Proxy Policy Steering adapts VLA models without weight access, a post says

A post describes Proxy Policy Steering (PPS) as a way to specialize vision-language-action (VLA) models for new tasks at inference time—when the model is being run—without accessing their weights. According to the post, PPS trains two lightweight proxy policies and uses their velocity-space residual to steer the frozen base model instead of fine-tuning it directly.