Research

Paper

AI LLM March 19, 2026

Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models

Authors

Bryce Grant, Xijia Zhao, Peng Wang

Abstract

Vision-Language-Action (VLA) models combine perception, language, and motor control in a single architecture, yet how they translate multimodal inputs into actions remains poorly understood. We apply activation injection, sparse autoencoders (SAEs), and linear probes to six models spanning 80M--7B parameters across 394,000+ rollout episodes on four benchmarks. The visual pathway dominates action generation across all architectures: injecting baseline activations into null-prompt episodes recovers near-identical behavior, while cross-task injection steers robots toward source-task positions (99.8\% of X-VLA episodes align with the source trajectory), exposing spatially bound motor programs tied to scene coordinates rather than abstract task representations. Language sensitivity depends on task structure, not model design: when visual context uniquely specifies the task, language is ignored; when multiple goals share a scene, language becomes essential (X-VLA \texttt{libero\_goal}: 94\%$\to$10\% under wrong prompts vs.\ \texttt{libero\_object}: 60--100\% regardless). In all three multi-pathway architectures (\pizhalf{}, SmolVLA, GR00T), expert pathways encode motor programs while VLM pathways encode goal semantics ($2\times$ greater behavioral displacement from expert injection), and subspace injection confirms these occupy separable activation subspaces. Per-token SAE processing is essential for action fidelity on most architectures, though mean-pooling improves fidelity on X-VLA. Contrastive identification recovers 82+ manipulation concepts, and causal ablation reveals sensitivity spanning 28--92\% zero-effect rates independent of representation width. We release \textbf{Action Atlas} (https://action-atlas.com) for interactive exploration of VLA representations across all six models.

Metadata

arXiv ID: 2603.19233
Provider: ARXIV
Primary Category: cs.RO
Published: 2026-03-19
Fetched: 2026-03-20 06:02

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.19233v1</id>\n    <title>Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models</title>\n    <updated>2026-03-19T17:59:55Z</updated>\n    <link href='https://arxiv.org/abs/2603.19233v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.19233v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Vision-Language-Action (VLA) models combine perception, language, and motor control in a single architecture, yet how they translate multimodal inputs into actions remains poorly understood. We apply activation injection, sparse autoencoders (SAEs), and linear probes to six models spanning 80M--7B parameters across 394,000+ rollout episodes on four benchmarks. The visual pathway dominates action generation across all architectures: injecting baseline activations into null-prompt episodes recovers near-identical behavior, while cross-task injection steers robots toward source-task positions (99.8\\% of X-VLA episodes align with the source trajectory), exposing spatially bound motor programs tied to scene coordinates rather than abstract task representations. Language sensitivity depends on task structure, not model design: when visual context uniquely specifies the task, language is ignored; when multiple goals share a scene, language becomes essential (X-VLA \\texttt{libero\\_goal}: 94\\%$\\to$10\\% under wrong prompts vs.\\ \\texttt{libero\\_object}: 60--100\\% regardless). In all three multi-pathway architectures (\\pizhalf{}, SmolVLA, GR00T), expert pathways encode motor programs while VLM pathways encode goal semantics ($2\\times$ greater behavioral displacement from expert injection), and subspace injection confirms these occupy separable activation subspaces. Per-token SAE processing is essential for action fidelity on most architectures, though mean-pooling improves fidelity on X-VLA. Contrastive identification recovers 82+ manipulation concepts, and causal ablation reveals sensitivity spanning 28--92\\% zero-effect rates independent of representation width. We release \\textbf{Action Atlas} (https://action-atlas.com) for interactive exploration of VLA representations across all six models.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.RO'/>\n    <published>2026-03-19T17:59:55Z</published>\n    <arxiv:comment>Accepted to Multimodal Intelligence Workshop @ ICLR</arxiv:comment>\n    <arxiv:primary_category term='cs.RO'/>\n    <author>\n      <name>Bryce Grant</name>\n    </author>\n    <author>\n      <name>Xijia Zhao</name>\n    </author>\n    <author>\n      <name>Peng Wang</name>\n    </author>\n  </entry>"
}