World-action models (WAMs) are an emerging approach that co-trains a policy with future prediction, using the predictive power of video models to supply physical priors for action generation. A new arXiv paper, FACT (Failure-Aware Causal Training), builds on this idea by introducing a training scheme that is explicitly failure-aware and causal.

The abstract explains the motivation: because WAMs rely on future prediction, errors or failures in that prediction can propagate into action generation. FACT is designed to address this by making the training process aware of such failures, rather than treating all predictions equally.

The provided source is a replacement (v2) of the arXiv paper, but the abstract is truncated mid-sentence, so the specific mechanism, experiments, and results are not available in this excerpt. As a single source, there are no conflicting accounts to compare; this summary is limited to what the abstract states.