Drop-Then-Recovery: How Redundant Are Vision-Language-Action Models?

Guoheng Sun, Kaixi Feng, Shwai He, Xiaochuan Gong, Yexiao He, Ziyao Wang, Zheyu Shen, Wanghao Ye, Ramana Rao Kompella, Gaowen Liu, Ang Li

arXiv:2606.27755 · 2026-07-01 공개 · arXiv · PDF

fine-tuning vision-language-action libero robotic-manipulation openvla capacity-allocation transformer-block-removal gateprobe

Abstract

Vision-Language-Action (VLA) models enable instruction-driven robotic manipulation, but they inherit oversized language backbones from pretrained VLMs whose capacity far exceeds what is needed for short robotic instructions. This raises a basic question: how much of a VLA model is actually necessary for closed-loop control? In this work, we study architectural redundancy in VLA models by using transformer block removal as a controlled intervention. We introduce Drop-Then-Recovery (DTR), an analysis protocol that removes selected blocks from a pretrained VLA model and then fine-tunes the resulting model to measure whether the removed capacity was necessary for downstream control. To make this intervention reliable, we propose GateProbe, a one-shot virtual-gate sensitivity metric that ranks blocks by their contribution to the downstream action loss. Across multiple VLA architectures, manipulation benchmarks and even real-robot industrial scenarios, we find a strong asymmetry in post-removal recoverability: \textit{language backbones are highly redundant for standard robotic manipulation tasks, whereas vision and action pathways are substantially less tolerant to removal}. On LIBERO, removing half of the LLM blocks even improves OpenVLA-OFT from 95.0% to 98.3% under the same downstream fine-tuning budget, and retaining only two language blocks still recovers baseline-level performance. These results suggest that current VLA benchmarks may exert limited pressure on deep language grounding and compositional instruction understanding, and that future VLA architectures should allocate capacity more deliberately across language, vision, and action components. The code is available at https://github.com/s1ghhh/VLADrop.

한국어 요약

한 줄 요약

VLA 모델에서 언어 백본이 과도하게 큰 반면, 시각 및 행동 경로는 필수적이라는 사실을 DTR과 GateProbe를 통해 실증적으로 분석했다.

핵심 기여도

핵심 아이디어

VLA 모델은 시각-언어-행동 통합 학습을 통해 로봇 조작을 수행하지만, 대규모 언어 백본이 과도하게 큰 문제가 있다. 연구팀은 **DTR**이라는 프로토콜을 통해 트랜스포머 블록을 제거한 후 미세조정(fine-tuning)을 통해 제거된 블록이 실제로 필요한지 평가하는 실험적 접근법을 제안했다. 이는 단순히 파라미터 수가 아닌, **실제 닫힌 루프 제어 성능**을 기준으로 중복성을 측정하는 핵심 차이점이다. 또한, **GateProbe** 알고리즘은 블록 제거 후 행동 손실(action loss)에 미치는 영향을 기준으로 블록의 중요도를 정렬하여, 효과적인 제거 집합을 선택한다.

기술적 접근법

주요 결과

의의 및 한계

실용적 활용