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

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 degradation and increased safety violations

Pruning Collapse in VLA Models

The study presents empirical evidence of pruning collapse in VLAs, showing that pruning, which is effective for LLMs, causes near-complete collapse of the success rate in embodied VLA models and an increase in safety violations. The authors observe that recent pruning algorithms reduced task success rate on manipulation tasks from 85.2% to 0.0% and on a navigation task from 43.0% to 0.0%, while also increasing the frequency of safety violations. This finding demonstrates that text-validated pruning does not directly transfer to embodied control. Furthermore, spectral analysis identifies structural properties of VLA architectures that could make them more fragile to pruning than language-only models because VLA layers show a noticeably flatter decay in singular value spectra, indicating energy spread across many directions.

GLUESTICK Recovery Method

GLUESTICK is introduced as a new post-pruning recovery method that recovers signal lost during pruning while preserving the efficiency benefits of sparsity. This method operates entirely in weight space, using the information discarded by pruning to nudge the model back toward more performant regions without any retraining. The correction is achieved by adding a lightweight correction term computed from singular values in the gap between the dense and pruned weights. This approach is pruning-agnostic and requires no additional training, introducing only a single interpretable hyperparameter that controls the tradeoff between efficiency and accuracy.

GLUESTICK Mechanism

For each linear layer with dense weight matrix Wdense ∈ R dout×din and its pruned version Wpruned (fixed, preserving the original 2:4/4:8 pattern), the gap matrix Wgap = Wdense − Wpruned captures lost information due to pruning. A truncated singular value decomposition (SVD) of the gap matrix is computed, yielding a low-rank approximation A = UrΣr ∈ R dout×r and B = Vr ∈ R din×r. During inference, GLUESTICK adds a lightweight correction around each pruned layer: h(x) = Wprunedx + A(B>x), which re-injects the dominant lost directions at low cost while leaving Wpruned unchanged. This correction adds only (din + dout)r extra parameters per layer, which is small compared to dindout in the dense case.

Experimental Results

GLUESTICK consistently recovers performance and improves safety across three VLA architectures, two widely used robotics benchmarks, and multiple robot embodiments. For LIBERO benchmarks, GLUESTICK–500 degrades by only −35.7% in manipulation tasks compared to Full Dense, recovering 50% of the success rate lost to pruning. On the VLN-CE-Isaac benchmark for navigation, GLUESTICK–500 fully restores the dense model’s performance, recovering 100% of the lost success rate and maintaining path length and final distance-to-goal nearly identical to the dense baseline. GLUESTICK–500 maintains a safety profile near parity with the dense policy, yielding 89% and 100% fewer unsafe episodes compared to the Full Sparse model for manipulation and navigation, respectively.

Analysis of Sensitivity

The analysis shows that vision backbones are disproportionately sensitive to pruning while offering little memory benefit, since they comprise less than 10% of total parameters. Pruning the vision backbone causes outsized harm relative to their limited contribution to overall model size because pruning vision components causes outsized harm relative to their limited contribution to overall model size. The choice of singular values is highly influential, with selecting the largest components by magnitude playing a central role in effective recovery. GLUESTICK exposes a single hyperparameter r that controls the trade off between memory efficiency and task recovery, where increasing the rank r improves success-rate recovery but at the cost of additional memory.

Conclusion

The paper presents the first systematic study of pruning VLA models and shows that pruning catastrophically degrades both task success and safety. GLUESTICK is a training-free, pruning-agnostic, and easily integrable post-pruning recovery method that reintroduces lost directions due to strict 2:4 sparsity, restoring performance and safety while retaining efficiency. The work lays the foundation for developing compression techniques that make powerful VLA models practical for real-world deployment on resource-constrained robotic platforms.

Appendix Details

The GLUESTICK Details section provides the implementation details, including Algorithm 1 (Offline) and Algorithm 2 (Online) for applying GLUESTICK to pruned layers. The experimental setting specifies that all experiments use an NVIDIA L40S GPU with 48 GB of VRAM. Safety definitions for manipulation include joint limit violations, arm collisions, object velocities, end-effector containment, and whole-body containment. The results are averaged across OpenVLA and WorldVLA models on the LIBERO benchmarks.

References

The paper cites numerous works related to VLA models, pruning methods like Magnitude and Wanda, and specific VLA architectures such as OpenVLA, NaVILA, WorldVLA, and RT. It also references foundational papers on model compression and low-rank approximation techniques like Eckart & Young (1936). The experimental setup involves benchmarks such as LIBERO, VLN-CE-Isaac, and Habitat for testing manipulation and navigation capabilities. The paper also cites works on safety in robotics and LLM fine-tuning alignment. The reference list includes sources for GPT-4, LLaMA-2, DINOv2, and various robotic control models. The authors also cite works on sparse deep neural networks and efficient inference engines like Eie. It also cites works related to pruning LLMs by Zhu et al. and Sun et al. The references cover a broad range of topics from foundational vision-language models to specific robotics benchmarks and safety concerns. The paper also cites works on model compression for large language models by Xunyu Zhu et al. The references include papers on safe lora and adversarial attacks in LLMs, such as Antidote. The paper also cites works on physical intelligence and open-world generalization for VLA models. The references cover the history of neural network approximation techniques from Eckart and Young. It also includes citations for specific robot platforms like Franka Panda, Unitree Go2, and H1 humanoid. The references provide context on the scaling laws of neural language models by Kaplan et al. The paper also cites works related to multimodal models like Gemini. It also includes citations for benchmarks and datasets used in robotics research, such as Habitat and LIBERO. The references provide context on the challenges of real-world reinforcement learning by Dulac-Arnold et al. The references cover the history of vision-language action models transfer web knowledge to robotic control by Zitkovich et al.

1. Bold header: Training-Free Post-Pruning Recovery for VLA Models

This improvement introduces GLUESTICK, a post-pruning recovery method that restores much of the original model’s functionality while retaining sparsity benefits. It performs a one-time interpolation between the dense and pruned models in weight-space to compute a corrective term, which is used during inference to recover lost capabilities with minimal overhead.

2. Bold header: Enhanced Manipulation Task Performance

By applying GLUESTICK, the system can achieve substantial performance gains on manipulation tasks, as evidenced by GLUESTICK-500 recovering 62% and 57% of lost performance on the Spatial and Goal suites, thereby substantially restoring manipulation performance while retaining pruning efficiency.

3. Bold header: Complete Restoration of Navigation Fidelity

GLUESTICK restores navigational capabilities by ensuring that the robot's behavior is not just successful but also efficient. It ensures that path length and final distance-to-goal remain nearly identical to the dense baseline, effectively closing the gap where pruned agents frequently veer into entirely different rooms.

4. Bold header: Safety Preservation During Compression

The system maintains a strong safety profile even after pruning, as GLUESTICK restores dominant weight-space directions that carry both task-relevant and safety-critical signal, resulting in 89% and 100% fewer unsafe episodes compared to the Full Sparse model.

5. Bold header: Adaptive Efficiency Trade-off Control

The introduction of a single hyperparameter allows practitioners to directly control the operational trade-off. This hyperparameter governs the tradeoff between efficiency and accuracy, enabling users to select a specific recovery level, such as GLUESTICK-200 or GLUESTICK-500, based on required memory versus recovery needs.

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/