We find that copying the VLM gradient from the final step and applying it uniformly to all steps outperforms backpropagating that gradient through the sampling trajectory on OneIG alignment.