mlsanityData to 7 Oct 2026

SpecPrune-VLA: Accelerating Vision-Language-Action Models via Action-Aware Self-Speculative Pruning

arXiv 2509.05614 ↗First posted 6 Sep 2025cs.CV cs.AI cs.ROcited by 39

On the map: Vision-language-action models → Efficient VLA inference and adaptation

Abstract (arXiv)

Pruning is a typical acceleration technique for compute-bound models by removing computation on unimportant values. Recently, it has been applied to accelerate Vision-Language-Action (VLA) model inference. However, existing acceleration methods focus on local information from the current action step and ignore the global context, leading to >20% success rate drop and limited speedup in some scenarios. In this paper, we point out spatial-temporal consistency in VLA tasks: input images in consecutive steps exhibit high similarity, and propose the key insight that token selection should combine local information with global context of the model. Based on this, we propose SpecPrune-VLA, a training-free, two-level pruning method with heuristic control. (1) Action-level static pruning. We leverage global history and local attention to statically reduce visual tokens per action. (2) Layer-level dynamic pruning. We prune tokens adaptively per layer based on layer-wise importance. (3) Lightweight action-aware controller: We classify actions as coarse- or fine-grained by the speed of the end effector and adjust pruning aggressiveness accordingly. Extensive experiments show that SpecPrune-VLA achieves up to 1.57× speedup in LIBERO simulation and 1.70× on real-world tasks, with negligible success rate degradation.

Key facts. Cited by 39 papers in the map (data through 7 Oct 2026). Placed in Efficient VLA inference and adaptation, one of 13 directions in Vision-language-action models. Builds on 40 of its 58 references that are in the map. First posted on arXiv 6 Sep 2025.

Builds on

40 of 58 references.

Closest ancestors · 4

Same direction, newest first

  1. 2025SP-VLA: A Joint Model Scheduling and Token Pruning Approach for VLA Model Acceleration15 Jun
  2. 2025EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models11 Jun
  3. 2025VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching4 Feb
  4. 2024Quantization-Aware Imitation-Learning for Resource-Efficient Robotic Control2 Dec

Other references · 36

  1. 2017Attention Is All You Needcited by 59,932
  2. 2021LoRA: Low-Rank Adaptation of Large Language Modelscited by 14,084
  3. 2023Llama 2: Open Foundation and Fine-Tuned Chat Modelscited by 12,865
  4. 2023Visual Instruction Tuningcited by 7,792
  5. 2023DINOv2: Learning Robust Visual Features without Supervisioncited by 7,764
  6. 2022Scalable Diffusion Models with Transformerscited by 5,931
  7. 2022Flow Matching for Generative Modelingcited by 5,426
  8. 2024Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolutioncited by 4,151
  9. 2023Sigmoid Loss for Language Image Pre-Trainingcited by 3,048
  10. 2023RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Controlcited by 2,993
  11. 2024OpenVLA: An Open-Source Vision-Language-Action Modelcited by 2,777
  12. 2024π0: A Vision-Language-Action Flow Model for General Robot Controlcited by 2,297
24 more
  1. 2022RT-1: Robotics Transformer for Real-World Control at Scalecited by 2,007
  2. 2022Fast Inference from Transformers via Speculative Decodingcited by 1,615
  3. 2023Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardwarecited by 1,490
  4. 2024Octo: An Open-Source Generalist Robot Policycited by 1,427
  5. 2023LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learningcited by 1,338
  6. 2023Open X-Embodiment: Robotic Learning Datasets and RT-X Modelscited by 849
  7. 2025Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Successcited by 844
  8. 2024An Image is Worth 1/2 Tokens After Layer 2: Plug-and-Play Inference Acceleration for Large Vision-Language Modelscited by 629
  9. 2024Evaluating Real-World Robot Manipulation Policies in Simulationcited by 484
  10. 2024CogACT: A Foundational Vision-Language-Action Model for Synergizing Cognition and Action in Robotic Manipulationcited by 411
  11. 2025UniVLA: Learning to Act Anywhere with Task-centric Latent Actionscited by 386
  12. 2024SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inferencecited by 378
  13. 2024LayerSkip: Enabling Early Exit Inference and Self-Speculative Decodingcited by 302
  14. 2024A Survey on Efficient Inference for Large Language Modelscited by 197
  15. 2025DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Modelscited by 177
  16. 2024FORA: Fast-Forward Caching in Diffusion Transformer Accelerationcited by 127
  17. 2024Fit and Prune: Fast and Training-free Visual Token Pruning for Multi-modal Large Language Modelscited by 125
  18. 2024DeeR-VLA: Dynamic Inference of Multimodal Large Language Models for Efficient Robot Executioncited by 116
  19. 2024What Matters in Building Vision-Language-Action Models for Generalist Robotscited by 88
  20. 2024Large Language Model Inference Acceleration: A Comprehensive Hardware Perspectivecited by 36
  21. 2024Attention Entropy is a Key Factor: An Analysis of Parallel Context Encoding with Full-attention-based Pre-trained Language Modelscited by 26
  22. 2025SpecEE: Accelerating Large Language Model Inference with Speculative Early Exitingcited by 23
  23. 2025AdaSkip: Adaptive Sublayer Skipping for Accelerating Long-Context LLM Inferencecited by 12
  24. 2025Accelerating LLM Inference Throughput via Asynchronous KV Cache Prefetchingcited by 4

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2025: 8; 2026: 31.

39 papers.

