Action-conditioned video generative models are designed to predict future frames given a command, but when trained on successful, goal-directed demonstrations they tend to ignore the commanded action. Instead, they render a plausible expert-like future no matter what is asked of them. This undermines their usefulness for control and planning.
The paper, "Training Controllable World Models via Action Influence Maximization," identifies this failure and proposes a training approach that maximizes the influence of the action on the generated outcome. The abstract does not detail the exact mechanism, but the intent is clear: force the model to actually use the action signal.
If successful, this could make world models more responsive to user commands, which is important for applications like robotics, autonomous driving, and interactive simulation. The work is at the preprint stage and has not yet been peer-reviewed.