Long-document understanding remains a major challenge for multimodal AI systems, as processing dozens of visually rich pages is computationally expensive and often leads to errors. A new paper from Hugging Face researchers introduces InSight-doc, an agentic visual perception framework that treats image resolution as an adaptive resource during reasoning, rather than a fixed input setting.
InSight-doc begins by analyzing documents at low resolution, then selectively zooms into high-resolution regions to gather finer evidence. This approach eliminates the need for an external retriever, as the model itself decides where to focus. The researchers trained the system using a custom corpus of 17.9K supervised fine-tuning examples with region-level zoom-in trajectories, plus 19.2K hard reinforcement learning examples.
According to the paper, the resulting InSight-doc-8B model outperforms its baseline by 4.3 to 16.4 accuracy points across document VQA benchmarks. On long documents, it reduces hallucination by more than 40% and inference latency by 41%–68%, all while maintaining an accuracy advantage.
The team has released the code, datasets, and model weights on GitHub, along with a live demo and a replayable demo. The work is part of a growing trend toward adaptive computation in vision-language models, where reasoning effort is allocated dynamically based on task complexity.