Research map / Architectures and efficiency / Pruning and efficient tuning
SpecPrune-VLA: Accelerating Vision-Language-Action Models via Action-Aware Self-Speculative Pruning
Hanzhen Wang, Jiaming Xu, Yushun Xiang, Jiayi Pan, Yongkang Zhou, Yong-Lu Li, Guohao Dai
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 global attention reuse from the previous step to static pruning (ablation) | SpecPrune-VLA · LIBERO | success rate (%): 93.4% -> 96.1% | Improved |
| use entropy-based layer weighting instead of average weighting in dynamic pruning (ablation) | SpecPrune-VLA · LIBERO | success rate (%): 92.0% -> 96.1% | Improved |
| add the action-aware controller on top of static and dynamic pruning (ablation) | SpecPrune-VLA · LIBERO-Spatial | success rate (%): 96.8 -> 97.4 | Improved |
| add action-aware controller (ablation) | SpecPrune-VLA · LIBERO-Spatial | latency: 70.8ms -> 72.3ms | Mixed |
| guide pruning with mid and deep layer attention instead of shallow layers (ablation) | post-hoc token pruning on OpenVLA-OFT · LIBERO pick and place | success rate | Improved |
| random token pruning beyond 12.5% sparsity (ablation) | OpenVLA-OFT · LIBERO pick and place | success rate | Hurt |
| use first two layers rather than first layer for local speculation (ablation) | SpecPrune-VLA · LIBERO | hit rate | Improved |
| add a third layer for local speculation (ablation) | SpecPrune-VLA · LIBERO | hit rate and latency | No clear effect |
| lower prune ratio alpha (more tokens pruned) (ablation) | SpecPrune-VLA · LIBERO | success rate vs speedup | Mixed |
| apply SpecPrune-VLA instead of FastV at similar token counts | OpenVLA-OFT · LIBERO | average success rate (%): 93.3% -> 96.1% | Improved |
| apply SpecPrune-VLA to OpenVLA-OFT | OpenVLA-OFT · LIBERO | average speedup: 1.00× -> 1.46× | Improved |
| apply SpecPrune-VLA on real robot | OpenVLA-OFT · Real-world tasks | latency (ms): 187.7 -> 110.2 | Improved |
Closest papers by meaning
- EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.94)
- When and What to Prune? Stage-Aware Visual Token Pruning for Efficient VLA (NeurIPS 2026 · similarity 0.93)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- Each Complexity Deserves a Pruning Policy (NeurIPS 2025 · similarity 0.91)
- CogVLA: Cognition-Aligned Vision-Language-Action Models via Instruction-Driven Routing & Sparsification (NeurIPS 2025 · similarity 0.91)
- Prune2Drive: A Plug-and-Play Framework for Accelerating Vision-Language Models in Autonomous Driving (CVPR 2026 · 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.
