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

summary

The gist

The gist The post-pruning recovery method GLUESTICK restores much of the original model’s functionality while retaining sparsity benefits, showing that pruning VLA models leads to drastic

In short

Pruning Vision-Language-Action (VLA) models drastically degrades performance and increases safety violations, unlike in language models. The GLUESTICK method is a training-free recovery technique that restores most lost capability by computing a one-time correction term based on weight space interpolation, preserving sparsity benefits.

Key concepts

Pruning Collapse in VLA Models: Pruning VLA models causes near-complete failure on embodied tasks, such as manipulation and navigation. This happens because pruning methods effective for language models do not directly translate to the complex control requirements of robotics. Spectral analysis shows VLA layers are structurally fragile to this type of compression.
GLUESTICK Recovery Method: GLUESTICK is a post-pruning recovery technique that fixes performance loss without retraining. It works by calculating a lightweight correction term derived from the gap between dense and pruned weights. This method is pruning-agnostic, meaning it works regardless of how the model was initially pruned.
Weight Space Interpolation: The core mechanism involves interpolating between the original dense model and its sparse version in weight space. This interpolation allows GLUESTICK to compute a corrective term that re-injecting dominant lost directions into the pruned layers during inference, effectively restoring lost signal at minimal computational cost.
Sparsity Benefits: The method aims to balance efficiency with accuracy. By using GLUESTICK, the model retains the memory and speed advantages gained from pruning while recovering most of its original performance. This is achieved by only adding a small number of extra parameters per layer during inference.

Terminology used across episodes

Reinforcement Learning (RL) — explained in 25 episodes
Sample Complexity — explained in 18 episodes
Federated Learning — explained in 17 episodes
Retrieval-Augmented Generation (RAG) — explained in 16 episodes
Interpretability — explained in 15 episodes
Quantum Fisher Information (QFI) — explained in 15 episodes

This episode discusses

The paper

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

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

FieldAI University of Harvard University

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/

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: Today's paper: "Don't Run with Scissors".

Dev: The gist The post-pruning recovery method GLUESTICK restores much of the original model’s functionality while retaining sparsity benefits,

Rosa: First, who's behind it and why it matters.

Paper summary: Rosa: So, to summarize what's happening with "Don't Run with Scissors: Pruning Breaks VLA Models but They Can Be Recovered," the core thesis is that pruning VLA models causes drastic performance degradation and increases safety violations >

Dev: The paper shows that for vision-language-action models, pruning isn't as straightforward as it is for language only models because the structural properties of these architectures are different >

Taro: They look at three components in these systems: a vision backbone, a multimodal projector, and a language backbone >

Rosa: And they use benchmarks like LIBERO and VLN-CE-Isaac to show that pruning can reduce task success rates to zero point zero percent on some manipulation tasks >

Dev: The authors highlight that the scaling trend where VLA capabilities grow with model size is especially problematic when we are constrained by hardware limits, like the eight to sixteen gigabytes of shared memory on an industry standard Jetson Orin NX >

Taro: They compare results across different pruning strategies, showing that methods like Magnitude and Wanda result in zero percent success in some cases while others perform better >

Rosa: The main contribution is GLUESTICK, which is a post-pruning recovery method that works by computing a one-time interpolation between the dense and pruned models >

Dev: This correction term is then used during inference by each pruned layer to re-inject lost capabilities with minimal overhead, and it requires no further training >

Taro: It’s pruning agnostic, meaning it doesn't matter which pruning algorithm you used before, as long as you have the dense and pruned versions of the weights >

Rosa: The results show that GLUESTICK recovers up to one hundred percent of the lost success rate and restores safety to near-dense levels while saving memory and improving speed >

Dev: So, in short, they prove that we can mitigate the catastrophic failure caused by pruning VLA models with a smart recovery technique >

Conclusion: Rosa: So looking at "Don't Run with Scissors: Pruning Breaks VLA Models but They Can Be Recovered," the authors are essentially showing that pruning is a dangerous shortcut when applied to these complex robotic systems >

Dev: They address the fact that while pruning was effective for smaller language models, it doesn't translate directly to embodied control because of how VLA models are structured >

Taro: The implication for autonomy is that we can still get the efficiency gains from compression techniques on VLA models if we use a method like GLUESTICK to fix the resulting collapse >

Rosa: It suggests that future work should focus on developing more robust pruning strategies specifically tailored to the structural properties of VLA weights, rather than just applying general LLM pruning methods >

Dev: And they leave us with this recovery method, GLUESTICK, as a way to make these powerful models practical for real-world deployment on resource-constrained robotic platforms by fixing the safety and performance issues introduced by pruning >

More episodes

← Home