DeepSeek Teaches AI to Point: Visual Reasoning Gets a 17-Point Leap With 'Thinking in Coordinates'

One of the most persistent failure modes of multimodal AI has been spatial reasoning: the inability to reliably locate, relate, and track objects within images in the way a human naturally would when scanning a diagram or photograph. DeepSeek-AI's May release, "Thinking with Visual Primitives," addresses that gap with a technique that embeds visual coordinates directly as special tokens inside the reasoning chain — allowing the model to, in effect, point to specific regions of an image mid-thought, much like a person circling items on a whiteboard while talking through a problem.

The benchmark results are striking. In testing on topological reasoning tasks — problems that require understanding spatial relationships between objects, not just identifying them — the system achieved a 17% improvement over its nearest multimodal competitor. The technique compounds with chain-of-thought prompting: as the model works through a visual problem, it can anchor each reasoning step to a coordinate, preventing the spatial drift that causes current multimodal models to confidently describe the wrong region of an image. The implications extend beyond visual question answering. Autonomous systems, medical imaging analysis, and robotics control all depend on grounding language reasoning to spatial reality.

The month produced a second significant research result with broader implications. Researchers at the University of Pennsylvania introduced "Mollifier Layers" — a technique integrating classical mathematical smoothing functions into neural networks to solve inverse partial differential equations with greater stability. Separately, a VARnet neural system reanalyzed NASA's NEOWISE archival data using wavelet decomposition and identified approximately 1.5 million previously undetected celestial objects, including quasars and supernovae hidden behind dust clouds. The astronomy result is notable not for what it reveals about AI capability — it used established methods — but for what it reveals about latent scientific value sitting in existing datasets, waiting for sufficiently capable retrieval systems to surface it.