Abstract
Recent Vision-Language-Action (VLA) models increasingly adopt chain-of-thought (CoT) reasoning, generating a natural-language plan before decoding motor commands. This internal text channel between the reasoning module and the action decoder has received no adversarial scrutiny. We ask: which properties of this intermediate plan does the action decoder actually rely on, and can targeted corruption of the reasoning trace alone -- with all inputs left intact -- degrade a robot's physical task performance? We design a taxonomy of seven text corruptions organized into three attacker tiers (blind noise, mechanical-semantic, and LLM-adaptive) and apply them to a state-of-the-art reasoning VLA across 40 LIBERO tabletop manipulation tasks. Our results reveal a striking asymmetry: substituting object names in the reasoning trace reduces overall success rate by 8.3~percentage points (pp) -- reaching -19.3~pp on goal-conditioned tasks and -45~pp on individual tasks -- whereas sentence reordering, spatial-direction reversal, token noise, and even a 70B-parameter LLM crafting plausible-but-wrong plans all have negligible impact (within \pm4~pp). This asymmetry indicates that the action decoder depends on entity-reference integrity rather than reasoning quality or sequential structure. Notably, a sophisticated LLM-based attacker underperforms simple mechanical object-name substitution, because preserving plausibility inadvertently retains the entity-grounding structure the decoder needs. A cross-architecture control using a non-reasoning VLA confirms the vulnerability is exclusive to reasoning-augmented models, while instruction-level attacks degrade both architectures -- establishing that the internal reasoning trace is a distinct and stealthy threat vector invisible to input-validation defenses.
