Thinking in Frames: How Visual Context and Test-Time Scaling Empower Video Reasoning

TL;DR

This work reframes visual reasoning as video-based planning, using generated frame sequences as intermediate reasoning traces between an initial state and a goal under constraints. By evaluating two regimes—MazeNavigation (low visual change, discrete actions) and TangramPuzzle (high visual change, continuous manipulation)—the study demonstrates that visual context serves as a robust geometric prior and that increasing the video-inference budget at test time enhances zero-shot generalization in sequential planning. The approach outperforms text-based baselines in both IID and OOD settings, uncovers a scaling law where more frames improve reasoning, and reveals task-dependent limits tied to geometric fidelity. The findings suggest video generation is a scalable, interpretable paradigm for visual reasoning with practical implications for robust, geometry-aware planning in embodied systems.

Abstract

Vision-Language Models have excelled at textual reasoning, but they often struggle with fine-grained spatial understanding and continuous action planning, failing to simulate the dynamics required for complex visual reasoning. In this work, we formulate visual reasoning by means of video generation models, positing that generated frames can act as intermediate reasoning steps between initial states and solutions. We evaluate their capacity in two distinct regimes: Maze Navigation for sequential discrete planning with low visual change and Tangram Puzzle for continuous manipulation with high visual change. Our experiments reveal three critical insights: (1) Robust Zero-Shot Generalization: In both tasks, the model demonstrates strong performance on unseen data distributions without specific finetuning. (2) Visual Context: The model effectively uses visual context as explicit control, such as agent icons and tangram shapes, enabling it to maintain high visual consistency and adapt its planning capability robustly to unseen patterns. (3) Visual Test-Time Scaling: We observe a test-time scaling law in sequential planning; increasing the generated video length (visual inference budget) empowers better zero-shot generalization to spatially and temporally complex paths. These findings suggest that video generation is not merely a media tool, but a scalable, generalizable paradigm for visual reasoning.

Figures (13)

• Figure 1: Video generation models as visual reasoners, empowered by (1) enriched visual context for improved geometric control and (2) visual test-time scaling that allocates a larger inference-frame budget and enables stronger performance on long-horizon, complex sequential planning tasks, together demonstrating robust generalization across diverse scenarios.
• Figure 2: Generated solution of TangramPuzzle by different system variants. For Qwen-3-VL, we visualize the layout based on the predicted coordinates and rotations. We crop the main area for the predictions from image editing model and video generation model. For video generation model, we only select the last frame as illustration here. For full details, please refer to Figure \ref{['appfig:tangram showcase']}.
• Figure 3: Visual Test-Time Scaling for MazeNavigation using unseen icon with more inference budget. Row 1 shows the performance curve when increasing the total number of frames per video; Row 2 shows the performance curve when changing the scaling factor $\kappa$ to allocate a different number of frames per discrete step in the maze solution. Detailed results for both settings are shown in Figure \ref{['appfig:maze scaling frames']} and \ref{['appfig:maze scaling kappa']}.
• Figure 4: Agent Icons for MazeNavigation during training and visual OOD evaluation.
• Figure 5: Illustration of different variants for TangramPuzzle.