Planning at Inference: MCTS Test-Time Scaling for Long Video Generation

TL;DR

Frames long-video generation as a sequential decision-making problem and applies AlphaGo-style Monte Carlo Tree Search at inference time to pick the next clip continuation, instead of greedily extending a chunked rollout. Each MCTS node is a candidate continuation; children are evaluated with look-ahead rollouts and backpropagated rewards from a verifier, and a Multi-Tree MCTS variant improves exploration in the continuous video state space. The method is modular — no retraining of the base video model — and on Cosmos-Predict2 and other backbones beats Best-of-N, Greedy, and Beam search on object permanence, temporal coherence, and text–video alignment, producing coherent videos exceeding 20 seconds.

Key claims

Long-video generation can be cast as a sequential decision-making problem where each “move” is a candidate continuation chunk and MCTS over a tree of continuations beats Best-of-N, Greedy, and Beam search on object permanence, temporal coherence, and text–video alignment [Abstract].
A Multi-Tree MCTS variant is needed to make tree search work in the continuous generation space, vs. the discrete action space of board games [Abstract].
The method is modular and can be applied to existing backbones (Cosmos-Predict2 is the headline example) without retraining the base video model [Abstract].
Produces high-quality videos exceeding 20 seconds, surpassing Sora and Kling by 18% and 47% respectively (metric not specified in the Abstract excerpt) while maintaining comparable visual fidelity [Abstract].
Honest caveat from the authors’ tweet: the compute cost is increased; the authors argue it may be acceptable “once video model capability exceeds certain usability threshold” [author tweet].

Method

The task is sequential clip-by-clip extension of a base video DiT (Cosmos-Predict2 in the headline configuration). At each step, instead of committing to a single continuation, MCTS expands a tree of candidate continuations, runs look-ahead rollouts through some horizon, and backpropagates a reward signal up the tree. The search policy then picks the continuation with the best aggregated value, that clip is committed, and the search re-roots. Because the continuation space is continuous (every diffusion sample is unique), a Multi-Tree MCTS variant maintains multiple parallel search trees to improve exploration over what a single UCT-style tree would cover. No fine-tuning of the base generator is required.

Results

Beats Best-of-N, Greedy, and Beam search on object permanence, temporal coherence, and text–video alignment, on Cosmos-Predict2 and other (unspecified) backbones [Abstract].
Generates coherent videos > 20 seconds.
+18% over Sora and +47% over Kling on the (unspecified) head-to-head metric, at comparable visual fidelity [Abstract].
The authors flag two limitations explicitly: results are “limited by the quality of current generators and verifiers” [Abstract] — i.e. the reward model upper-bounds the search — and compute cost is increased relative to single-pass extension [author tweet].

Why it’s interesting

This is the first paper filed that applies game-playing-style MCTS as a test-time scaling axis on top of a video DiT, and it’s the cleanest mapping yet of Inference-Time Scaling‘s sequence-budget framing onto long-video generation: where DMS (Inference-Time Hyper-Scaling with KV Cache Compression) spends extra compute by compressing the KV cache so a reasoning model can write longer chains, this paper spends extra compute by tree-searching over candidate clips so an autoregressive video generator can extend longer without semantic drift. It contrasts cleanly with Self-Refining Video Sampling, which spends extra NFEs within a single diffusion sample (iterated Predict-and-Perturb at each fixed noise level) rather than across rollout branches — both are “scale the generator at inference” but they target orthogonal axes: refinement vs. search. The Abstract’s note that “results are limited by the quality of current generators and verifiers” is also a direct datapoint for the VLM-as-Evaluator story: a search-based TTS scheme is only as good as the reward model driving it, the same bottleneck WMReward (Autoregressive Video Generation discussion) identifies for inference-time alignment of frozen video AR bases.