OpenVLA

Stanford / TRI / UC Berkeley / Google DeepMind (open collaboration)

Structures found in this family (2)

Linear Direction, Linear Separability

By model (1)

OpenVLA-7B · 7B

Linear Direction, Linear Separability

Observations (5)

A value-like success direction is recoverable from frozen VLAs

Sparse logit-lens directions in VLA FFN value vectors steer robot actions

OpenVLA linearly encodes a state-transition vector in middle layers

Ablating general SAE feature directions in a VLA collapses grasp success

Symbolic object-relation and action states are linearly decodable above 90% accuracy from nearly every layer of a real LIBERO-spatial-finetuned OpenVLA-7B checkpoint

Papers

What Frozen VLAs Already Know About Success: A Probing Study of Value-Like Structure in Foundation Robot Policies (2026), Mechanistic Interpretability for Steering Vision-Language-Action Models (2025), Emergent World Representations in OpenVLA (2025), Sparse Autoencoders Reveal Interpretable and Steerable Features in VLA Models (2026), Probing a Vision-Language-Action Model for Symbolic States and Integration into a Cognitive Architecture (2025)