World models provide foresight for sequential decision making by predicting the consequences of actions, showing strong promise for policy learning. Yet reasoning about counterfactual actions remains difficult because standard predictive objectives often downplay information crucial for evaluating action quality, and unexecuted actions lack observed futures for direct supervision.
To overcome these limitations, we introduce the World Action Critic (WAC) model. At each step the model proposes multiple action candidates and jointly learns
- Candidate‑conditioned future state prediction;
- Corresponding consequence evaluation.
During roll‑out, critic supervision directly shapes intermediate predictive representations, encouraging them to retain information needed to distinguish action quality. Observed futures supervise predictions for executed actions, while external task‑quality targets supply evaluation signals for counterfactual candidates. Joint optimization lets critic signals from these candidates propagate into predictive features without requiring their observed futures.
Our theoretical analysis shows that critic‑guided feature retention reduces information loss and yields a regret bound for candidate selection $R_T \le O(\sqrt{T})$. Applied to autonomous driving, WAC achieves 93.4 PDMS on the NAVSIM‑v1 benchmark, demonstrating the effectiveness of jointly learning predictive foresight and consequence evaluation.
Review: By tightly coupling action evaluation with future prediction, WAC fills a critical gap in traditional world models, offering a more reliable policy learning framework for high‑risk decision‑making scenarios.