Researchers introduce Learning-to-Theorize, a new AI paradigm that builds explicit explanatory theories of the world from observations rather than simply predicting future states. The Neural Theorizer (NEO) model represents understanding as executable, compositional programs whose learned primitives can be recombined to explain novel phenomena, enabling explanation-driven generalization.
This research represents a fundamental shift in how AI systems approach world understanding, moving beyond the dominant paradigm of prediction-based world models. Rather than treating understanding as accurate future forecasting in latent space, the authors draw inspiration from cognitive science to frame it as theory construction—the ability to build internal causal models that explain observations and generalize to new contexts.
The NEO model instantiates this approach through a learned Language of Thought, where theories are represented as executable programs with composable primitives. This architectural choice enables systematic recombination of learned building blocks to explain novel phenomena, a capability that mirrors human reasoning more closely than standard neural networks. The emphasis on explainability and compositional structure addresses long-standing limitations in deep learning: brittleness to distribution shift and lack of interpretability.
For the broader AI community, this work challenges the current trajectory toward larger scale prediction models. It suggests that understanding may require explicit causal reasoning and modular structure rather than scaling up pattern matching. This has implications for AI safety and robustness—systems built on explicit theories may be more interpretable and controllable than end-to-end deep networks.
The immediate impact remains academic, focused on improving generalization and explainability in visual reasoning tasks. However, if theory-building approaches demonstrate superior performance on out-of-distribution problems, they could influence architecture design in robotics, scientific discovery, and AI agents. The research opens questions about integrating symbolic and neural computation at scale.
- →Learning-to-Theorize proposes understanding as explicit theory construction rather than future prediction, grounded in cognitive science.
- →The Neural Theorizer model uses executable programs with learned primitives to represent compositional theories of the world.
- →This approach enables explanation-driven generalization, improving performance on novel phenomena not seen during training.
- →The work challenges the current paradigm of prediction-based world models by emphasizing interpretability and causal structure.
- →Potential applications include more robust AI systems for robotics, scientific discovery, and agents requiring out-of-distribution reasoning.