Subliminal-Style Action Traits in Vision-Language-Action Robot Policies

Subliminal learning is the surprising phenomenon in which a student model acquires a teacher's hidden trait after training on teacher-generated data whose surface content is unrelated to that trait \citep{cloud2025subliminal}. Existing evidence is strongest in language-model settings, where both inputs and outputs are text. We ask whether an analogous effect can appear in robot foundation models, where observations are visual, proprioceptive, and linguistic, while outputs are continuous action trajectories. We study popular open vision-language-action (VLA) policies, including SmolVLA and X-VLA, and test whether action-level traits can pass from a trait-finetuned teacher to a fresh student through neutral teacher-generated trajectories. We construct two offline action traits, a low-motion joint-style code and a trajectory smoothness signature, by augmenting real robot demonstration actions while leaving observations and task prompts unchanged. Across native model-task pairs, we find several positive transfer cases, especially for joint-style traits, including X-VLA on Google Robot and Soft-FOLD tasks and SmolVLA on Stanford Kuka and LIBERO. Smoothness transfer is less stable, with a positive Soft-FOLD case but repeated negative results on LIBERO and Google Robot. These results suggest that subliminal-style trait transfer is a useful probe for VLA policies and a promising direction for constructive style transfer, bias auditing, and alignment-aware robot policy distillation.