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

2025-10-09

AI Summary

Pruning vision-language-action (VLA) models for deployment on resource-constrained devices often incurs severe performance degradation and a sharp increase in safety violations, while existing mitigation strategies rely heavily on costly retraining. To address this, we propose GLUESTICK—a zero-shot, training-free recovery method that reconstructs functionality lost during pruning. GLUESTICK performs a one-time interpolation between dense and pruned model weights in parameter space, generating layer-wise correction terms applied during inference to precisely compensate for functional deficits—without altering the sparse architecture. The approach is broadly applicable across pruning strategies and requires tuning only a single hyperparameter. Extensive experiments across diverse VLA architectures and robotic manipulation/navigation tasks demonstrate that GLUESTICK significantly improves task success rates, reduces safety violations, and preserves memory efficiency—achieving a robust balance between model sparsity and behavioral fidelity.

Technology Category

• Natural Language Processing: Safety and Robustness
• Computer Vision: Large Vision Models
• Intelligent Robots: Manipulation

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Security and Privacy: Security and privacy of machine learning and AI applications

Abstract

Vision-Language-Action (VLA) models have advanced robotic capabilities but remain challenging to deploy on resource-limited hardware. Pruning has enabled efficient compression of large language models (LLMs), yet it is largely understudied in robotics. Surprisingly, we observe that pruning VLA models leads to drastic degradation and increased safety violations. We introduce GLUESTICK, a post-pruning recovery method that restores much of the original model's functionality while retaining sparsity benefits. Our method performs a one-time interpolation between the dense and pruned models in weight-space to compute a corrective term. This correction is used during inference by each pruned layer to recover lost capabilities with minimal overhead. GLUESTICK requires no additional training, is agnostic to the pruning algorithm, and introduces a single hyperparameter that controls the tradeoff between efficiency and accuracy. Across diverse VLA architectures and tasks in manipulation and navigation, GLUESTICK achieves competitive memory efficiency while substantially recovering success rates and reducing safety violations. Additional material can be found at: https://gluestick-vla.github.io/.

Problem

• Pruning causes severe performance degradation in VLA models
• VLA models suffer increased safety violations after pruning
• GLUESTICK recovers pruned VLA functionality while maintaining sparsity

Innovation

• GLUESTICK recovers pruned VLA models via weight interpolation
• Method requires no training and is pruning-algorithm agnostic
• Uses corrective term during inference to restore functionality