ENACT evaluates embodied cognition through egocentric interaction modeling, focusing on how well models understand the world from a first-person perspective. The project features a streamlined dataset pipeline, making it easier to work with. Users can explore the dataset through a dedicated viewer and compare model performances on a leaderboard. This leaderboard distinguishes between proprietary and open-weight models, allowing for direct comparison with human performance in tasks involving forward and inverse world modeling. Key findings from the research reveal that models perform better on inverse tasks compared to forward tasks, indicating a stronger capacity for language-based reasoning over action-oriented visual understanding. The study highlights limitations in spatial memory when models face partial visibility, and emphasizes that current Vision-Language Models (VLMs) fall short of human-level performance in real-world scenarios involving mobile manipulation. In probing tasks, the models showed insensitivity to variations in rendering, suggesting that their weaknesses lie in multi-step reasoning rather than image quality. The analysis also points to a bias in VLMs toward human-centric perspectives, as evidenced by performance drops when models encounter non-standard viewpoints. This reliance on human-like visual cues restricts their adaptability to robots with different visual systems. Interestingly, the models demonstrated robustness against variations in robot appearance but exhibited a pronounced right-handed bias, mirroring human handedness patterns. This research offers a nuanced understanding of the challenges and limitations of current VLMs in embodying cognition effectively.