Study finds the OpenVLA robot model ignores 'do not' in its instructions

Sudhanshu Srivastava and Faisal Aqlan have tested whether OpenVLA, an open vision-language-action robot model, obeys prohibitions such as not putting a bowl on the stove, in a paper on arXiv. On the LIBERO-Goal tasks, adding a negation made no detectable difference to which task the robot carried out: it did the forbidden task in 79% of negated runs and completed 78% of plain tasks.

Using activation patching, probes and steering, the authors traced where the model represents the goal it pursues. A steering vector at layer 8 redirected 12 of 18 runs, while fixed edits at layers 12 to 20 redirected none of 158. They recommend enforcing prohibitions outside the instruction and watching the model's internal goal read-outs, and say goal edits need to come before layer 12 or patch every action token.