Abstract:
Cognitive science is good at modeling how minds update beliefs within a representational vocabulary that is already in place. It is much weaker at modeling how that vocabulary itself changes. Symbolic systems hand-design their vocabularies and inherit no process by which experience could have produced them. Connectionist systems learn representations, but the usual analytic lens is convergence to a final code rather than the accumulation of reusable structure that reshapes later learning. Bayesian systems specify a hypothesis space and perform inference within it, leaving open how that space was carved up in the first place and how it should be revised as the agent encounters new tasks. The shared failure is not the absence of representation, but the absence of representational dynamics: a process-level account of how vocabulary is built, maintained, and shared, and of how each of these changes what becomes easy to learn next.
This thesis argues that these dynamics become tractable once a particular object is named: a generative library of reusable, executable, compositional, and inspectable building blocks over which hypotheses are composed. The central claim is that intelligence is the resource-rational lifecycle management of the library. The claim has bite because the properties that make something a library impose three demands on any adequate account. First, library entries accumulate. What has already been compressed determines what can be represented cheaply next, so construction must be modeled as path-dependent rather than as convergence to a fixed code. Second, a library is finite, and the value of an entry depends on uses it has not yet seen. Curation therefore cannot be reduced to local utility; it requires a criterion defined over the agent's own representational state and the future computations that state makes possible. Third, library entries are inspectable and shareable. Communication is therefore not merely the transmission of information, but the alignment of representational resources; an explanation succeeds only when it can be compiled into action by a listener whose library may differ from the speaker's.
These demands are not three separable problems. A library that is cheap to build but undisciplined in what it keeps will drift. A library that is well curated but opaque will not survive contact with another agent. A library that is communicable but built without regard to learning history will encode the wrong things. The lifecycle is the constraint that construction, curation, and communication must be optimized together.
The thesis develops this argument through a common methodology: latent inference over generative models of agents, with compression and two-part description length as the cross-stage criterion of representational value. The empirical chapters span controlled human sequence learning, large-scale educational data, simulated agents driven by competing intrinsic motivations, and human studies of explanation under uncertainty. Each chapter is built so that its model and its evidence speak to one stage of the lifecycle using the same yardstick as the others. The contribution is therefore not a model of any single cognitive capacity, but a unified account of representational dynamics in which construction, curation, and communication are stages of the same process and become optimizable only together.
generative library