Sparse Feature Policy Unlearning Mitigates State Hallucination in Vision-Language-Action Models

Sparse Feature Policy Unlearning Mitigates State Hallucination in Vision-Language-Action Models. By Jiho Lee and four others, from Chung-Ang University, the Georgia Institute of Technology, and one other institution. Published on the seventh of October. No code repository is linked to it.

Vision-language-action models generalize well in robotic manipulation, drawing on the rich representations of pretrained vision-language models. Yet real-world deployment is held back by unreliable behaviours that keep recurring. The authors study one of them, which they call state hallucination. The robot carries on as if some robot-object state had been achieved when it has not. The policy behaves as though a step is done, while the scene in front of it says otherwise.

Two analyses point at where the failure lives. First, state hallucination coincides with weakened attention to the task-relevant parts of the image. Second, a mechanistic look through sparse autoencoders shows that particular sparse features, associated with hallucination, are active when these failures occur. If the failure has identifiable features, those features give a handle for removing it. That is the basis for SOUL, which stands for Sparse feature pOlicy UnLearning.

SOUL selectively unlearns the policy knowledge tied to state hallucination behaviours. The sparse features found in hallucination failures become explicit forgetting targets. The sparse features found in successful behaviours become retention targets. So the unlearning has both a direction to move away from and a set of things it must keep. That pairing matters for a robot policy, because the knowledge behind a hallucinated step and the knowledge behind a correct one sit in the same model. Forgetting without a retention target would risk taking useful manipulation skill along with the failure. With both targets defined from the model's own features, the edit is aimed rather than broad.

Experiments span different vision-language-action architectures, in simulation and in the real world. SOUL substantially reduces hallucinated failures and improves task success, without substantially compromising the manipulation capabilities the model already had.

The authors frame the result as a case for interpretability with a practical use. Feature analysis here does more than explain a failure after the fact. It supplies the targets for editing the policy, a basis for selectively modifying undesirable knowledge in a robot policy while leaving the rest of it in place.