An arXiv preprint, 'Thinking with Images without Coordinates' (arXiv:2610.14265), addresses a growing trend in multimodal large language models. The authors note that recent work has shifted toward a 'thinking with images' paradigm, where models reason over visual inputs directly rather than through textual descriptions. This approach is meant to improve reasoning faithfulness.

However, the abstract points out that existing methods in this paradigm rely heavily on explicit coordinate-based grounding—essentially, they still need to know where things are in an image. The paper identifies something about this reliance, though the abstract cuts off before stating the full finding. Based on the available text, the work appears to question whether coordinate-based grounding is necessary for faithful visual reasoning.

The preprint is new and has not yet been peer-reviewed. As with all arXiv submissions, the claims should be treated as preliminary.