Despite recent progress in Vision Language Action (VLA) models, but their transformer back-
bones largely remain black boxes, limiting safety, debuggability, and resource-aware deployment.
This thesis presents an action-centric mechanistic analysis of...
Despite recent progress in Vision Language Action (VLA) models, but their transformer back-
bones largely remain black boxes, limiting safety, debuggability, and resource-aware deployment.
This thesis presents an action-centric mechanistic analysis of a modern VLA policy, ω0-FAST,
trained on LIBERO tabletop manipulation tasks, with the goal of uncovering the internal circuit
that decodes multimodal context into continuous robot actions.
We first study feedforward value vectors to locate where in depth the model allocates ca-
pacity to promote action tokens versus preserve the pretrained semantic backbone. Next, we
group tokens by modality and role (images, text instructions, proprioceptive state, and previ-
ous actions) and measure modality-wise attention patterns and head-output magnitudes. This
reveals early visual and instruction preprocessing heads, mid-layer heads that integrate visual,
state, and action-history signals, and late heads that format representations into action logits,
together forming an action-decoding circuit.
To directly link individual heads to control behavior, we introduce a KNN-based head–action
attribution score that evaluates how well each head’s representation predicts decoded contin-
uous action chunks, both globally and for specific motion and contact factors. Combining this
geometric signal with head-output norms and attention mass, we define an Action Relevance
Score (ARS) and a Hybrid Importance Score (HIS) that act as head-importance metrics for
identifying components that are functionally necessary for action decoding.
As a concrete downstream application of this circuit view, we use ARS/HIS to drive safe head
pruning. Pruning only the lowest-scoring heads under ARS/HIS preserves almost all LIBERO
performance at moderate pruning ratios, whereas magnitude-, attention-, or geometry-only cri-
teria quickly collapse control. These results show that VLA policies organize into structured
sensorimotor circuits at the head level, and that circuit-based importance scores provide a prac-
tical handle for safe head pruning, FLOP-aware compression, and future interpretability-driven
steering and safety interventions in embodied AI systems.