Aditi Raghunathan

Aditi Raghunathan is an Assistant Professor in the Computer Science Department at Carnegie Mellon University, with affiliations to the Machine Learning Department. She holds a PhD from Stanford University (2021), advised by Percy Liang, and a B.Tech in Computer Science from IIT Madras (2016). Her research focuses on advancing the scientific understanding of frontier models, addressing their reliability, safety, and unlearning challenges.

• Education
  • PhD, Stanford University (2021)
  • B.Tech, IIT Madras (2016)
• Research Themes
  • Overtraining effects in LLMs
  • Memorization sinks for unlearning
  • Algorithmic creativity beyond next-token prediction
  • Scalable frameworks for AI safety

Her work at ICML 2025 revealed critical limitations in LLM pre-training, including catastrophic overtraining and imperfect unlearning due to entangled memorization-generalization circuits. She has developed seed-conditioning and multi-token learning to enhance structured diversity and creativity in models.

Scientific Recognition includes:

• NSF CAREER Award
• Arthur Samuel Best Thesis Award at Stanford
• Google Research Scholar
• Forbes 30 Under 30 in Science

She supervises PhD and Master’s students such as Jacob Mitchell Springer, Christina Baek, and Taeyoun Kim, and actively collaborates with CMU graduate students. Her teaching includes courses on Trustworthy AI and Graduate Artificial Intelligence.