The project on GitHub presents the code base for the paper titled *Visual Generation Unlocks Human-Like Reasoning through Multimodal World Models*. It explores how unified multimodal models (UMMs) enhance reasoning by integrating visual generation into world modeling. Humans use both verbal and visual channels to create mental models that support reasoning, planning, and decision-making. In contrast, many current large language models (LLMs) mainly rely on verbal reasoning. By employing UMMs, this research shifts the focus to visual generation, claiming it leads to improved reasoning in tasks grounded in the physical world. Key findings include the formalization of atomic capabilities of world models and the introduction of the visual superiority hypothesis, which suggests that visual world modeling provides richer information than purely verbal approaches. The authors designed a new evaluation suite called VisWorld-Eval, which features seven tasks that assess specific capabilities of world models. These tasks range from paper folding to real-world spatial reasoning, with a mix of synthetic and real-world domains. The suite is intended to rigorously evaluate multimodal reasoning. Results from controlled experiments using the BAGEL framework indicate that interleaving visual and verbal chain-of-thought reasoning outperforms conventional verbal-only approaches. For instance, the Gemini 3 Flash model achieved an overall accuracy of 60.5% across five tasks in the VisWorld-Eval suite, with particularly strong performance in multi-hop manipulation. In contrast, models like GPT 5.1 and BAGEL-7B-MoT showed significantly lower accuracy, emphasizing the advantages of integrating visual data in reasoning tasks. The project invites feedback and collaboration, encouraging users to cite the paper if they find the work useful.