Reasoning via Video: The First Evaluation of Video Models' Reasoning Abilities through Maze-Solving Tasks

TL;DR

VR-Bench introduces a principled framework for evaluating reasoning via video by grounding tasks in maze-solving trajectories. It demonstrates that fine-tuned open-source video models can achieve robust spatial-temporal reasoning and generalization, outperforming vision-language models on complex scenarios. The benchmark also reveals a test-time scaling effect, where diverse sampling during inference improves reliability by roughly 10–20%. Together, these findings position reasoning via video as a scalable and expressive paradigm for spatial reasoning tasks, with strong implications for future benchmarks and embodied reasoning settings.

Abstract

Video Models have achieved remarkable success in high-fidelity video generation with coherent motion dynamics. Analogous to the development from text generation to text-based reasoning in language modeling, the development of video models motivates us to ask: Can video models reason via video generation? Compared with the discrete text corpus, video grounds reasoning in explicit spatial layouts and temporal continuity, which serves as an ideal substrate for spatial reasoning. In this work, we explore the reasoning via video paradigm and introduce VR-Bench -- a comprehensive benchmark designed to systematically evaluate video models' reasoning capabilities. Grounded in maze-solving tasks that inherently require spatial planning and multi-step reasoning, VR-Bench contains 7,920 procedurally generated videos across five maze types and diverse visual styles. Our empirical analysis demonstrates that SFT can efficiently elicit the reasoning ability of video model. Video models exhibit stronger spatial perception during reasoning, outperforming leading VLMs and generalizing well across diverse scenarios, tasks, and levels of complexity. We further discover a test-time scaling effect, where diverse sampling during inference improves reasoning reliability by 10--20%. These findings highlight the unique potential and scalability of reasoning via video for spatial reasoning tasks.

Figures (20)

• Figure 1: Overview of VR-Bench. (A) Maze Types. VR-Bench comprises five maze types—Regular Maze, Irregular Maze, 3D Maze, Trapfield, and Sokoban—covering both 2D and 3D settings as well as diverse task structures, yielding a broad range of spatial reasoning scenarios. (B) Reasoning via Video Paradigm. VR-Bench adopts a chain-of-frame reasoning paradigm wiedemer2025video, requiring models to produce frame-by-frame inferences that capture sequential visual reasoning. (C) Benchmark Performance. Leading VLMs and video models are evaluated on four core metrics across all maze types, revealing clear differences in spatial reasoning capability. (D) Additional Analysis. VR-Bench also supports evaluations on difficulty generalization, texture generalization, maze-type generalization, and test-time scaling, enabling a comprehensive assessment of model robustness and generalization.
• Figure 2: Variations of difficulty level and maze texture
• Figure 3: Bad case visualization and VLM-as-judge schematic
• Figure 4: Model performance (PR and SR) on Irregular Maze and Trapfield across difficulty levels. Each curve represents a baseline, while the dashed and dotted lines indicate VLM and Video Model averages. Results for other maze types are in the Appendix.
• Figure 5: Performance on Irregular Maze using Wan-R1 under test-time scaling. Results are shown across different sampling numbers ($K \in {1,4,8,12,16}$) and difficulty levels (Easy, Medium, Hard). Results for other maze types are in the Appendix.