Title: Adapting Models to User Intent for Safe and Controllable Generation

Date: Tuesday, April 21st , 2026

Time: 2:00 p.m. - 4:00 p.m. ET

Location (in-person + virtual): Coda C1303 Glenwood [Virtual Link]

Kartik Sharma

Ph.D. candidate

School of Computational Science and Engineering

College of Computing

Georgia Institute of Technology 

Committee:

Dr. Srijan Kumar (Advisor, Computational Science and Engineering, Georgia Institute of Technology)

Dr. Rakshit Trivedi (Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology)

Dr. Chao Zhang (Computational Science and Engineering, Georgia Institute of Technology)

Dr. Bo Dai (Computational Science and Engineering, Georgia Institute of Technology)

Dr. Polo Chao (Computational Science and Engineering, Georgia Institute of Technology)

Abstract:

As artificial intelligence (AI) is increasingly deployed in high-stakes, open-world environments, the demand for systems that are both highly capable and strictly trustworthy has become paramount. Modern generative models require high test-time plasticity to remain responsive to complex and dynamic human intentions. However, this same architectural flexibility inadvertently leaves them vulnerable to manipulation by bad actors and unintended changes. To ensure AI remains an empowering tool rather than a fragile liability, this dissertation presents a comprehensive framework for adapting models safely to user intent. First, this work introduces interpretable test-time mechanisms that empower users to precisely steer model behavior. These interventions are designed to enforce hard constraints in graph diffusion models and guide large language and imitation learning models toward exemplary and directed behavior descriptions. However, because unconstrained plasticity inherently alters a model’s risk profile, the research then systematically diagnoses the vulnerabilities in the systems, exposing vulnerabilities in dynamic graph models and multimodal language models. To safely support user-driven controllability without succumbing to these adversarial risks, the final phase demonstrates how structural adaptivity of system prompts can protect a language model’s core without compromising its benign behavior. Ultimately, this research works towards resolving the modern stability-plasticity dilemma, transforming vulnerable networks into deeply steerable and secure extensions of human intent.