AutoPruner: Agentic Depth Pruning for Vision-Language-Action Models

Tao Lin, Yikeng Chen, Qiwei Liang, Zinan Xie, Zhengtu Liang, Renjing Xu · N/A · 2026

Summary

Abstract Summary

Layer pruning reduces the size and inference cost of vision-language-action (VLA) models, but selecting a deployable configuration requires deciding both which layers to remove and how to recover task performance. These decisions depend on the model and task, and a failed trial can reflect either the deletion set or insufficient recovery. We present AutoPruner, an agent-driven framework that jointly searches pruning structures and recovery recipes through closed-loop experimentation. Offline layer interventions measure changes in action output to provide functional evidence. An LLM agent combines this evidence with evaluation history and retained diagnoses to propose deletion sets and recovery recipes within a search budget. Each proposal is executed and evaluated on the target tasks, and the resulting feedback guides the next experiment. Across OpenVLA-OFT, StarVLA, and FastWAM on four LIBERO suites, AutoPruner removes 16.7--87.5% of backbone layers while maintaining task success above 90%, with measured inference speedups of up to 2.50$\times$. Further analysis shows that functional probes identify action-sensitive layers despite high representation similarity, and that recovery can make previously failing structures viable.

Key Points

Layer pruning reduces the size and inference cost of vision-language-action (VLA) models, but selecting a deployable configuration requires deciding both which layers to remove and how to recover task performance.
These decisions depend on the model and task, and a failed trial can reflect either the deletion set or insufficient recovery.
We present AutoPruner, an agent-driven framework that jointly searches pruning structures and recovery recipes through closed-loop experimentation.
Offline layer interventions measure changes in action output to provide functional evidence
An LLM agent combines this evidence with evaluation history and retained diagnoses to propose deletion sets and recovery recipes within a search budget.

Method

As shown in Figure 2 , AutoPruner closes the pruning and recovery loop with an LLM agent. The LLM agent acts as the proposal controller, selecting the next deletion set and recovery recipe from the accumulated evidence and recovery library. Tools prune, recover, and evaluate each policy. Feedback guides subsequent proposals toward deeper pruning under a success threshold and a round budget.

For a VLA policy $\pi$ with $L$ backbone layers, a deletion set $D\subseteq\{0,\ldots,L-1\}$ produces the pruned policy $\pi_{D}$.

Method — AutoPruner: Agentic Depth Pruning for Vision-Language-Action Models — Fig. 2: AutoPruner pipeline. Given a target success rate and search budget, an LLM agent uses layer probes and trial history to select pruning masks and recovery strategies. Closed-loop evaluations guide the search, and a selected configuration meeting the target is exported as a physically pruned checkpoint.

Experiments

We evaluate pruning and recovery on OpenVLA-OFT [ 9 ] , StarVLA [ 3 ] , and FastWAM [ 10 ] across the Object, Spatial, Goal, and Long suites of LIBERO [ 35 ] . Each model–suite pair is a deployment with its dense policy as the reference. Recorded OpenVLA-OFT agent runs span 4–6 proposal rounds, with GPT-5.5 using evaluation feedback to revise pruning and recovery configurations.

The Object case study evaluates four proposals over 200 episodes each (800 total).

Experiments — AutoPruner: Agentic Depth Pruning for Vision-Language-Action Models — Fig. 3: OpenVLA-OFT recovery recipes. The three levels evaluate the pruned policy directly ($r_{0}$), adapt it to demonstrations ($r_{1}$), or add supervision from the dense teacher ($r_{2}$).

Analysis

Figure 4 (a) shows how high representation similarity can conceal action-sensitive layers. StarVLA L6 has adjacent-layer CKA of 0.9986 yet produces a large action deviation when bypassed. CKA ranks it first for deletion, which reduces task success below 90%. FastWAM L0 likewise remains action-sensitive at CKA 0.9995.

By measuring action changes directly, functional probes identify sensitive layers that appear redundant under CKA.

Analysis — AutoPruner: Agentic Depth Pruning for Vision-Language-Action Models — Fig. 4: Functional sensitivity and layer interactions. (a) Bypassing StarVLA L6 causes large action deviation despite high adjacent-layer CKA [ 38 ] . (b) Percentage of evaluated layer pairs with joint action deviation below the sum of individual deviations. (c) Partial Spearman correlations with action sensitivity. Filled markers show readout deviation controlling for initial deviation; hollow markers reverse these roles. CKA uses calibration inputs; pairwise measurements use 120, 128, and 60 examples for OpenVLA-OFT, StarVLA, and FastWAM, respectively.

Sources and Demonstrations

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