Research

Human–LLM representational alignment

I'm researching representational alignment between humans and LLMs, with the goal of making your LLM an extension of your own mind.

Every entity, human or machine, represents the world and think of concepts differently. This makes conversation imprecise. In the case of human-LLM communication, this leads to the model misrepresenting the user's intent and goals for a task. My work attempts to automate the process of aligning the model with the user, at a fundamental logical level, such that whether we are querying information, learning about a topic, or performing a task, the model reasons about the scenario in the same way you would, without explicitly having to state your own logic every time.

In short, the goal is to automate the process of making the model see your point of view. Why? Because it makes the LLM a better assistant to the human.