Introducing AlphaGo search, new MCTS video generation framework with longer video duration than Sora According to Beating Monitoring, researchers from the University of Waterloo, Brown University, and other institutions proposed a new Test Time Scaling framework called Planning at Inference in their paper submitted to ICLR 2026, which for the first time applies AlphaGo's Monte Carlo Tree Search (MCTS) algorithm to the generation of long videos. This framework models the task of generating long videos as a sequential decision-making problem. The system introduces MCTS in the inference stage and uses look ahead rollouts and backpropagation rewards to evaluate multiple video continuation segments, fundamentally solving the semantic drift and error accumulation problems commonly faced in traditional block or single generation. In order to achieve efficient exploration in the continuous video generation space, the research team specially designed the Multi Tree MCTS (Multi Tree Monte Carlo Tree Search) variant. Compared to the traditional method of using a single search tree under a fixed computing power budget, the multi tree architecture can conduct extensive searches in the continuous state space with more reasonable pruning and branching coefficients, significantly improving exploration efficiency. More importantly, Planning at Inference has a highly modular feature and belongs to a fully plug and play inference optimization solution. Developers do not need to retrain or fine tune the underlying large model, and can directly deploy this solution on existing video generation bases. In experiments based on NVIDIA's open-source video prediction model Cosmos-Predict2, Planning at Inference demonstrated strong generative performance. In the evaluation of long video generation, this solution successfully generated high-quality coherent videos of over 20 seconds. Test data shows that in core indicators such as object persistence, temporal coherence, and text video alignment, the quality of MCTS search generation has significantly improved compared to traditional baseline methods such as Greedy Search, Beam Search, and Best-of-N. Compared to the leading closed source models in the industry, this method generates videos that are 18% longer than Sora and 47% longer than Kling in terms of duration, while maintaining comparable image precision and visual fidelity. Although the search mechanism brings excellent picture coherence, introducing multi tree search in the inference stage also incurs high computational overhead. The researchers admitted that the current Planning at Inference framework is significantly slower in generation speed than traditional autoregressive direct generation, which to some extent limits the possibility of real-time deployment. However, with the efficiency evolution of the underlying video generation base and the continuous growth of computing hardware computing power, the scaling route of reasoning in exchange for image quality at the cost of computing is expected to become a key technological path for the practical engineering of long video generation after the basic capabilities of large models break through specific thresholds. [Original link]