# Video Models Reason Early: Exploiting Plan Commitment for Maze Solving

arXiv: https://arxiv.org/abs/2603.30043 · HTML: https://mlsanity.com/paper/2603.30043/

## Key facts

- Cited by 9 papers in the map (data through 14 Oct 2026).
- Placed in Multimodal reasoning, one of 17 directions in LLM reasoning.
- Builds on 29 of its 43 references that are in the map.
- First posted on arXiv 31 Mar 2026.

## Placed in

- LLM reasoning / [Multimodal reasoning](https://mlsanity.com/lr/multimodal-reasoning/)
- Diffusion models / [Video generation](https://mlsanity.com/diffusion/video-generation/)

## Builds on (most cited first, 29 of 29)

- [Flow Matching for Generative Modeling](https://mlsanity.com/paper/2210.02747/) (2022)
- [Scaling Rectified Flow Transformers for High-Resolution Image Synthesis](https://mlsanity.com/paper/2403.03206/) (2024)
- [Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow](https://mlsanity.com/paper/2209.03003/) (2022)
- [SAM 2: Segment Anything in Images and Videos](https://mlsanity.com/paper/2408.00714/) (2024)
- [Wan: Open and Advanced Large-Scale Video Generative Models](https://mlsanity.com/paper/2503.20314/) (2025)
- [Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets](https://mlsanity.com/paper/2311.15127/) (2023)
- [Prompt-to-Prompt Image Editing with Cross Attention Control](https://mlsanity.com/paper/2208.01626/) (2022)
- [Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters](https://mlsanity.com/paper/2408.03314/) (2024)
- [eDiff-I: Text-to-Image Diffusion Models with an Ensemble of Expert Denoisers](https://mlsanity.com/paper/2211.01324/) (2022)
- [Cosmos World Foundation Model Platform for Physical AI](https://mlsanity.com/paper/2501.03575/) (2025)
- [Structure and Content-Guided Video Synthesis with Diffusion Models](https://mlsanity.com/paper/2302.03011/) (2023)
- [Diffusion Models already have a Semantic Latent Space](https://mlsanity.com/paper/2210.10960/) (2022)
- [Minigrid & Miniworld: Modular & Customizable Reinforcement Learning Environments for Goal-Oriented Tasks](https://mlsanity.com/paper/2306.13831/) (2023)
- [Perception Prioritized Training of Diffusion Models](https://mlsanity.com/paper/2204.00227/) (2022)
- [Inference-Time Scaling for Diffusion Models beyond Scaling Denoising Steps](https://mlsanity.com/paper/2501.09732/) (2025)
- [A General Framework for Inference-time Scaling and Steering of Diffusion Models](https://mlsanity.com/paper/2501.06848/) (2025)
- [Video models are zero-shot learners and reasoners](https://mlsanity.com/paper/2509.20328/) (2025)
- [Learning to Navigate in Complex Environments](https://mlsanity.com/paper/1611.03673/) (2016)
- [Dynamical Regimes of Diffusion Models](https://mlsanity.com/paper/2402.18491/) (2024)
- [A Phase Transition in Diffusion Models Reveals the Hierarchical Nature of Data](https://mlsanity.com/paper/2402.16991/) (2024)
- [Vision Language Models are Biased](https://mlsanity.com/paper/2505.23941/) (2025)
- [Critical windows: non-asymptotic theory for feature emergence in diffusion models](https://mlsanity.com/paper/2403.01633/) (2024)
- [Video-T1: Test-Time Scaling for Video Generation](https://mlsanity.com/paper/2503.18942/) (2025)
- [A Very Big Video Reasoning Suite](https://mlsanity.com/paper/2602.20159/) (2026)
- [Manifolds, Random Matrices and Spectral Gaps: The geometric phases of generative diffusion](https://mlsanity.com/paper/2410.05898/) (2024)
- [Reasoning via Video: The First Evaluation of Video Models' Reasoning Abilities through Maze-Solving Tasks](https://mlsanity.com/paper/2511.15065/) (2025)
- [DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models](https://mlsanity.com/paper/2512.24165/) (2025)
- [Cross-fluctuation phase transitions reveal sampling dynamics in diffusion models](https://mlsanity.com/paper/2511.00124/) (2025)
- [Beyond the Last Frame: Process-aware Evaluation for Generative Video Reasoning](https://mlsanity.com/paper/2512.24952/) (2025)

## Built on by (most cited first, 9 of 9)

- [VideoCoCo: Code-as-CoT for Physically-Consistent Video Generation via an Agentic Dual-Engine System](https://mlsanity.com/paper/2607.27380/) (2026)
- [VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning](https://mlsanity.com/paper/2608.26105/) (2026)
- [VLMs are Good Teachers for Video Reasoning via Adaptive Test-Time Optimization](https://mlsanity.com/paper/2606.02564/) (2026)
- [VGI-Bench: Probing Visual Intelligence in Video Generation Models](https://mlsanity.com/paper/2608.19583/) (2026)
- [The Seriality Gap in Video Diffusion Models](https://mlsanity.com/paper/2607.13031/) (2026)
- [Why Do Video Diffusion Models Violate Physics? Unveiling the Flaws in Attention Mechanisms](https://mlsanity.com/paper/2609.23658/) (2026)
- [Visual prompt engineering for video models](https://mlsanity.com/paper/2607.25537/) (2026)
- [Visual General Intelligence: A White Paper](https://mlsanity.com/paper/2608.25924/) (2026)
- [Learning via Self-Consistency for Diffusion-based Video Reasoning](https://mlsanity.com/paper/2609.36826/) (2026)

## Abstract (arXiv)

Video diffusion models exhibit emergent reasoning capabilities like solving mazes and puzzles, yet little is understood about how they reason during generation. We take a first step towards understanding this and study the internal planning dynamics of video models using 2D maze solving as a controlled testbed. Our investigations reveal two findings. Our first finding is early plan commitment: video diffusion models commit to a high-level motion plan within the first few denoising steps, after which further denoising alters visual details but not the underlying trajectory. Our second finding is that path length, not obstacle density, is the dominant predictor of maze difficulty, with a sharp failure threshold at 12 steps. This means video models can only reason over long mazes by chaining together multiple sequential generations. To demonstrate the practical benefits of our findings, we introduce Chaining with Early Planning, or ChEaP, which only spends compute on seeds with promising early plans and chains them together to tackle complex mazes. This improves accuracy from 7% to 67% on long-horizon mazes and by 2.5x overall on hard tasks in Frozen Lake and VR-Bench across Wan2.2-14B and HunyuanVideo-1.5. Our analysis reveals that current video models possess deeper reasoning capabilities than previously recognized, which can be elicited more reliably with better inference-time scaling.