Don't Run with Scissors: Pruning Breaks VLA Models but They Can Be Recovered

Jason Jabbour; Dong-Ki Kim; Max Smith; Jay Patrikar; Radhika Ghosal; Youhui Wang; Ali Agha; Vijay Janapa Reddi

cs.ROOct 9, 2025

Don't Run with Scissors: Pruning Breaks VLA Models but They Can Be Recovered

Authors: Jason Jabbour, Dong-Ki Kim, Max Smith, Jay Patrikar, Radhika Ghosal, Youhui Wang, Ali Agha, Vijay Janapa Reddi, +1 more

Organizations: FieldAI · Harvard University

Abstract

Vision-Language-Action (VLA) models are a promising paradigm for field robotics, but their deployment is bottlenecked by strict edge SWaP-C constraints. Compression techniques such as quantization, distillation, and pruning are well-established for LLMs and VLMs. Pruning, however, remains particularly underexplored for VLA models. We surprisingly find that pruning VLA models yields drastic performance degradation and increased safety violations. Through spectral analysis, we identify structural properties of VLA weights that help explain this fragility. Based on this insight, we introduce GLUESTICK, a post-pruning recovery method that restores much of the original model's capability while preserving sparsity benefits. GLUESTICK computes a one-time corrective term by interpolating between the dense and pruned models in weight space. At inference, each pruned layer applies this correction to recover lost performance with minimal overhead. Our approach requires no additional training and is agnostic to the pruning algorithm. Across diverse VLA architectures and domains, GLUESTICK recovers up to 100% of the success rate and restores safety to near-dense levels, while saving up to 5 GB of memory and achieving up to 1.75x speedup. Additional material can be found at: https://gluestick-vla.github.io/

Figures & tables

Figure 1: VLA models break under pruning, and GLUESTICK fixes them. Pruning methods unexpectedly cause task and safety failures in VLAs: colliding with an object in a navigation task (left) , or dropping a bowl in a manipulation task (right) . Our training-free post-pruning method, GLUESTICK , restores the lost functionality by adding a lightweight correction computed purely in weight space, while largely retaining the memory and speed benefits of pruning.

Figure 3: Overview of GLUESTICK. (Top) A VLA model is pruned with a standard algorithm (e.g., Wanda) to enforce 2:4 sparsity in linear layers. (Middle) Offline, the gap between the dense and pruned weights is computed and dominant lost directions are extracted via SVD, yielding lightweight corrections. (Bottom) At inference, these correction terms are applied alongside the pruned weights, adding back lost signal.

Table 4

Table 5

Model	Device	Method	Latency ( $\downarrow$ )	Speedup ( $\uparrow$ )
OpenVLA	NVIDIA Thor	Full Dense	$796.50\pm 25.85$ ms	—
		Full Sparse	$434.47\pm 3.63$ ms	$1.83\times$
		Sparse Language Backbone	$439.45\pm 13.37$ ms	$1.81\times$
		GLUESTICK-200	$453.58\pm 31.29$ ms	$1.75\times$
		GLUESTICK-500	$537.49\pm 28.19$ ms	$1.48\times$
WorldVLA	NVIDIA Thor	Full Dense	$2816.57\pm 65.73$ ms	—

Table 6: Edge inference latency. End-to-end inference on NVIDIA Jetson Thor and Orin NX. Pruning accelerates inference; GLUESTICK retains speedup with minimal overhead.

\dagger

Since dense exceeds device memory, speedup not compared but still accelerated; sparsity enables deployment.

Figure 4: Memory Savings and Rank Tradeoff. Left: VRAM savings relative to Full Dense. All methods use NaVILA. Right: Rank ablation. All methods use OpenVLA and LIBERO Spatial.

Method	Dense	Dense/Int8	Sparse	Sparse/Int8	Sparse/Int8/GLUESTICK
Success Rate (%)	85.2	84.8	31.2	31.8	62.2

Table 7: GLUESTICK composes with quantization. Evaluating OpenVLA on LIBERO Spatial.

Table 9

Method	2:4	4:8
Magnitude	11.0	25.4
Magnitude + GLUESTICK	48.8	61.8

Table 10: GLUESTICK is pruning-algorithm and sparsity-pattern agnostic. Magnitude-pruned OpenVLA language backbone evaluated on LIBERO Spatial (values show success %).

Figure 5: Real-World Manipulation on Unitree G1. With the pruned model, the G1’s arm collides with its own body and the task fails, while with GLUESTICK the G1 successfully places the apple in the bin ( left ). Pruning results in 90% unsafe episodes and 0% task success, while GLUESTICK fully restores safety and reaches 60% success ( right ).

Figure 6: Real-World Navigation on Boston Dynamics Spot. Our real Spot robot navigating in a real office environment (left) . GLUESTICK reduces pruned NaVILA deviation by 50% (right) .

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Correlation of layer-wise spectra with pruning sensitivity. Pruning the top, middle, and bottom 50 layers ranked by AUC/

k

change pre/post VLA fine-tuning hurts success most, moderately, and least, respectively.

