Go to main content
Formats
Format
BibTeX
MARCXML
TextMARC
MARC
DataCite
DublinCore
EndNote
NLM
RefWorks
RIS

Files

Abstract

The rapid advancement of large language models (LLMs) offers transformative potential across a wide range of applications, but concurrently raises critical safety concerns, under- scoring the importance of aligning AI systems with human values. This thesis investigates a series of methodologies designed to enhance AI alignment through novel optimization strate- gies that improve performance, robustness, and trustworthiness. By addressing limitations intraditionalreinforcementlearningfromhumanfeedback(RLHF)pipelines, ourworkintro- duces streamlined alternatives and complementary techniques to optimize alignment while maintaining computational efficiency and effectiveness. In Chapter 2, we propose f-DPO (generalized Direct Preference Optimization), which extends the DPO framework by incorporating diverse divergence constraints, such as Jensen- Shannon and forward KL divergences. Through an analytical exploration of the Karush- Kuhn-Tucker conditions, we derive simplified relationships between reward functions and optimal policies under these divergences. Empirical evaluations demonstrate that f-DPO balances alignment performance and generative diversity more effectively than RLHF-based ProximalPolicyOptimization(PPO).Italsoachieveslowerexpectedcalibrationerror(ECE) while providing practical benefits, such as improved divergence efficiency. Chapter 3 extends alignment optimization by addressing the limitations of single- sample preference comparison. We introduce Multi-sample Direct Preference Optimization (mDPO) and Multi-sample Identity Preference Optimization (mIPO), frameworks that op- timize group-wise characteristics to improve distributional properties such as diversity and bias reduction. These approaches significantly enhance generative diversity in LLMs and mitigate demographic biases in diffusion models. Moreover, multi-sample methods exhibit robustness to noisy human-labeled preference data, making them particularly effective for fine-tuning in real-world scenarios where label quality may be imperfect. More importantly, it offers more controllability for improving the alignment of generative models over f-dpo. In Chapter 4, we move beyond the traditional supervised learning in chapters 2 and 3, and address a critical challenge in reward modeling for online RLHF: spurious correlations that can distort alignment objectives. We introduce a causal reward modeling framework that integrates causal inference techniques to mitigate these biases. By enforcing counter- factual invariance, this approach ensures reward predictions remain unaffected by irrelevant or confounding variables. Experiments on synthetic and real-world datasets demonstrate significant improvements in addressing biases such as length preference, sycophancy, and concept biases, ultimately enhancing the fairness and reliability of alignment. Together, these contributions advance the theoretical and practical foundations of AI alignment. Byofferingscalableandrobustmethodologies, thisthesisbridgesthegapbetween current capabilities and the long-term goal of developing safe and trustworthy AI systems. The proposed approaches not only improve alignment workflows but also address critical shortcomings in existing pipelines, paving the way for more reliable, diverse, and human- aligned AI systems.

Details

Actions

PDF

from
to
Export
Download Full History