The compositionality continuum as a principle for studying the neural basis of intelligence

Compositionality has long been regarded as a uniquely human capacity and a cornerstone of general intelligence. It is most clearly exemplified in natural language, where a finite set of words is combined through grammatical rules to generate an unbounded number of sentences. Recent advances across disciplines challenge the view that compositionality is exclusive to human cognition and question whether an explicit symbolic structure is required to implement it. Large language models, for example, achieve impressive compositional abilities through scale alone. Neuroscience has also revealed that animals use compositional neural codes when applying knowledge to novel situations, while theory and modeling are beginning to clarify how neural networks without explicit symbolic structures implement compositional computations. High-density neural recordings from animals performing compositional tasks, combined with reverse engineering of neural network models, now enable us to test fundamental questions, such as how biological brains implement compositional solutions and when compositional behavior can emerge from scale rather than from explicit compositional mechanisms. We propose that compositional mechanisms exist along a continuum defined by the expressivity of computation-specific building blocks and the complexity of systematic recombination rules. Comparing implementations across biological and artificial systems will determine whether general intelligence requires explicit compositionality or can emerge from scaled mechanisms alone.