Parallel Predictive World Models for Accurate and Efficient Long-Horizon Planning
PPWM replaces step-by-step rollout with parallel finite-horizon prediction, and reports both lower error and faster planning.
The paper says autoregressive world-model planning keeps feeding predicted states back into the model, which slows long horizons and compounds decoded-state errors. PPWM instead conditions each horizon on the causal action prefix, lets future representations interact before decoding, and removes that decoded-state feedback path. In four visual-control tasks, it reports the lowest long-horizon prediction error and the best CEM simulator success among the tested interfaces. It also reports more than a 3x average CEM planning speedup over the autoregressive LeWM baseline. ArXiv · AI/CL/LG's note
The paper says autoregressive world-model planning keeps feeding predicted states back into the model, which slows long horizons and compounds decoded-state errors. PPWM instead conditions each horizon on the causal action prefix, lets future representations interact before decoding, and removes that decoded-state feedback path. In four visual-control tasks, it reports the lowest long-horizon prediction error and the best CEM simulator success among the tested interfaces. It also reports more than a 3x average CEM planning speedup over the autoregressive LeWM baseline. ArXiv · AI/CL/LG's note
score 5