Sort:

By area:

  1. 2025A Survey on Efficient Vision-Language-Action Modelscited by 34
  2. 2025Efficient Vision-Language-Action Models for Embodied Manipulation: A Systematic Surveycited by 25
  3. 2025AVA-VLA: Improving Vision-Language-Action models with Active Visual Attentioncited by 22
  4. 2026FASTER: Rethinking Real-Time Flow VLAscited by 20
  5. 2025Bridging the Semantic-Action Gap in Visual Token Pruning for Efficient VLA Inferencecited by 14
  6. 2026DyQ-VLA: Temporal-Dynamic-Aware Quantization for Embodied Vision-Language-Action Modelscited by 10
  7. 2026KERV: Kinematic-Rectified Speculative Decoding for Embodied VLA Modelscited by 9
  8. 2026VLA-Forget: Vision-Language-Action Unlearning for Embodied Foundation Modelscited by 9
  9. 2026TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAMcited by 9
  10. 2026DySL-VLA: Efficient Vision-Language-Action Model Inference via Dynamic-Static Layer-Skipping for Robot Manipulationcited by 8
  11. 2026Realtime-VLA FLASH: Speculative Inference Framework for Diffusion-based VLAscited by 8
  12. 2025Compressor-VLA: Instruction-Guided Visual Token Compression for Efficient Robotic Manipulationcited by 7
27 more
  1. 2026Shallow-π: Knowledge Distillation for Flow-based VLAscited by 7
  2. 2025Token Expand-Merge: Training-Free Token Compression for Vision-Language-Action Modelscited by 6
  3. 2026FocusVLA: Focused Visual Utilization for Vision-Language-Action Modelscited by 5
  4. 2025FlashEdit: Decoupling Speed, Structure, and Semantics for Precise Image Editingcited by 4
  5. 2025An Anatomy of Vision-Language-Action Models: From Modules to Milestones and Challengescited by 4
  6. 2026RAPID: Redundancy-Aware and Compatibility-Optimal Edge-Cloud Partitioned Inference for Diverse VLA Modelscited by 4
  7. 2026Spatial Memory for Out-of-Vision Manipulation in Vision-Language-Actioncited by 4
  8. 2026SAFE-Pruner: Semantic Attention-Guided Future-Aware Token Pruning for Efficient Vision-Language-Action Manipulationcited by 4
  9. 2026vla.cpp: A Unified Inference Runtime for Vision-Language-Action Modelscited by 4
  10. 2026Drop-Then-Recovery: How Redundant Are Vision-Language-Action Models?cited by 3
  11. 2026Jetson-PI: Towards Onboard Real-Time Robot Control via Foresight-Aligned Asynchronous Inferencecited by 3
  12. 2026FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inferencecited by 3
  13. 2026ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Modelscited by 2
  14. 2026PolicyTrim: Boosting Intrinsic Policy Efficiency of Vision-Language-Action Modelscited by 2
  15. 2026The Gate, Not the Cache: Gate Provenance Bounds the Closed-Loop Reliability of Training-Free VLA Token Skippingcited by 2
  16. 2026Environment-Aware Adaptive Pruning with Interleaved Inference Orchestration for Vision-Language-Action Modelscited by 1
  17. 2026DepthCache: Depth-Guided Training-Free Visual Token Merging for Vision-Language-Action Model Inferencecited by 1
  18. 2026Beyond Attention Magnitude: Leveraging Inter-layer Rank Consistency for Efficient Vision-Language-Action Modelscited by 1
  19. 2026Finetuning Vision-Language-Action Models Requires Fewer Layers Than You Thinkcited by 1
  20. 20262D or 3D: Who Governs Salience in VLA Models? -- Tri-Stage Token Pruning Framework with Modality Salience Awarenesscited by 0
  21. 2026The Speedup Paradox: Rethinking Inference Speed-Quality Trade-off in Embodied Taskscited by 0
  22. 2026Just Noticeable Difference Modeling for Token Compression in Vision-Language-Action Modelscited by 0
  23. 2026rMuscle: Robotic Muscle Memory for Efficient Vision-Language-Action Model Inferencecited by 0
  24. 2026Catch Me If You Can: Real-Time Feedback Denoising for Responsive VLAscited by 0
  25. 2026Sparse-WAM: Accelerating World Action Models via Action-Guided Sparse Imaginationcited by 0
  26. 2026When and What to Prune? Stage-Aware Visual Token Pruning for Efficient VLAcited by 0
  27. 2026VLA-ACL: Action-Consistent Visual Token Pruning for Efficient Vision-Language-Action Modelscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025SQAP-VLA: A Synergistic Quantization-Aware Pruning Framework for High-Performance Vision-Language-Action Models
  2. 2025The Better You Learn, The Smarter You Prune: Towards Efficient Vision-language-action Models via Differentiable Token Pruning
  3. 2025Think Twice, Act Once: Token-Aware Compression and Action Reuse for Efficient Inference in Vision-Language-Action Models
  4. 2025Action-aware Dynamic Pruning for Efficient Vision-Language-Action Manipulation
  5. 2026BFA++: Hierarchical Best-Feature-Aware Token Prune for Multi-View Vision Language Action Model
  6. 2026Beyond Token Importance: Preserving Spatial Scaffolds for Efficient Vision-Language-Action Inference
  7. 2026Look Before Acting: Enhancing Vision Foundation Representations for Vision-Language-Action Models
  8. 2026Learning to Accelerate Vision-Language-Action Models through Adaptive Visual Token Caching