Abstract
Transformer-based models, large language models (LLMs), and related modular architectures have demonstrated a remarkable ability to adapt to new tasks either in-context, by conditioning on a few demonstrations, through fine-tuning, by updating a small subset of parameters, or through selective routing to specialized components. However, the mechanisms that enable such sample-efficient adaptation, and principled methods for exploiting unlabeled data, personalization, and non-stationary environments, remain only partially understood. This dissertation develops a unified theoretical and algorithmic framework for sample-efficient adaptation in transformer-based and modular learning systems.
The first part studies how transformers implement learning-to-optimize (L2O) algorithms for in-context sparse recovery tasks. For LASSO-type problems, we show that a K-layer decoder-only transformer can be explicitly constructed to realize a LISTA-type algorithm with an error that decreases linearly with K. Unlike classical LISTA variants that must be trained and tested under the same measurement matrix, the resulting ``LISTA with Varying Measurements'' (LISTA-VM) implemented by the transformer provably generalizes across different measurement matrices and accommodates a varying number of demonstrations. Experiments with both small transformers and GPT-2 validate the theory and highlight the robustness and efficiency of the learned in-context procedure.
The second part introduces an augmented in-context learning framework that combines a small set of labeled examples with a block of unlabeled inputs in the same prompt. For multi-class linear classification, we demonstrate that, under chain-of-thought prompting, a multi-layer encoder-based transformer can emulate an expectation-maximization (EM) style algorithm, leveraging both labeled and unlabeled data to achieve provable gains in in-context accuracy. When trained with teacher forcing, the transformer parameters converge to the desired solution at a linear rate, and the resulting classifier enjoys excess risk bounds that strictly improve over purely supervised in-context learning.
The third part focuses on personalized reinforcement learning from human feedback (RLHF). We propose a shared low-rank adaptation approach in which Low-Rank Adaptation (LoRA) is applied in a joint parameter space of all user-specific reward functions. This construction exploits shared low-rank structure while allowing individual-specific adaptations, leading to sample-complexity guarantees for recovering both common and personalized components of human preferences. Empirical results on real-world preference datasets demonstrate that the resulting P-ShareLoRA algorithms improve preference-model accuracy and alignment quality compared with standard global or purely local LoRA baselines.
The fourth part studies continual learning under concept drift through a mixture-of-experts (MoE) framework. We model streaming data as partially observed samples drawn from a union of low-dimensional subspaces whose structure evolves over time. In the stationary setting, we show that the underlying low-rank experts are identifiable under partial observations, yielding correct expert selection and exact completion. In the online setting, we show that an MoE-style continual learner can track drifting subspaces with provable contraction of expert error over time while updating only the routed expert, thereby mitigating cross-concept interference. Synthetic repeated-drift experiments further validate the theory and demonstrate stable adaptation relative to single-subspace online baselines.
Collectively, these four parts provide a principled foundation for designing adaptive learning systems that efficiently leverage limited labeled data, auxiliary unlabeled samples, heterogeneous human feedback, and evolving data streams.