DPWM Argues World Models Should Skip Recursive Rollout Entirely and Predict the Endpoint in One Forward Pass
Most world models train on few-step prediction losses but deploy by recursively rolling out their own predictions — a mismatch where local errors amplify and transitions with very different downstream influence get weighted identically. The Direct Prediction World Model compresses an action sequence of arbitrary length into a single embedding and predicts the endpoint observation in one forward pass, avoiding recurrent rollout in both inference and gradient propagation at horizons where unrolled autoregressive training destabilizes. Crucially, the authors show recurrent baselines gain similarly when retrained with the same endpoint objective — evidence the training objective, not the backbone, drives long-horizon accuracy.
↳ Follow the thread