A group of over 21 researchers from Google DeepMind, Harvard, and other institutions published a white paper titled Visual General Intelligence: A White Paper. The document, available on arXiv under the identifier 2608.25924, proposes shifting the focus of artificial general intelligence development from text-based models to visual data.
The authors argue that AI systems should prioritize learning from images, video, and geometric data to better understand and interact with the physical world. The research suggests that generative video models and self-supervised learning techniques could serve as a necessary foundation for AGI that language-only models currently lack.
The paper emerged from discussions at the CVPR 2026 Visual General Intelligence Workshop. Key contributors include Robert Geirhos from Google DeepMind and Yilun Du from Harvard. The work does not introduce a specific model or architecture but outlines a research agenda and potential benchmarks for future development.
This publication continues a series of DeepMind research papers on AGI, following the 2024 work Levels of AGI and the June 2026 report From AGI to ASI. The authors frame vision and language as complementary channels for intelligence rather than competing modalities, though they emphasize the need for a broader approach to reach AGI milestones.
