Present but Not Remembered: Auditing How Frozen VLAs Encode, Deploy, and Steer Visual History

Across Octo-Small, Octo-Base, and CronusVLA, linear probes recover previous-frame content, while predicting the residual left by a current-frame predictor remains difficult. The reported probe sweeps yield residual-target $R^2$ scores of at most $0.019$, alongside raw-target scores of $0.142$–$0.389$. We examine how this shared pattern relates to action outputs using representation interventions on BridgeData V2 observation pairs. Blacking out the current image increases the action displacement caused by replacing history in Octo and decreases it in CronusVLA. In the detailed Octo-Small analysis, the occlusion-related increase is concentrated in early layers. A separate Octo-Small injection study finds a small shift toward the action associated with the injected history ($+0.069$, $95\%$ CI $[0.043,0.094]$); estimates for clean-action recovery and an ambiguous-observation test remain uncertain. Cross-model evidence thus covers residual prediction and occlusion responses, while recovery is an Octo-Small case study. These open-loop results motivate evaluating visual memory through its added predictive value, the action changes it produces, and recovery toward a specified action target.