Multimodal models often struggle with fine-grained perception, where small details are easily lost in the global context of an image. Current solutions rely on "Thinking-with-Images," a method where models iteratively zoom into specific regions during inference to find small pieces of evidence. While effective, this approach is computationally heavy and slow, requiring multiple tool calls and repeated visual processing that drive up latency for real-world applications.

To solve this, researchers introduced "Region-to-Image Distillation." This technique effectively moves the zooming process from the conversation stage to the training stage. By using a powerful teacher model to analyze micro-cropped images and generate high-quality labels, the team distilled that "zoomed-in" knowledge back into a smaller student model. This allows the student to perceive tiny details in a single glance without needing to zoom in manually during use.

The team's "ZwZ" models achieve SOTA performance on several benchmarks, demonstrating that complex agentic behaviors can be internalized for faster execution. Alongside the models, the researchers released ZoomBench, a new VQA benchmark specifically designed to measure the gap between global and regional visual understanding. This work paves the way for more efficient AI agents capable of high-precision visual reasoning in areas like GUI navigation and document analysis.

Multimodal models often struggle with fine-grained perception, where small details are easily lost in the global context of an image. Current solutions rely on "Thinking-with-Images," a method where models iteratively zoom into specific regions during inference to find small pieces of evidence. While effective, this approach is computationally heavy and slow, requiring multiple tool calls and repeated visual processing that drive up latency for real-world applications.

To solve this, researchers introduced "Region-to-Image Distillation." This technique effectively moves the zooming process from the conversation stage to the training stage. By using a powerful teacher model to analyze micro-cropped images and generate high-quality labels, the team distilled that "zoomed-in" knowledge back into a smaller student model. This allows the student to perceive tiny details in a single glance without needing to zoom in manually during use.

The team's "ZwZ" models achieve SOTA performance on several benchmarks, demonstrating that complex agentic behaviors can be internalized for faster execution. Alongside the models, the researchers released ZoomBench, a new VQA benchmark specifically designed to measure the gap between global and regional visual understanding. This work paves the way for more efficient AI agents capable of high-precision visual reasoning in areas like GUI navigation and document analysis.