gen‑ai.news
← Back
Video

World models that ignore human beliefs predict the wrong actions, new research shows

World models that ignore human beliefs predict the wrong actions, new research shows

World models - AI systems trained to simulate how environments unfold - have made significant progress in recent years, with projects like Sora and Genie demonstrating that models can learn surprisingly coherent physical dynamics from video. But a new line of research argues that these systems are missing something fundamental: the mental states of the people inside those environments. Without accounting for beliefs, desires, and intentions, the argument goes, a world model cannot reliably predict what a person will do next.

The new framework, called Mental World Modeling, proposes augmenting standard world models with explicit representations of mental variables alongside physical ones. The idea draws on longstanding concepts in cognitive science - particularly "theory of mind," the human capacity to attribute mental states to others and use them to anticipate behavior. Translating this into a computational framework means the model must track not just where objects are and how they move, but what an agent believes about those objects and what they are trying to achieve.

The research offers a striking empirical finding: language models that are relatively weak in terms of scale or general capability can outperform stronger models on action-prediction tasks when they are equipped with the Mental World Modeling approach. This suggests that the architecture and the information being modeled matter more than raw model size for this class of problems. It also points to a clear gap in how current world models are evaluated - benchmarks that focus on physical plausibility alone will not surface these shortcomings.

Perhaps the most useful insight from the work is its diagnosis of where the hardest problem lies. The biggest bottleneck is not representing mental states in isolation, but rather modeling how physical and mental states co-evolve - how a change in the environment updates someone's beliefs, and how those updated beliefs in turn shape their next action. That tight coupling between the physical and the mental is something current systems are poorly equipped to handle, and addressing it is likely to require both new training data and new modeling approaches tailored to that interaction.

Enjoy this story? Get the next one in your inbox.

Twice a week: the most important stories in generative image and video AI, distilled into a 2-minute read.

Free. Unsubscribe any time. No spam, ever.

Your next read

Google Opens Genie 4 World Model to Developers via Gemini API
Video

Google Opens Genie 4 World Model to Developers via Gemini API

Google has made Genie 4 available through the Gemini API, allowing developers to build interactive 720p environments with native audio and real-time controls. The world model can generate coherent, navigable spaces from prompts, moving it from a research curiosity toward a practical tool. Access is also rolling out to Google AI Ultra subscribers through the Project Genie interface.

Could This AI-Powered Operating System Actually Speed Up Your Video Production Flows?
Video

Could This AI-Powered Operating System Actually Speed Up Your Video Production Flows?

Lemonlight, a video production company with over 50,000 brand videos in its portfolio, has launched Hero - an AI-powered operating system designed to unify the full production pipeline in one platform. The tool aims to connect pre-production, production, post-production, and generative AI under a single roof. It targets production companies, small crews, and solo creators looking to reduce workflow fragmentation.