Model	Method	Success (%)	VRAM (GB)	$\Delta$ Success	$\Delta$ VRAM
OpenVLA	Full Dense	85.2	16.12	+0.0	+0.00
	Full Sparse	0.0	10.17	-85.2	-5.95
	Sparse Language Backbone	31.2	10.56	-54.0	-5.56
	75% Sparse Language Backbone	25.4	11.96	-59.8	-4.16
	GLUESTICK-200	49.0	11.18	-36.2	-4.94
	GLUESTICK-500	60.2	11.97	-25.0	-4.15

Appendix

Table 11: LIBERO Spatial.

\Delta

columns are relative to each model’s Full Dense baseline on this benchmark. VRAM measured during inference on same hardware. Corrections stored in Int8.

Model	Method	Success (%)	VRAM (GB)	$\Delta$ Success	$\Delta$ VRAM
OpenVLA	Full Dense	85.8	16.12	+0.0	+0.00
	Full Sparse	13.4	10.17	-72.4	-5.95
	Sparse Language Backbone	50.8	10.56	-35.0	-5.56
	75% Sparse Language Backbone	49.0	11.96	-36.8	-4.16
	GLUESTICK-200	66.7	11.18	-19.1	-4.94
	GLUESTICK-500	71.2	11.97	-14.6	-4.15

Appendix

Table 12: LIBERO Object.

\Delta

columns are relative to each model’s Full Dense baseline on this benchmark. VRAM measured during inference on same hardware. Corrections stored in Int8.

Model	Method	Success (%)	VRAM (GB)	$\Delta$ Success	$\Delta$ VRAM
OpenVLA	Full Dense	77.0	16.12	+0.0	+0.00
	Full Sparse	0.8	10.17	-76.2	-5.95
	Sparse Language Backbone	20.0	10.56	-57.0	-5.56
	75% Sparse Language Backbone	20.8	11.96	-56.2	-4.16
	GLUESTICK-200	30.4	11.18	-46.6	-4.94
	GLUESTICK-500	47.0	11.97	-30.0	-4.15

Appendix

Table 13: LIBERO Goal.

\Delta

columns are relative to each model’s Full Dense baseline on this benchmark. VRAM measured during inference on same hardware. Corrections stored in Int8.

Model	Method	Success (%)	VRAM (GB)	$\Delta$ Success	$\Delta$ VRAM
OpenVLA	Full Dense	55.8	16.12	+0.0	+0.00
	Full Sparse	0.0	10.17	-55.8	-5.95
	Sparse Language Backbone	12.4	10.56	-43.4	-5.56
	75% Sparse Language Backbone	11.4	11.96	-44.4	-4.16
	GLUESTICK-200	16.2	11.18	-39.6	-4.94
	GLUESTICK-500	26.6	11.97	-29.2	-4.15

Appendix

Table 14: LIBERO Long.

\Delta

columns are relative to each model’s Full Dense baseline on this benchmark. VRAM measured during inference on same hardware. Corrections stored in Int8. It is worth noting that WorldVLA completely collapsed after pruning on LIBERO Long, producing invalid action tokens and failing to generate meaningful outputs. Although task success rates did not improve, GLUESTICK was able to restore the model to producing more valid outputs.

Figure 8: Navigation Trajectory Quality. Distribution of path lengths (left) and final distances to goal (right) for pruned NaVILA models and GLUESTICK on VLN-CE-Isaac.

Method	Success ( $\uparrow$ )	$\boldsymbol{\Delta}$ Unsafe ( $\downarrow$ )	PL ( $\downarrow$ )	DG ( $\downarrow$ )	$\boldsymbol{\Delta}$ VRAM ( $\downarrow$ )
Full Dense	$43.0$	$+0.0$	$11.7$	$5.9$	$+0.00$
Full Sparse	$0.0$	$+23.0$	$17.6$	$9.5$	$-5.74$
Sparse Language Backbone	$23.0$	$+2.0$	$14.8$	$8.5$	$-5.68$
% Sparse Language Backbone	$25.0$	$+12.0$	$14.9$	$8.4$	$-4.59$
GLUESTICK-200	$41.0$	$-1.0$	$12.5$	$6.5$	$-5.36$
GLUESTICK-500	$\mathbf{44.0}$	$\mathbf{-4.0}$	$\mathbf{11.9}$	$\mathbf{5.9}$	$\mathbf{-4.60}$

Appendix

Table 15: Navigation results. Success denotes task success rate (%).

\Delta

Unsafe and

\Delta

VRAM are reported relative to Full Dense; lower values are better. All methods use the NaVILA model. PL = Path Length. DG = Final Distance to Goal.

Figure 9: Component Sensitivity : Success-rate drop per million parameters when pruning only the language or vision backbone. Vision pruning is $4.8\times$ (OpenVLA) and $37\times$ (NaVILA) more damaging per parameter.

Figure 10: Calibration Set Visualization. LIBERO Spatial calibration trajectories. Full trajectories (left) . 5% window around the first gripper-closing event, with red segments marking closed-gripper states (right) .

