June 22nd, 2026 Research
Teaching VLMs to think visually
Reasoning in pixel space
Our latest research shows vision-language models reason better when they think visually: pointing at regions, sketching boxes, and cropping to verify — instead of answering from a single glance.
Models trained with visual tools outperform text-only chain-of-thought on grounding-heavy benchmarks.
The visual scratchpad
We give the model a small toolbox — point, box, crop, zoom — and reward it for checking its own claims against the pixels before answering.
What the numbers say
Across referring expression and visual QA suites, visually-grounded reasoning beats longer text rationales at equal token budgets.
What this means for builders
These techniques ship in the Loop models: ask for locations and the model shows its work as verifiable geometry, not prose.