Multimodal Large Language Models (MLLMs) have revolutionized how we interact with images, yet they often hit a wall when asked to understand how a scene shifts from different perspectives. If a camera moves just a few inches, traditional systems frequently struggle with depth errors and geometric warping, causing their internal representation of space to falter. A research team from KAIST has introduced "token warping" as a compelling solution to this persistent spatial blind spot.

Instead of attempting to distort raw pixels—a process that often introduces messy geometric distortions—the researchers propose manipulating image tokens, the abstract mathematical representations the model uses internally. By employing "backward token warping," the model defines a grid on the target viewpoint and retrieves corresponding data for each point. This method is far more stable than trying to "stretch" or "shift" the image itself, allowing the model to better maintain the semantic coherence of the scene.

The team validated this approach with ViewBench, a custom benchmark designed to stress-test spatial reasoning. The results are clear: this method consistently outperforms pixel-wise approaches and standard spatial fine-tuning. By bridging the gap between flat images and dynamic spatial understanding, this advancement moves AI closer to navigating and interpreting physical environments with human-like spatial awareness.

Multimodal Large Language Models (MLLMs) have revolutionized how we interact with images, yet they often hit a wall when asked to understand how a scene shifts from different perspectives. If a camera moves just a few inches, traditional systems frequently struggle with depth errors and geometric warping, causing their internal representation of space to falter. A research team from KAIST has introduced "token warping" as a compelling solution to this persistent spatial blind spot.

Instead of attempting to distort raw pixels—a process that often introduces messy geometric distortions—the researchers propose manipulating image tokens, the abstract mathematical representations the model uses internally. By employing "backward token warping," the model defines a grid on the target viewpoint and retrieves corresponding data for each point. This method is far more stable than trying to "stretch" or "shift" the image itself, allowing the model to better maintain the semantic coherence of the scene.

The team validated this approach with ViewBench, a custom benchmark designed to stress-test spatial reasoning. The results are clear: this method consistently outperforms pixel-wise approaches and standard spatial fine-tuning. By bridging the gap between flat images and dynamic spatial understanding, this advancement moves AI closer to navigating and interpreting physical environments with human-like spatial awareness.