Latent World Models' Planning Fails Due to Objective, Not Prediction.

When latent world models struggle with long-horizon planning, it's commonly assumed that the underlying predictor is at fault, degrading over extended time steps. However, this study challenges that assumption, demonstrating that the primary bottleneck often lies with the planner's objective function. Researchers reproduced LeWorldModel on a TwoRoom task and found that the world model's predictions remained relatively accurate far into the future, while the planner's effective horizon was much shorter. The core issue identified was that the cross-entropy-method planning, which minimizes squared latent distance, does not accurately track true spatial distance beyond a certain point. This flawed objective can even incentivize moving away from the goal. Crucially, the necessary information for successful planning was still present within the latent embeddings. By replacing only the objective function—without any retraining or GPU usage—the success rate for long-horizon goals dramatically increased from 26.0% to 98.0%, matching short-horizon performance. This highlights that a "best cost" objective is not necessarily the most accurate one, but rather one that captures relevant properties like reachability over mere proximity.