Towards Zero-Shot Task Transfer with Neurosymbolic World Models

Imagine an artificial intelligence agent learning to navigate a complex world, much like a child learning to walk. In the field of machine learning, these agents often rely on "world models," which are internal mental maps that allow them to predict what will happen next if they take a specific action. By simulating these future scenarios in their minds, the agents can plan their moves without needing to physically try every possibility in the real world. This approach has become incredibly powerful, enabling computers to master games and control robots by learning from high-dimensional sensory data, such as streams of video images. However, a significant hurdle remains: these mental maps are usually built for a single, specific job. If the goal changes—say, from reaching a red door to reaching a blue one—the agent often has to relearn its entire understanding of the world, because its internal map is tightly woven with the specific reward it was chasing during training.

Researchers at KU Leuven in Belgium have developed a new approach to solve this rigidity, creating a system that can adapt to new goals instantly, without needing to interact with the environment again. Their work introduces a type of "neurosymbolic" world model, a hybrid system that combines the pattern-recognition strength of neural networks with the clear, logical structure of symbolic reasoning. Instead of learning a single, tangled representation of the world, these models learn to separate the raw sensory details of an environment from the specific, high-level facts that determine success. This separation allows the agent to keep its understanding of how the world moves and changes, while simply swapping out the rule for what counts as a "win." The result is an agent that can learn a complex environment once and then apply that knowledge to entirely new tasks, a capability known as zero-shot transfer, simply by redefining the goal in terms of the same underlying facts.

The core innovation lies in how the model processes information. Traditional world models learn a compressed, abstract version of the world that is difficult for humans to interpret and hard to repurpose. In contrast, the new neurosymbolic models force the system to identify and predict specific, structured properties of the state, such as the coordinates of an agent or the location of an object. These properties act as a bridge. The model uses its neural network to understand the messy, visual world and translate it into these clear, symbolic facts. Once the model has these facts, it uses a separate, logical component to predict rewards. Because the reward prediction depends only on these clear facts, researchers can simply replace the reward function with a new one that targets a different goal, and the model immediately understands how to plan for it. The agent does not need to see the new goal or try it out; it simply re-runs its mental simulations with the new rule, using the same internal map of how the world works.

To test this idea, the researchers trained their models on several distinct environments, including a grid-based navigation task where an agent must find a goal, a 3D version of the same task with continuous movement, and a puzzle game called Sokoban where boxes must be pushed to specific spots. In these training phases, the agents learned to navigate and achieve a specific objective, such as reaching a fixed goal location. The researchers then challenged the models with new tasks where the environment's rules of movement remained exactly the same, but the goals changed. In one scenario, the goal might move to a different room; in another, the agent might need to reach a specific configuration of objects rather than a single point. The models were able to adapt to these new objectives without any further interaction with the real environment. They achieved this through two methods: either by planning new paths using a search algorithm that explored the possibilities in their mind, or by refining their decision-making policy entirely within their internal simulations.

The results showed that this approach not only worked but did so more efficiently than standard methods. The neurosymbolic models learned the training tasks with fewer examples, suggesting that the structured way they process information helps them grasp the environment faster. More importantly, when faced with new tasks, they adapted instantly. While traditional models would require a period of retraining or a massive amount of new data to figure out the new goal, the neurosymbolic agents simply applied their existing knowledge of the world to the new objective. This was true even when the researchers provided only partial information about the symbolic facts during training, or when they provided no direct supervision for those facts at all, relying on the structure of the model itself to guide the learning.

This work demonstrates a significant step toward more flexible and reusable artificial intelligence. By decoupling the understanding of how the world works from the specific goals an agent is trying to achieve, the researchers have created a system that behaves more like a general learner than a specialized tool. The agent does not just memorize a path to a destination; it learns the terrain and the rules of movement, allowing it to navigate to any destination defined within that terrain. While the system still requires that the new goals be describable using the same set of high-level facts learned during training, it removes the need for costly retraining whenever a task changes. This suggests a future where intelligent systems can be deployed in dynamic environments, ready to tackle new challenges the moment they arise, simply by being told what to look for, rather than having to learn how to see it all over again.