 68.81 | Improved |
| use staged instead of joint quantization (ablation) | SPEED-Q (InternVL3-1B, 4-bit) · OCRBench | accuracy: 77.0 -> 79.7 | Improved |
| SPEED-Q instead of MBQ at 2-bit | InternVL3-1B · Avg. of 8 benchmarks (2-bit) | average accuracy: 3.81 -> 51.48 | Improved |
| SPEED-Q instead of MBQ at 4-bit | InternVL3-8B · Avg. of 8 benchmarks (4-bit) | average accuracy: 75.35 -> 75.04 | No clear effect |
| SPEED-Q instead of MBQ at 2-bit | InternVL3-8B · Avg. of 8 benchmarks (2-bit) | average accuracy: 61.66 -> 68.46 | Improved |
Closest papers by meaning
- VLM-PTQ: Efficient Post-Training Quantization for Large Vision-Language Models (CVPR 2026 · similarity 0.87)
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.87)
- Agile-Quant: Activation-Guided Quantization for Faster Inference of LLMs on the Edge (AAAI 2024 · similarity 0.87)
- QSVD: Efficient Low-rank Approximation for Unified Query-Key-Value Weight Compression in Low-Precision Vision-Language Models (NeurIPS 2025 · similarity 0.87)
- Gated Relational Alignment via Confidence-based Distillation for Efficient VLMs (ICML 2026 · similarity 0.87)
- SPHQuant: Efficient extreme low bit weight quantization for Vision-Language Models (NeurIPS 2026 · 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.
