My Master's thesis explores how language models can be used to solve multi-step reasoning problems via task decomposition as the complexity of the tasks increases. My interests at the moment include:

• Knowledge representation in deep models (e.g., latent embedding spaces), particularly using formal languages.
• Deep architectures and methods for solving complex reasoning problems (e.g., mathematics, programming tasks, planning, etc.).
• Reinforcement and imitation learning, particularly meta-learning the RL/IL process outside of classical algorithmic approaches.

My long term ambition is to design models capable of not just imitating training data, but of discovering and displaying new knowledge. Today, even the largest models struggle to extrapolate complex reasoning patterns in data and exhibit valuable novel behavior.