Kartik Sharma

I am a final-year Ph.D. candidate at Georgia Institute of Technology, being advised by Prof. Srijan Kumar and Dr. Rakshit Trivedi. My research is broadly focused on safe and controllable adaptation of machine learning models, where my goal is to build machine learning (🤖) models that are robust to perturbations (🦺) and can be easily controlled (🎛️) by users.

My motivation stems from a fundamental challenge: these models are inherently probabilistic and non-deterministic, which can lead to behavior that deviates from users’ expectations and hinders widespread adoption. I address this by ensuring models remain stable under reasonable perturbations while changing precisely as directed through explicit controls. My work develops methods for controllable generation (test-time steering of diffusion models and LLMs), human-centered adaptation (modeling diversity in behavior and intentions), and adversarial robustness (defending against attacks while maintaining utility). To this end, I have worked in relational (e.g., graph), ordered (e.g., language, decision process), and dynamic (e.g., dynamic graph, multi-agent) environments.

Looking forward, I’m working toward control mechanisms that manage the trade-off between utility and safety—analogous to traction control in vehicles, which enables performance while preventing catastrophic outcomes. Specifically, I aim to: (1) design user controls for monitoring and intervention in multi-agent systems where behaviors emerge from agent interactions, (2) develop methods to handle complex user intentions including modeling ambiguity and contradictions in natural language interfaces, (3) create online adaptation mechanisms that account for evolving user intentions to minimize regret from over-optimization, and (4) provide worst-case guarantees that ensure true satisfaction of controls beyond surface-level compliance.

I have had the pleasure of gaining valuable industrial experience throughout my Ph.D.: working with Amazon RISC (Regulatory Intelligence, Safety and Compliance) Science team in 2025, Microsoft Research for Industry group in 2024, and Visa Research in 2024. Before this, I graduated from IIT Delhi in 2021, where I was fortunate to be advised by Prof. Sayan Ranu and mentored by Prof. Sourav Medya.