Inference-time Physics Alignment of Video Generative Models with Latent World Models

CVPR 26 2026 67.8 method

TLDR

Using latent world models as rewards to improve physics plausibility of video generation at inference time, achieving first place in PhysicsIQ Challenge.

Reasoning

Strengths: novel inference-time alignment approach, strong empirical results including competition win and human preference study. Weaknesses: reliance on a specific latent world model (VJEPA-2), limited discussion of limitations or generalization.

Read-first score

Read-first score 67.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 44.

Recency 8%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=True; dataset=False; markers=github

Topical relevance 42%

62.9

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=evaluation,validation

Field roles

FrontierReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 200.

Keyword Scores

world model

video world model

world dynamics prediction

world simulator

generative world model

model-based reinforcement learning world model

interactive world model

Deep Analysis

Innovations

Identifying that physics plausibility shortfall in video generation is due to suboptimal inference strategies, not just pre-training
Introducing WMReward, an inference-time alignment method that uses a latent world model as a reward to search and steer multiple candidate denoising trajectories
Demonstrating that scaling test-time compute via latent world model guidance improves physics plausibility across diverse generation settings

Methodology

The authors treat improving physics plausibility as an inference-time alignment problem. They leverage the strong physics prior of a latent world model (VJEPA-2) as a reward to search and steer multiple candidate denoising trajectories, enabling scaling test-time compute for better generation performance. Evaluation is conducted across image-conditioned, multiframe-conditioned, and text-conditioned generation settings, with validation from human preference study and the ICCV 2025 Perception Test PhysicsIQ Challenge.

Key Results

The approach substantially improves physics plausibility across multiple generation settings, validated by human preference study. In the ICCV 2025 Perception Test PhysicsIQ Challenge, it achieves a final score of 62.64%, winning first place and outperforming the previous state of the art by 7.42%.

Resources

Paper Code