arXiv:2607.09776v1 Announce Type: new Abstract: When adapting Vision Language Action (VLA) models to downstream tasks, multiple rounds of post training are required because a single round of data cannot resolve all issues, making continuous iterations necessary to progressively address the weakness