Research

My current research focuses on large language models (LLMs) and vision-language models (VLMs), with particular emphasis on efficient adaptation, safety, robustness, distributed learning, and learning from demonstrations or feedback.

Current Research

Parameter-Efficient Fine-Tuning and Prompt Tuning

Foundation models are powerful but expensive to adapt to new tasks and domains. We develop parameter-efficient methods for LLMs and VLMs that update only a small fraction of model parameters through low-rank factors, prompts, adapters, or related structured representations. Our work studies how to make adaptation more computationally efficient while preserving accuracy and generalization, with particular interest in few-shot learning, multimodal adaptation, low-rank structure, and model compression.

Representative papers: MMLoP: Multi-Modal Low-Rank Prompting for Efficient Vision-Language Adaptation; Decentralized Low-Rank Fine-Tuning of Large Language Models; INTERLACE: Interleaved Layer Pruning and Efficient Adaptation in Large Vision-Language Models.

Illustration of safe behavior in LLMs and VLMs under benign and harmful inputs.

Safety of Foundation Models and Vision-Language Models

As LLMs and VLMs are deployed in increasingly consequential settings, they must remain useful while resisting harmful or adversarial inputs. We study the mechanisms behind jailbreaks, overrefusal, and failures caused by noisy or adversarial multimodal prompts, and develop defenses that improve safety without unnecessarily degrading performance on benign inputs. A particular focus is medical VLMs, where the tradeoff between under-defense and over-defense is especially important.

Representative papers: Deactivating Refusal Triggers: Understanding and Mitigating Overrefusal in Safety Alignment; Enhancing the Safety of Medical Vision-Language Models by Synthetic Demonstrations; Benchmarking and Mitigating the Impact of Noisy User Prompts in Medical VLMs via Cross-Modal Reflection.

Illustration of federated learning across multiple devices and a central server.

Federated and Distributed Learning

Federated learning enables models to be trained across distributed, privacy-sensitive data without moving raw data to a central location. Our recent work brings these ideas to foundation models, where communication and memory costs make efficient adaptation particularly important. We study communication-efficient and personalized fine-tuning of LLMs and VLMs, decentralized low-rank adaptation, and learning under heterogeneous clients, while continuing to investigate fundamental issues such as stragglers, robustness, and data heterogeneity.

Representative papers: Communication-Efficient and Tensorized Federated Fine-Tuning of Large Language Models; pFedMMA: Personalized Federated Fine-Tuning with Multi-Modal Adapter for Vision-Language Models; Robust Decentralized Learning with Local Updates and Gradient Tracking.

Illustration of adversarial robustness and perturbations in representation space.

Robust Machine Learning and Foundation Models

We study machine-learning systems that remain reliable under adversarial perturbations, distribution shifts, and other forms of model or data uncertainty. Our earlier work characterized fundamental limits of adversarial robustness and the tradeoff between standard and robust performance. More recently, we have extended these questions to foundation models and parameter-efficient adaptation, including robust low-rank fine-tuning of VLMs and optimization methods that explicitly account for conflicts between clean and adversarial objectives.

Representative papers: Few-Shot Adversarial Low-Rank Fine-Tuning of Vision-Language Models; Conflict-Aware Adversarial Training; Asymptotic Behavior of Adversarial Training in Binary Linear Classification.

Reinforcement and Imitation Learning

We are interested in learning from interaction and demonstrations, especially when online exploration is expensive, unsafe, or impractical. A recurring theme in our work is that demonstrations and annotations may come from people with very different levels of expertise. We developed expertise-aware imitation and inverse reinforcement learning methods that identify and exploit reliable behavior rather than treating all demonstrators equally. More recently, we have carried this idea into foundation-model adaptation: our REALM framework jointly learns an LLM and the reliability of heterogeneous annotators from noisy supervision.

Representative papers: REALM: Reliable Expertise-Aware Language Model Fine-Tuning from Noisy Annotations; Inverse Reinforcement Learning by Estimating Expertise of Demonstrators; Imitation Learning by Estimating Expertise of Demonstrators.

Past Projects

Fairness in Machine Learning

Sequential decision-making and learning methods that account for the long-term effects of algorithmic decisions on different groups, including dynamic notions of fairness and the tradeoff between predictive performance and equitable outcomes.

Intelligent Transportation and Mixed Autonomy

Modeling and control of transportation networks in which autonomous and human-driven vehicles coexist, including routing, pricing, capacity, traffic control, and mechanisms for influencing system-level behavior.

Coded and Distributed Computing

Coding and scheduling methods for speeding up large-scale distributed computation and machine learning in the presence of stragglers, heterogeneous workers, failures, and communication bottlenecks.

Robust Scheduling for Large-Scale Networks

Scheduling and control of stochastic networks with unknown or time-varying parameters, with applications to data centers, communication systems, and other large networked infrastructures.

Sparse Recovery and Sparse-Graph Codes

Low-complexity algorithms based on sparse-graph codes for compressed sensing, phase retrieval, group testing, sparse covariance estimation, and related high-dimensional inference problems.