Domain	Avg. Path-Length Differences
Manipulation	0.018 m
Navigation	3.292 m

Appendix

Table 16: Average path-length difference between successful and failed episodes in manipulation and navigation tasks. Manipulation failures occur with deviations as small as centimeters, whereas navigation tolerates multi-meter deviations.

Model	Total Parameters	Linear Parameters	Conv. Parameters
OpenVLA	7,541,237,184	7,407,513,280 (98.23%)	1,281,664 (0.02%)
WorldVLA	7,042,582,528	6,744,440,832 (95.77%)	27,310,848 (0.39%)
NaVILA	8,494,180,416	7,962,922,816 (93.75%)	678,528 (0.01%)

Appendix

Table 17: Parameter composition of recent VLA models. Across OpenVLA, WorldVLA, and NaVILA, the vast majority of parameters ( 93–98% ) are contained in linear layers, while convolutional components represent well under 1% of total model size.

Rank	200	400	500	800	1000	2000	2500
Success Rate (%)	0	0	0	0	0	0.4	21.0

Appendix

Table 18: Effect of replacing OpenVLA weights with an approximation using SVD (no pruning).

Method	Total FLOPs (T)
Full Dense	29.32
50% 2:4 Pruned OpenVLA	16.54
50% 2:4 Pruned OpenVLA + GLUESTICK-200	18.51
50% 2:4 Pruned OpenVLA + GLUESTICK-500	21.47

Appendix

Table 19: FLOP analysis for a single inference pass through OpenVLA. FLOP counters for the following operations: aten.convolution , aten.add , aten.addmm , aten._scaled_dot_product_flash_attention , aten.mm , and aten.bmm . In these experiments the language backbone of OpenVLA was pruned with Wanda.

Explore similar work

Recovering Aggressively Pruned Vision-Language-Action Models with Offline Hidden-State Distillation
Sep 17, 2026Chiyoung Kim, Sanghyuk Roy Choi, Minhyeok LeeEfficient VLA Models Vision-Language-Action Models
Chung-Ang University, Seoul, Republic of Korea
SAFE-Pruner: Semantic Attention-Guided Future-Aware Token Pruning for Efficient Vision-Language-Action Manipulation
May 28, 2026Shilin Ma, Chubin Zhang, Changyuan Wang +6Efficient VLM Inference Robotic Manipulation
Tsinghua Shenzhen International Graduate School, Tsinghua University · GigaAI
Revisiting Parameter Redundancy in Vision-Language-Action Models: Insights from VLM-to-VLA Adaptation
Jun 30, 2026Fengnian Zhang, Tao Huang, Siyu Xu +2VLM Adaptation Vision-Language-Action Models
Computer Network Information Center, Chinese Academy of Sciences · University of Chinese Academy of Sciences · School of Computer Science, Shanghai Jiao Tong University +1

Sep 17, 2026J/K move · Enter open · S save

Recovering Aggressively Pruned Vision-Language-Action Models with Offline Hidden-State Distillation

Chiyoung Kim, Sanghyuk Roy Choi, Minhyeok Lee

Chung-Ang University, Seoul, Republic of Korea

Efficient VLA Models Vision-Language-Action Models VLMs for Robotics Knowledge Distillation

Open paper AskPDF arXiv

Vision-language-action (VLA) models let robots follow language instructions, but their language backbones of several billion parameters are the main obstacle to running them on robot hardware. Structured pruning reduces that backbone, and removing 63% of it from OpenVLA-OFT drops LIBERO-Long success from 93.2% to 0.8%. A recent approach restores such a model with supervised fine-tuning followed by reinforcement learning, which needs online rollouts and hundreds of GPU-hours. We recover most of the lost success entirely offline. Width pruning narrows the blocks but keeps the residual stream at its original size, so teacher and student hidden states have the same shape and are matched directly, without a projector. Training against a cache built in one teacher pass lifts the 63%-reduced student to within 3.5 points of the teacher in about 8 GPU-hours. A sweep over nine ratios locates where the recovery objective starts to matter. Up to 45% reduction the two do not differ significantly on OpenVLA-OFT. Hidden-state distillation then adds +2.1 to +4.5 points there between 63% and 87%, and +9.4 to +22.1 points on CogACT from 63% onward. At 81% on CogACT, a tripled recovery budget narrows the distilled student's gap to the teacher to 3.9 points on average, while supervised recovery stays more than 20 points below. At matched compression, width pruning yields higher success and depth pruning lower latency. On a 6-DoF manipulator, the distilled student at 72% reduction reaches 77.5% success against 59.5% for supervised recovery, runs 2.23x faster on-board than the teacher, and uses 62% less memory.

Abstract

Explore similar work

Recovering Aggressively Pruned Vision-Language-Action Models with Offline Hidden-State Distillation

SAFE-Pruner: Semantic Attention-Guided Future-Aware Token Pruning for Efficient Vision-Language-Action Manipulation

Revisiting Parameter Redundancy in Vision-Language-Action Models: Insights from VLM-to-VLA Adaptation