Redefining Context for Powerful Test-Time Adaptation Using Unlabeled Data
Speaker
Sharut Gupta
MIT CSAIL
Host
Yifei Wang
MIT CSAIL
Abstract:
Foundation models, while powerful, often struggle under distribution shifts in unfamiliar domains, typically requiring costly data collection and retraining to maintain performance. Test-Time Adaptation (TTA) has emerged as a promising approach to address these limitations, enabling models to adapt dynamically to new target domains at test time. In this talk, I will present TTA approaches by rethinking the notion of “context”—an abstract concept drawn from in-context learning—to address two fundamental challenges: improving out-of-distribution generalization and aligning representations with varying task-specific inductive biases, such as fairness constraints. Specifically, we explore two ways of leveraging unsupervised in-context learning, allowing models to use unlabeled data to adapt their behavior flexibly. First, we will demonstrate how using unlabeled domain data as context can align models with diverse distributions, enhancing their robustness in changing environments. Next, we will extend this idea to further improve this alignment by enforcing task-specific inductive priors. Together, these approaches showcase the potential of unsupervised, context-driven TTA to address key challenges of current-generation foundation models. Finally, we will explore the broader implications of this context-driven perspective for building world models, planning, and robust decision-making.
Bio:
Sharut Gupta is a third-year PhD candidate in Electrical Engineering and Computer Science (EECS) at MIT, advised by Prof. Stefanie Jegelka. Her research interests focus on multi-modal representation learning, robustness, and out-of-distribution generalization. She received her Bachelor’s and Master’s (Dual) degrees from the Indian Institute of Technology Delhi (IIT Delhi), where she completed her thesis research with Prof. Yoshua Bengio on "A Causal Perspective on Efficient Distributed Systems”. Sharut is a recipient of the MIT Presidential Fellowship and has completed research internships at FAIR (Meta AI) and Google DeepMind.
Foundation models, while powerful, often struggle under distribution shifts in unfamiliar domains, typically requiring costly data collection and retraining to maintain performance. Test-Time Adaptation (TTA) has emerged as a promising approach to address these limitations, enabling models to adapt dynamically to new target domains at test time. In this talk, I will present TTA approaches by rethinking the notion of “context”—an abstract concept drawn from in-context learning—to address two fundamental challenges: improving out-of-distribution generalization and aligning representations with varying task-specific inductive biases, such as fairness constraints. Specifically, we explore two ways of leveraging unsupervised in-context learning, allowing models to use unlabeled data to adapt their behavior flexibly. First, we will demonstrate how using unlabeled domain data as context can align models with diverse distributions, enhancing their robustness in changing environments. Next, we will extend this idea to further improve this alignment by enforcing task-specific inductive priors. Together, these approaches showcase the potential of unsupervised, context-driven TTA to address key challenges of current-generation foundation models. Finally, we will explore the broader implications of this context-driven perspective for building world models, planning, and robust decision-making.
Bio:
Sharut Gupta is a third-year PhD candidate in Electrical Engineering and Computer Science (EECS) at MIT, advised by Prof. Stefanie Jegelka. Her research interests focus on multi-modal representation learning, robustness, and out-of-distribution generalization. She received her Bachelor’s and Master’s (Dual) degrees from the Indian Institute of Technology Delhi (IIT Delhi), where she completed her thesis research with Prof. Yoshua Bengio on "A Causal Perspective on Efficient Distributed Systems”. Sharut is a recipient of the MIT Presidential Fellowship and has completed research internships at FAIR (Meta AI) and Google DeepMind.