Kumar Tanmay

Background

About Kumar Tanmay

I am a Master's student in SEAS at Harvard and a Graduate Research Assistant at MIT EECS and Media Lab, specializing in the interpretability of transformer-based models focusing on the factual and in-context recall capabilities in multilingual settings. My research focuses on understanding the limitations and capabilities of contemporary AI systems, such as Large Language Models (LLMs), through mechanistic interpretability and model behavior control techniques like model editing and representation engineering. I am particularly interested in the mechanisms behind LLM behaviors and the democratization of these models by ensuring they are secure and unbiased without extensive fine-tuning. My work at Microsoft involved training mid-sized language models like Phi and Turing Bletchely in multilingual and multimodal settings respectively and integrating them into products like Bing Search and Microsoft Copilot. At Stanford's Sustainability Lab, I collaborated on diverse AI projects, ranging from self-supervised learning in geographical scenarios to poverty detection using reinforcement learning. I have also contributed to healthcare AI, developing an FDA-approved system for detecting abnormalities in chest X-rays. My work spans the intersection of responsible and safe AI, multilingualism, interpretability, and practical applications in diverse sectors.