Beyond Myopic World Models: Long-Horizon End-to-End Training for Direct Future Prediction
The paper argues that long-horizon world models should be trained on the endpoint they must predict, not on short recursive steps.
Li, Xia, Hao, and Chen introduce DPWM, a non-recursive model that compresses an action sequence into one embedding and predicts the final observation in a single pass. The authors say this avoids error amplification from rolling out a model’s own predictions and makes end-to-end training practical at longer horizons. In their benchmarks, DPWM improves endpoint prediction over recursive baselines, with larger gains as the horizon grows. They also report that recurrent baselines improve when trained with the same endpoint objective, pointing to the loss function as the key factor. ArXiv · AI/CL/LG's note
Li, Xia, Hao, and Chen introduce DPWM, a non-recursive model that compresses an action sequence into one embedding and predicts the final observation in a single pass. The authors say this avoids error amplification from rolling out a model’s own predictions and makes end-to-end training practical at longer horizons. In their benchmarks, DPWM improves endpoint prediction over recursive baselines, with larger gains as the horizon grows. They also report that recurrent baselines improve when trained with the same endpoint objective, pointing to the loss function as the key factor. ArXiv · AI/CL/LG's note
score 4