A group of researchers—including former xAI world model lead Ethan He, NVIDIA senior researchers Linnan Wang and Ashwath Aithal, and ex-NVIDIA deep learning intern Ritvik Bale—has submitted a paper to ICLR 2026 proposing a new test-time scaling framework called Planning at Inference. It brings Monte Carlo Tree Search (MCTS), the algorithm that powered AlphaGo, to long video generation for the first time as a plug-and-play module.
The framework treats long video generation as a sequential decision problem. At inference time, it uses MCTS with look-ahead rollouts and backpropagated rewards to evaluate multiple possible continuations, addressing the semantic drift and error accumulation that plague traditional chunked or single-pass generation.
To efficiently explore the continuous video generation space, the team designed a Multi-Tree MCTS variant. Unlike a single search tree under a fixed compute budget, this variant automatically spawns new trees when the search tree exceeds a preset size, expanding the search space and reducing sensitivity to poor initialization. Crucially, Planning at Inference is fully modular: developers can drop it into any existing video generation model without retraining or fine-tuning.
In experiments with NVIDIA’s open-source video prediction model Cosmos-Predict2, Planning at Inference generated high-quality coherent videos over 20 seconds long. On key metrics like object persistence, temporal coherence, and text-to-video alignment, MCTS search dramatically outperformed traditional baselines such as greedy search, beam search, and Best-of-N. Compared to leading closed-source models, the approach produced videos 18% longer than OpenAI’s Sora and 47% longer than Kuaishou’s Kling, while maintaining comparable detail and visual fidelity.
But there’s a catch: the search mechanism comes with a heavy compute cost. The researchers admit that Planning at Inference is significantly slower than standard autoregressive generation, limiting real-time deployment for now. Still, as base models become more efficient and hardware gets faster, trading compute for quality could become the key path to practical long video generation — once the underlying model crosses a certain capability threshold.