ziyuguo99/atlas

ATLAS is an AI research project investigating how visual reasoning models can be made more efficient. The central finding, described in an accompanying academic paper, is that a single discrete word, meaning one token of text output, is sufficient to bridge two competing approaches to visual reasoning: "agentic" methods (where an AI takes multiple steps, like an agent working through a problem) and "latent" methods (where reasoning happens inside the model's hidden layers without explicit steps visible to users). The problem it addresses is that current AI systems either reason explicitly through many steps (which is slow) or implicitly in ways that are hard to interpret (which limits control). ATLAS proposes a middle path where one discrete output token is enough to unlock the benefits of both approaches. As of the repository's release date, the code, trained model weights, and dataset are pending a company review and were not yet publicly available, only the paper and visual diagrams were included. Researchers working on multimodal AI (systems that process both images and text) or visual question answering would be the primary audience. No programming language is listed because no code has been released yet.