Existing video multimodal large language models (MLLMs) structure their visual input primarily along the time axis, treating observations as a sequence of frames. But embodied spatial reasoning—such as navigating a room or manipulating objects—demands an understanding of how things are arranged in space, not just when they appear. The abstract notes that this temporal bias can leave visual representations locally incomplete, suggesting a mismatch between the model's internal organization and the task's requirements.
The paper, titled "Thinking from Above," proposes a shift in perspective: instead of ordering tokens by time, it represents observations using bird's-eye-view visual tokens. This spatial layout mirrors the top-down view often used in planning and reasoning about physical environments, potentially making the representation more aligned with the geometry of the task.
The significance lies in efficiency and effectiveness. By reorganizing tokens spatially, the model may avoid redundant temporal processing and focus on the spatial relationships that matter. The abstract is truncated, but the title and framing indicate that this bird's-eye-view tokenization is designed to make spatial reasoning more efficient, offering a possible path forward for embodied AI systems that must act in the real world.