For decades, the "Holy Grail" of artificial intelligence has been compositional generalization: the ability to take familiar building blocks—like the concepts "red," "blue," "car," and "bike"—and understand a novel combination never seen during training, such as a "blue car."
The traditional debate centered on architecture. Many argued that standard neural networks are "monolithic" and cannot generalize this way unless we explicitly build modular, symbolic structures into them. However, recent research suggests that the secret to compositionality lies not in the model's architecture, but in the structure of the training data.
The Geometry of Compositional Support
To understand why a model generalizes to new combinations, we have to look at the "support" of the training distribution. Matthias Bethge, Wieland Brendel, and their colleagues propose a mathematical framework where compositionality is a property of the data-generating process.
They argue that a model can learn the identity of an individual component (like "redness") even through complex interactions (like occlusions) if the training data provides sufficient variety. If the model sees "red" in enough different contexts, it can mathematically track the influence of that specific component. Compositional generalization is guaranteed when the training distribution covers the full range of states for every individual component, even if specific combinations are missing.
Connected Support as the "Glue"
If we view tasks as combinations of modules, how does a network know how to "stitch" them together? Simon Schug, Yassir Akram, and their team demonstrate that for a modular system to generalize, the training data must have connected support.
Imagine you have two modules, A and B. If you only ever see A with C, and B with D, the model might learn internal representations for A and B that are incompatible—they exist in different "coordinate spaces." To generalize to the combination (A, B), the training data must include overlapping sets of components (e.g., seeing A with C, and B with C). This overlap acts as the "glue" that aligns the internal representations into a unified space, preventing the model from learning isolated, incompatible solutions.
Scaling into Systematicity
A common critique is that while small models might need these strict conditions, large-scale models seem to "just work." Florian Redhardt and Simon Schug show that scaling actually enables this process.
They provide evidence that standard Multi-Layer Perceptrons (MLPs) can efficiently represent compositional structures if they are large enough and the data is sufficiently connected. The key mechanism is the formation of linear internal representations. When a model is scaled up, it becomes more likely to find these linear structures, which allow it to "interpolate" between known combinations to solve unknown ones. Successful generalization is fundamentally tied to the model's ability to bridge gaps in the training distribution using these internal coordinates.
From Linear to Exponential Generalization
The power of this "connected support" becomes most apparent in multi-step reasoning. Amirhesam Abedsoltan, Jingzhao Zhang, and their co-authors explored whether a model trained on $D$ tasks can generalize to $D^T$ tasks.
By framing tasks as autoregressive compositions (like a Chain-of-Thought), they found that a model can solve an exponential number of unseen tasks if the training data demonstrates every sub-operation across all possible positions in a sequence. If a specific sub-task is missing from a particular "stage" of the process during training, the model fails. This reinforces the idea that structural diversity—ensuring no component is isolated from any position or context—is the primary driver of compositional intelligence.