Research map / Video / Video-language understanding
ParallelVLM: Lossless Video-LLM Acceleration with Visual Alignment Aware Parallel Speculative Decoding
Quan Kong, Yuhao Shen, Yicheng Ji, Huan Li, Cong Wang
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 |
|---|---|---|---|
| raise draft pruning ratio from 0.0 to 0.9 (ablation) | ParallelVLM · LLaVA-OV 7B/72B | speedup: 2.94x -> 3.36x | Improved |
| raise draft pruning ratio from 0.0 to 0.9 (ablation) | ParallelVLM · Qwen2.5-VL 7B/32B | speedup: 2.08x -> 2.42x | Improved |
| raise draft pruning ratio from 0.0 to 0.9 (ablation) | draft model speed ratio · LLaVA-OV 7B/72B | speed ratio c: 5 -> 9 | Improved |
| push pruning ratio to 1.0 (prune all draft visual tokens) (ablation) | ParallelVLM · LLaVA-OV 7B/72B VideoDetailCaption | speedup: 3.43x -> 2.53x | Hurt |
| push pruning ratio to 0.99 (ablation) | ParallelVLM · LLaVA-OV 7B/72B VideoDetailCaption | speedup: 3.43x -> 2.90x | Hurt |
| use UV-Prune verifier-guided pruning instead of SpecVLM attention-guided pruning (ablation) | ParallelVLM · pruning alignment, alpha=0.9 | accepted length M: 0.2 drop for SpecVLM | Improved |
| add FlashAttention (ablation) | ParallelVLM · LLaVA-OV-72B, 8K sequence | speedup: 1.22x | Improved |
| add FlashAttention (ablation) | ParallelVLM · LLaVA-OV-72B, 32K sequence | speedup: 1.22x -> 1.41x | Improved |
| use ParallelVLM instead of SpecVLM | lossless SD · LLaVA-OV 7B/72B | avg speedup: 2.74x -> 3.36x | Improved |
| use ParallelVLM instead of OnSD | lossless SD · LLaVA-OV 7B/7B | avg speedup: 0.84x -> 1.55x | Improved |
| use ParallelVLM instead of FastV token pruning | LLaVA-OV-72B · 10% retention | speedup: 1.63x -> 3.36x | Improved |
| use ParallelVLM instead of FastV token pruning | LLaVA-OV-72B · 10% retention | token acceptance ratio: 84.6% -> 99.1% | Improved |
Closest papers by meaning
- ViSpec: Accelerating Vision-Language Models with Vision-Aware Speculative Decoding (NeurIPS 2025 · similarity 0.87)
- Sharp Eyes and Memory for VideoLLMs: Information-Aware Visual Token Pruning for Efficient and Reliable VideoLLM Reasoning (AAAI 2026 · similarity 0.87)
- DyCoke: Dynamic Compression of Tokens for Fast Video Large Language Models (CVPR 2025 · similarity 0.87)
- Scaling LLM Speculative Decoding: Non-Autoregressive Forecasting in Large-Batch Scenarios (AAAI 2026 · similarity 0.86)
- FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025 · similarity 0.86)
- ApET: Approximation-Error Guided Token Compression for Efficient VLMs (CVPR 2026 · similarity 0.86)
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.
