Research map / Vision-language and multimodal / Multimodal LLMs
SAFE-Pruner: Semantic Attention–Guided Future-Aware Token Pruning for Efficient Vision-Language-Action Manipulation
Shilin Ma, Chubin Zhang, Changyuan Wang, Yuji Wang, Yue Wu, Zixuan Wang, Jingqi Tian, Zheng Zhu, Yansong Tang
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 future-layer saliency forecast (ablation) | SAFE-Pruner on OpenVLA-OFT · LIBERO (4 suites avg) | success rate: 94.5 -> 96.4 | Improved |
| add future-layer saliency forecast (ablation) | SAFE-Pruner on OpenVLA-OFT · LIBERO (4 suites avg) | FLOPs (T): 2.141 -> 1.722 | Improved |
| adaptive key timestep selection instead of fixed interval (ablation) | SAFE-Pruner on OpenVLA-OFT · LIBERO (4 suites avg) | success rate: 95.8 -> 96.4 | Improved |
| adaptive key timestep selection instead of fixed interval (ablation) | SAFE-Pruner on OpenVLA-OFT · LIBERO (4 suites avg) | latency (ms): 51.98 -> 37.22 | Improved |
| SAFE-Pruner vs best prior pruning | OpenVLA-OFT · LIBERO (4 suites avg) | success rate: 94.5 (FastV) -> 96.4 | Improved |
| SAFE-Pruner vs FastV on real robot | pi0.5 Astribot S1 · real robot avg | success rate: 69.0 -> 79.3 | Improved |
| SAFE-Pruner vs unpruned model | OpenVLA-OFT · LIBERO (4 suites avg) | success rate: 96.8 -> 96.4 | Hurt |
| random token pruning (ablation) | CogACT · SIMPLER Visual Matching | success rate: 74.8 -> 20.9 | Hurt |
| combine SAFE-Pruner with HiF8 quantization (ablation) | OpenVLA-OFT · LIBERO (4 suites avg) | GPU memory (GB): 18.31 -> 11.03 | Improved |
| SAFE-Pruner on SIMPLER Variant Aggregation vs VLA-Cache | CogACT · SIMPLER Variant Aggregation | success rate: 62.3 (VLA-Cache) vs 61.9 | 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.
- VLA-Cache — VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025) · better on 1 of 1 shared result
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
- When and What to Prune? Stage-Aware Visual Token Pruning for Efficient VLA (NeurIPS 2026 · similarity 0.93)
- LearnPruner: Rethinking Attention-based Token Pruning in Vision Language Models (ICLR 2026 · similarity 0.92)
- Each Complexity Deserves a Pruning Policy (NeurIPS 2025 · similarity 0.91)
- TopV: Compatible Token Pruning with Inference Time Optimization for Fast and Low-Memory Multimodal Vision Language Model (CVPR 2025 · similarity 0.91)
- ZOO-Prune: Training-Free Token Pruning via Zeroth-Order Gradient Estimation in Vision-Language Models (CVPR 2026 · similarity 0.91)
- Balanced Token Pruning: Accelerating Vision Language Models Beyond Local Optimization (NeurIPS 2025 · similarity 0.90)
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.
