论文研究如何恢复结构化剪枝后视觉语言动作模型的操作能力。方法缓存一次教师模型前向得到的隐状态，用离线隐状态蒸馏训练剪枝学生模型，无需在线采样轨迹；在 OpenVLA-OFT 上剪掉 63% 参数后，约 8 GPU 小时即可把成功率恢复到距教师 3.5 个百分点以内。

Vision-language-action (VLA) models let robots follow language instructions, but their language backbones of several billion parameters are the main obstacle to running them on robot hardware. Structured pruning reduces that backbone, and removing 63% of it from OpenVLA-OFT drops LIBERO-Long success from 93.2% to 0.8%. A recent approach restores such a model with supervised fine-tuning followed by reinforcement learning, which needs online rollouts and hundreds of GPU-hours. We recover most of the lost success entirely offline. Width pruning narrows the blocks but keeps the residual stream at its original size, so teacher and student hidden states have the same shape and are matched directly, without a projector. Training against a cache built in one teacher pass lifts the 63%-reduced student to within 3.5 points of the teacher in about 8 GPU-hours. A sweep over nine ratios locates where the recovery objective starts to matter. Up to 45% reduction the two do not differ significantly on OpenVLA-OFT. Hidden-state distillation then adds +2.1 to +4.5 points there between 63% and 87%, and +9.4 to +22.1 points on CogACT from 63% onward. At 81% on CogACT, a tripled recovery budget narrows the distilled student’s gap to the teacher to 3.9 points on average, while supervised recovery stays more than 20 points below. At matched compression, width pruning yields higher success and depth pruning lower latency. On a 6-DoF manipulator, the distilled student at 72% reduction reaches 77.5% success against 59.5% for supervised recovery, runs 2.23× faster on-board than the teacher, and uses 62% less memory.

VLA 的数十亿参数语言骨干难以在机器人硬件运行；大幅剪枝会严重损害任务成功率，而已有恢复方法需要在线 rollout 和大量 GPU 时间。

On a VLA, however, aggressive pruning does not simply degrade the policy.

利用宽度剪枝保留教师与学生隐状态的相同维度，直接做离线隐状态蒸馏，无需投影层、在线 rollout 或奖励信号。

METHOD At the target ratios, recovery needs to bring back most of the policy, and the dense teacher is the model that already performs the task successfully.

先对教师模型执行一次前向并缓存隐状态，再以缓存监督被剪枝学生模型；系统比较九种剪枝比例、两种骨干以及宽度剪枝和深度剪枝的恢复效果。

METHOD At the target ratios, recovery needs to bring back most of the policy, and the dense teacher is the model that already performs the task successfully.

在 OpenVLA-OFT 剪枝 63% 时，离线蒸馏约用 8 GPU 小时将学生恢复到距教师 3.5 个百分点内；在真实 6 自由度机械臂上，72% 剪枝的蒸馏模型成功率为 77.5%，监督恢复为 59.5%，且板载运行快 2.23 倍、内存少 62%。

Width pruning leaves teacher and student hidden states directly comparable, which turns recovery into offline supervision from a cached teacher, with no rollouts and no reward. • When hidden-state distillation is beneficial.A sweep over nine compression ratios and two backbones locates the reduction at which hidden-state distillation begins to outperform supervised recovery, and shows that a larger recovery budget moves that point further out. • Width versus depth at matched compression.With both choices under one recovery protocol, width pruning yields higher success than a CKA-guided depth baseline at every measured point on both backbones, and depth pruning yields lower latency. arXiv:2609.19579v1 [cs.RO] 17 Sep 2026