Training to improve planning reliability and extend the effective horizon in video diffusion models

Determine whether training approaches can produce video diffusion models that plan more reliably and over longer horizons than currently observed, overcoming the sharp failure cliff on trajectories exceeding the single-generation window.

Background

The study demonstrates that path length, not obstacle density, is the dominant driver of maze difficulty, with a marked failure threshold at approximately 12 steps. This suggests that current models are limited by an effective generation horizon, motivating chaining across multiple generations to solve longer tasks.

While inference-time strategies (Early Planning Beam Search and Chaining with Early Planning) substantially improve performance without retraining, it remains unresolved whether training can directly enhance planning reliability and extend the effective planning horizon.