A neural mechanism for compositional generalization of structure in humans

A human ability to adapt to the dynamics of novel environments relies on abstracting and generalizing from past experiences. Previous research has focused on how humans generalize from isolated sequential processes, yet we know little about mechanisms that enable adaptation to more complex dynamics, including those that govern much everyday experience. Here, using a novel sequence learning task based on graph factorization, coupled with simultaneous magnetoencephalography (MEG) recordings, we asked how reuse of experiential “building blocks” enables inference and generalization. Behavioral evidence was consistent with participants decomposing task experience into subprocesses, involving abstracting dynamical subprocess structures away from their sensory specifics and transferring these to a new task environment. Neurally this transfer was underpinned by a representational alignment of abstract subprocesses across task phases, evident in an enhanced neural similarity among stimuli that adhered to the same subprocesses, a temporally evolving mapping between predictive representations of subprocesses and a generalization of the dynamic roles that stimuli occupied within graph structures. Decoding strength for dynamical role representations predicted behavioral success in transfer of subprocess knowledge, consistent with a role in supporting behavioral adaptation in new environments. Our findings reveal neural dynamics that support compositional generalization, consistent with a structural scaffolding mechanism that facilitates efficient adaptation within new contexts.