Researchers from the University of Waterloo and Brown University have introduced a novel framework, Planning at Inference, which applies the Monte Carlo Tree Search (MCTS) algorithm to long-form video generation. This approach, detailed in a paper submitted to ICLR 2026, models video generation as a sequential decision problem, using MCTS to evaluate video continuations and address issues like semantic drift and error accumulation. The framework features a Multi-Tree MCTS variant, allowing for efficient exploration in continuous video generation spaces. It is designed to be modular and can be integrated with existing video generation models without retraining. Experiments using NVIDIA's Cosmos-Predict2 model showed that Planning at Inference produces high-quality videos over 20 seconds long, outperforming traditional methods in metrics such as object persistence and temporal coherence. The framework generates videos 18% longer than Sora and 47% longer than Kling, though it incurs significant computational overhead, limiting real-time deployment potential.