Research map / Vision-language and multimodal / Robot action and navigation
EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EfficientVLA, a training-free framework that combines LLM layer pruning, task-aware diversity-based visual token pruning, and static caching of diffusion action-head features to accelerate CogACT (1.93x speedup, 28.9% FLOPs) on SIMPLER.
Problem
Diffusion-based Vision-Language-Action models have high computational and memory costs at inference, and existing acceleration methods target isolated modules, limiting end-to-end speedup for robot deployment.
Approach
Jointly exploit depth-wise redundancy in the LLM (cosine-similarity-based layer importance pruning), task-relevance plus diversity visual token selection, and static N-step caching of attention/MLP features in the diffusion action head, all without retraining.
Key results
- On CogACT in SIMPLER Visual Matching, EfficientVLA (L=22, T=56) reaches 1.93x speedup and 28.9% FLOPs with a 0.6-point average success drop (74.8% to 74.2%).
- VLA-Cache on CogACT achieves only 1.38x speedup, while FastV reaches 1.21x, per the paper's comparison.
- On CogACT-Large, EfficientVLA reports about 2.0x inference speedup with success dropping from 76.7% to 76.1%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply EfficientVLA with 22 LLM layers and 56 visual tokens | CogACT · SIMPLER Visual Matching | success rate (Average): 74.8% -> 74.2% | Hurt |
| Apply EfficientVLA with 22 LLM layers and 56 visual tokens | CogACT · SIMPLER Visual Matching | speedup: 1.00× -> 1.93× | Helped |
| Apply EfficientVLA with 22 LLM layers and 56 visual tokens | CogACT · SIMPLER Variant Aggregation | success rate (Average): 61.3% -> 61.2% | Hurt |
| Apply EfficientVLA with 22 LLM layers and 56 visual tokens | CogACT · SIMPLER Variant Aggregation | speedup: 1.00× -> 1.91× | Helped |
| Replace guided token selection with random retention of 112 visual tokens | CogACT · SIMPLER Visual Matching, pick coke can | success rate: 91.3% -> 9.7% | Hurt |
| Retain 56 visual tokens instead of 256 (ablation) | EfficientVLA · SIMPLER Visual Matching, pick coke can | inference time: 0.2342 -> 0.1866 | Helped |
| Retain 56 visual tokens instead of 256 (ablation) | EfficientVLA · SIMPLER Visual Matching, pick coke can | success rate: 91.3% -> 95.0% | Helped |
| Increase action-cache interval from 1 to 5 (ablation) | EfficientVLA · SIMPLER Visual Matching, pick coke can | inference time: 0.2342 -> 0.1909 | Helped |
| Increase action-cache interval from 1 to 5 (ablation) | EfficientVLA · SIMPLER Visual Matching, pick coke can | success rate: 91.3% -> 93.7% | Mixed |
| Enable visual token pruning only (ablation) | EfficientVLA · SIMPLER Visual Matching, pick coke can | success rate: 91.3% -> 95.6% | Helped |
| Enable visual token pruning only (ablation) | EfficientVLA · SIMPLER Visual Matching, pick coke can | speedup: 1.00× -> 1.25× | Helped |
| Enable action caching only (ablation) | EfficientVLA · SIMPLER Visual Matching, pick coke can | speedup: 1.00× -> 1.23× | Helped |
Surprising findings
- Pruning visual tokens and LLM layers sometimes raised success on the pick coke can task (e.g., 91.3% to 95.6% with 56 tokens), which the authors attribute to parameter redundancy.
Limitations
- Training-free approach may not reach the compression or speedup of training-aware methods.
- Fixed cache interval N trades acceleration against action fidelity; adaptive caching is not explored.
- Evaluation is limited to CogACT and the SIMPLER Google robot tasks due to few open diffusion-based VLAs.
Experimental setup
Datasets: VLA training samples (used for layer importance calibration; unspecified)
Benchmarks: SIMPLER
Metrics: success rate FLOPs inference time speedup
Baselines: CogACT Random Dropping FastV VLA-Cache
Builds on: CogACT FastV VLA-Cache PruneNet
Closest papers by meaning
- CogVLA: Cognition-Aligned Vision-Language-Action Models via Instruction-Driven Routing & Sparsification (NeurIPS 2025 · similarity 0.93)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.93)
- Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
- D-VLA: A High-Concurrency Distributed Asynchronous RL Framework for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · similarity 0.91)
- Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
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.