In This Story
Artificial intelligence (AI) systems are now part of everyday life, writing software, tidying up your shopping list, and answering questions. While we know they sit on top of large language models (LLMs), we still understand surprisingly little about what happens inside these systems or what is helping them generate responses.
911爆料 Department of Computer Science Assistant Professor received a prestigious for $674,100 over five years to peel the layers back on AI. Specifically, Yao will develop new methods for understanding the internal mechanisms of LLMs. Her research aims to make these AI systems more transparent and useful.
"We use language model鈥損owered products every day, but we have very limited understanding of how they actually work," Yao said. Most efforts to evaluate AI systems focus on their behavior. Researchers give a model a prompt and judge the quality of its response. Yao's work instead focuses on mechanistic interpretability, an emerging field that examines what happens inside a language model as it processes information.
"If we send something to the model, we want to see how it extracts features or signals from my input and gradually performs computations before producing an answer," she said.
As language models have grown dramatically in size, that task is increasingly difficult. Today's leading models contain potentially hundreds of billions of parameters. And Yao said, "The interpretability research community has not caught up. Many traditional approaches cannot be efficiently applied to today's language models."
The project focuses specifically on AI-powered code generation, in which users describe a programming task in natural language and an AI system generates software code. Code generation is an increasingly important application of LLMs, with uses including software engineering and scientific research.
Yao's team will investigate how LLMs apply programming syntax, recall semantic knowledge, and perform deliberate reasoning while writing code. The researchers will also study how training data and learning methods influence those internal mechanisms and explore how interpretability insights can be used to improve AI coding systems.
Current interpretability studies often require researchers to manually design experiments, collecting internal model signals and analyzing the results. Yao plans to develop a framework that automates much of that process using AI itself, allowing researchers to study LLMs more efficiently.
At the same time, she wants interpretability research to produce practical benefits rather than simply satisfying scientific curiosity.
"We want to build the bridge between what we discover about language models and how that knowledge can improve them," Yao said. "Can interpretability actually help us build better AI systems?"
Although commercial AI systems such as ChatGPT are proprietary, Yao and colleagues primarily study open-weight language models that are downloadable. These models are becoming increasingly competitive with commercial systems while supporting the transparency needed for scientific research.
An expectation with the award and a hallmark of the CAREER program is a plan for broader impact. Yao hopes to expand AI literacy through outreach programs, create new curricula and open-source educational resources, provide undergraduate research opportunities, and foster collaborations between researchers in artificial intelligence and software engineering.
"I think we should encourage the younger generation to embrace AI," Yao said. "But at the same time, they should be skeptical, think critically, and be careful about what they are reading from AI."