- The problem
- Robots and other systems often need quick judgments about where things are, how objects relate, or how far apart they are. Writing a text answer one word at a time and then converting it into a choice, number, or location adds delay and extra work.
- The idea
- MJev-Spatial has the model choose directly from possible answers instead of writing a response. For a distance or image location, it first chooses a broad range or region and then narrows that choice, like zooming in on a map. It uses existing parts of the model to score the choices rather than adding a separate answer-making component.
- What it got
- Across twelve spatial-understanding tests, it scores 65.64% overall and ranks first on seven. It runs 5.6 times faster than the original text-generating model it was built from.