AI Agents Learn to Test Their Own Hypotheses About How the World Works
Most AI systems today are like really good students who can ace a specific test but struggle when the subject changes. CDS PhD student Anthony GX-Chen, CDS-associated Professor Rob Fergus, and Kenneth Marino of Google DeepMind and The University of Utah have built AI agents that work more like scientists — they form hypotheses about how their environment works, test those hypotheses, and build up knowledge they can use in new situations.
GX-Chen and his team’s approach is detailed in “Efficient Exploration and Discriminative World Model Learning with an Object-Centric Abstraction,” which was presented at ICLR 2025. It tackles a basic problem in AI: most systems only learn to complete specific tasks rather than understand the rules governing their world. “Can we design an AI so that when we drop it into some environment, it can just go about discovering how that environment works, and come back to us with how it works,” GX-Chen said.
The key is how these agents think about the world. Instead of trying to predict everything that might happen next — which gets overwhelming fast — the agents focus on objects and test simple hypotheses: “If I try to change this object in this way, will it work?”
Think of it like a curious child in a new playground, systematically testing whether the swing moves when pushed, whether the slide is slippery, or whether the seesaw balances with different weights. The AI agent forms hypotheses about these relationships and tests them one by one.
This hypothesis-testing approach proved remarkably effective. In challenging video game environments where other AI systems failed after millions of attempts, these agents succeeded using far fewer tries. They could even transfer what they learned to completely new environments.
What makes this approach particularly compelling is its interpretability. Unlike black-box AI systems that provide little insight into their reasoning, GX-Chen’s team’s method produces clear, readable models of how environments function. “It gives you this highly interpretable way of interpreting how this environment works, which the agent discovers automatically,” he explained.
The applications go beyond gaming. GX-Chen discussed how a general method for discovering sequential outcomes could be relevant for a wide variety of settings, such as in medical treatments. Here, doctors “get to try a variety of different drugs and treatment and chemotherapy” and “the orders in which you try them is also important.” Each treatment is essentially a hypothesis test: will this drug help this patient in their current state?
The research compliments the growing capabilities of foundation models — the large AI systems behind tools like ChatGPT — which have become increasingly adept at extracting meaningful abstractions from complex data. By leveraging these models to identify objects and their properties, this new system can focus on learning the relationships between them rather than starting from scratch with raw sensory data.
The team’s work represents a step toward AI systems that can operate more autonomously in complex, real-world environments. Rather than having to specify domain specific knowledge for each new task, these systems could potentially adapt to novel situations by building their own understanding of the underlying mechanics at play.
As AI becomes more common in complex, real-world situations, this kind of scientific thinking could make the difference between systems that only work in controlled lab settings and ones that can handle the messy, unpredictable world outside.
By Stephen Thomas
