Finetuning Vision-Language-Action Models Requires Fewer Layers Than You Think

Gia-Binh Nguyen, Trong-Bao Ho, Thien-Loc Ha, Khoa Vo, Philip Lund Møller, Quang T. Nguyen, Long Dinh, Tuan Dam, Vu Duong, Tung M. Luu, Trung Le, Tran Nguyen Le, Minh Vu, An Thai Le, Ngan Le, Daniel Sonntag, James Zou, Jan Peters, Duy M. H. Nguyen, Ngo Anh Vien · · 2026

Summary

Abstract Summary

Vision-Language-Action (VLA) models pre-trained on massive video-robot datasets have revolutionized robotic manipulation, yet their multi-billion parameter architectures impose prohibitive computational burdens during downstream fine-tuning and real-time inference. In this work, we reveal a highly non-trivial architectural characteristic of these continuous control foundation policies (e.g., pi_0, GR00T-N1.5): despite being trained on diverse physical trajectories, they exhibit severe layer-wise representational redundancy. To exploit this, we introduce a structural compression pipeline that is entirely training-free, bypassing the need of existing methods to load full-scale models to learn optimized token reductions or dynamic layer selectors. Instead, using only a single forward pass via Centered Kernel Alignment to identify redundant layer features, we remove twin layers to permanently compress the model depth by up to 50% across both the VLM backbone and the continuous control policy head. Downstream fine-tuning of this streamlined architecture yields a dual acceleration benefit: a 40-50% reduction in training time and up to 30% faster real-time inference, while matching or exceeding full-scale base model performance. We comprehensively validate our method across three simulation benchmarks (LIBERO, RoboCasa, SimplerEnv) and 10 diverse real-world manipulation tasks across 4 unique robotic embodiments. These results prove that advanced VLAs require significantly fewer layers than previously assumed, offering a highly compute-efficient paradigm for scalable robot learning.

Key Points

Proposes Finetuning Vision-Language-Action Models Requires Fewer Laye for robotics.
Evaluated on real or simulated robotic tasks.
Primary contribution in ar-vr.

Preliminaries

Vision-Language-Action Model. Let $\mathbf{x}^{\text{lang}}$ denote a text instruction and $\mathbf{x}^{\text{img}}\in\mathbb{R}^{H\times W\times 3}$ denote the RGB observation. A continuous-action VLA policy predicts an action chunk $\mathbf{a}\in\mathbb{R}^{T_{a}\times d_{a}}$ conditioned on these multimodal inputs.

CKA-Guided Layer Pruning

Diagnosing Layer Redundancy in Deep VLAs. To better understand the internal mechanics driving deep VLA architectures, we initiate an empirical exploration into a central question: “how does information actually evolve across network depth?” .

Leveraging the CKA metric established in Section 1 , we trace the hidden state trajectories of representative continuous-control foundations as they map multimodal context to physical actions.

CKA-Guided Layer Pruning — Finetuning Vision-Language-Action Models Requires Fewer Layers Than You Think — Figure 2: CKA similarity profiles across $\pi_{0}$ and GR00T-N1.5 sub-modules. The heatmaps illustrate pairwise representation alignment among transformer layers inside the VLM backbones, action heads, and DiT blocks. The extensive, contiguous plateaus of high similarity (dark red) across both model families signify minimal representational changes between successive layers, pinpointing candidate zones for structured pruning.

Experimental Results and Analysis

We conduct a diverse set of experiments to examine the performance of pruned VLA models trained with our CLP models across simulation and real-world settings and robot embodiment configurations. Details are presented in the Appendix. We first investigate the fundamental architectural properties of layer redundancy in VLAs and dissect the internal mechanisms of our framework ( RQ1 & RQ2 ).

Limitation

A key limitation of CLP is its use of a global pruning criterion that does not explicitly account for modality-specific token dynamics in manipulation tasks. Our analysis suggests that action and state tokens exhibit distinct representations, motivating future token- or modality-aware pruning strategies.

In addition, the current work investigates CLP exclusively in the context of post-pretraining fine-tuning, leaving its application to the pretraining stage an open direction for future work.

Limitation — Finetuning Vision-Language-Action Models Requires Fewer Layers Than You Think — Figure 4: Robot experiments with folding shorts.

Sources and Demonstrations

Figures are reproduced from Nguyen et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.