Understanding the mathematical principles behind model representation capacity, training dynamics, and generalization.

Leveraging these principles to design better and more scalable architectures, optimizers, training/fine-tuning methods, and regularization techniques.

Aligning Large Language Models (LLMs), Vision-Language Models (VLMs), and their derivative Agents with specific human preferences and demands, with techniques like Reinforcement Learning from Human Feedback (RLHF) and Reinforcement Learning with Verifiable Reward (RLVR).

Exploring robust fine-tuning "recipes" within the RL framework to ensure that pre-trained capabilities are preserved while desired, human-aligned skills are effectively amplified.