Lingjing Kong

 Traditional machine learning is largely organized around passive prediction: given observations from a training distribution, a model learns to predict labels, tokens, or future observations. Modern large models have changed the stakes of this paradigm. They now read, generate, reason, and assist decisions, yet they can remain brittle, opaque, difficult to control, and unreliable outside their training conditions. This thesis argues that trustworthy AI requires more than larger predictors. It requires representations that expose reusable mechanisms, preserve them across environments, and reveal when new evidence is needed.

The work is organized around four data regimes. In learning from multiple distributions, domain variation separates invariant and changing latent factors, enabling transfer without target-domain labels. In learning with unstructured distributions, masked prediction turns internal dependence into a signal for recovering hierarchical concepts. In generalization beyond training distributions, a novel prediction or generation is justified only when it can be decomposed into mechanisms already supported by training data. In learning through feedback and interaction, reward, multitask experience, and model verification identify reusable reasoning modules, compose dynamics across tasks, and direct new data collection toward weak parts of a world model.

Together, these regimes support the thesis that causal principles such as minimality, hierarchy, modularity, and structured feedback can turn representation learning from black-box fitting into mechanism discovery and support expansion. The dissertation develops this claim as a self-contained argument: each technical chapter states the available data signal, formalizes the latent structure, derives what can be recovered or generalized, and tests the resulting mechanism. The final chapter distinguishes the demonstrated results from the larger goal of trustworthy AI systems whose transfer, interpretation, control, and self-improvement can be diagnosed rather than merely hoped for.