Researchers have developed a system called TOWN-VLA (Think Only When Needed) to control how prompts are used in vision-language-action manipulation. The system separates candidate generation from permission to alter the policy input, and uses a fixed compatibility rule to authorize certain instructions while restoring the original prompt otherwise. This approach improves task success rates by 3-5% compared to previous methods, and is enforceable even for frozen controllers. --- Why it matters: This matters because it allows for more efficient and effective use of prompts in vision-language-action manipulation tasks, which can be crucial in applications such as robotics or autonomous vehicles where small improvements in performance can have significant safety implications. Source: https://arxiv.org/abs/2608.23224