REASONING AS A DOUBLE-EDGED SWORD: ARCHITECTURE AND CROSS-STAGE ROBUSTNESS IN VISION-LANGUAGE-ACTION MODELS

This study investigates whether incorporating an explicit reasoning step—such as a text-based chain-of-thought (CoT) or a latent iterative loop—improves the robustness of Vision-Language-Action (VLA) models against adversarial perturbations. While the intuition suggests that reasoning before acting should help a model absorb noise, the authors find that the reasoning architecture itself can become a significant point of failure.

The researchers evaluated three distinct VLA architectures across the reasoning spectrum: a single-pass model (OpenVLA-OFT), a text-CoT model (DeepThinkVLA), and a latent-iterative model (RD-VLA). They subjected these models to a cross-stage attack matrix, injecting stochastic noise and white-box perturbations (PGD-10) at the vision, reasoning, and action stages. They also tested whether reasoning outputs could be monitored as a safety signal using plan-action consistency probes.

The latent-iterative model (RD-VLA) proved to be the least robust, with task success rates collapsing under both stochastic and white-box attacks. Crucially, the authors discovered that this fragility is structural rather than cumulative; varying the reasoning depth (the number of iterations) at inference time did not significantly change the amplification of perturbations. Furthermore, while monitoring the reasoning process seems effective under naive evaluation, these monitors fail completely when subjected to adaptive attacks, providing no meaningful defense against adversarial inputs.

As VLA models are increasingly deployed in robotics, understanding the trade-offs between reasoning capabilities and system robustness is critical. This work demonstrates that adding reasoning layers can inadvertently expand the attack surface and create structural vulnerabilities. It highlights that current behavioral monitoring techniques are insufficient for safety-critical applications, establishing a necessary performance floor that future defenses must overcome.