UC San Diego

Multimodal large language models (MLLMs) have demonstrated impressive capabilities across vision-language tasks, yet fundamental challenges remain in understanding their internal multimodal processing, aligning their behavior with human preferences, and enabling faithful reasoning. This dissertation addresses these challenges through three lines of work spanning multimodal understanding, alignment, and reasoning. We begin with multimodal understanding, where we develop an information-theoretic framework that traces how visual information is encoded, compressed, and shaped by textual instructions across MLLM layers. Building on these insights, we introduce CoMMIT, a coordinated instruction tuning approach that resolves the learning imbalance between visual encoders and the backbone LLM, and MDGD, which preserves visual knowledge during fine-tuning by decoupling modality-specific gradient updates. We then turn to multimodal alignment, proposing OCEAN, which models chain-of-thought reasoning as a Markov decision process and leverages knowledge graph exploration for offline evaluation and alignment, and IRPO, which extends preference optimization to in-context ranking lists. Finally, we advance multimodal reasoning through DeCoT, which applies causal intervention to debias chain-of-thought reasoning, and CTRLS, which formulates reasoning as latent state transitions with distributional reinforcement learning for structured exploration of diverse reasoning trajectories.