We delve into the high-dimensional latent space of LLMs to uncover how knowledge is encoded. Our research focuses on mechanistic interpretability to decode the 'inner circuits' of reasoning, enabling precise behavior steering and machine unlearning.

This research focuses on protecting AI models from manipulation. It explores ways to prevent attacks that exploit weaknesses, ensuring AI makes reliable and secure decisions.

We focus on improving AI’s ability to generate natural content while developing robust methods to detect AI-written text and deepfake speech, ensuring authenticity and preventing misuse.

Explores making AI decisions more transparent while addressing vulnerabilities in explanations. It aims to ensure AI insights are trustworthy, resistant to manipulation, and easy for end users to understand.

We apply our developed methodology in various critical NLP applications, from healthcare to education and social media analysis.