Research

Projects

Ongoing and completed research projects across our wings.

Project list

TRoPE-TUL: Trajectory-User Linking via Spatio-Temporal Rotary Position Embedding

TRoPE-TUL: Trajectory-User Linking via Spatio-Temporal Rotary Position Embedding

Trajectory-user linking identifies the likely user behind an anonymous mobility trace, but existing methods often lose fine-grained spatial, temporal, and semantic information or scale poorly.…

View project
Where the Preference Signal Lives: Constructed Negatives, Token Weighting, and Why Confidence Misses the Decisive Tokens

Where the Preference Signal Lives: Constructed Negatives, Token Weighting, and Why Confidence Misses the Decisive Tokens

This project studies how to make the reasoning produced by medical LLMs more factually reliable, not just their final answers. We construct controlled incorrect rationales with known clinically…

View project
How Good Must a Detector Be? Graded Semantic-Error Detection in Multi-Agent LLM Systems

How Good Must a Detector Be? Graded Semantic-Error Detection in Multi-Agent LLM Systems

This project studies when error detection and intervention can reliably improve multi-agent LLM systems, where mistakes from one agent may propagate through later interactions and affect the team’s…

View project
pFedBBN: A Personalized Federated Test-Time Adaptation with Balanced Batch Normalization for Class-Imbalanced Data

pFedBBN: A Personalized Federated Test-Time Adaptation with Balanced Batch Normalization for Class-Imbalanced Data

Personalized Federated Test-Time Adaptation studies how decentralized models can adapt to changing domains while also remaining robust to severe class imbalance across clients. Existing federated…

View project
FedPoisonTTP: A Threat Model and Poisoning Attack for Federated Test-Time Personalization

FedPoisonTTP: A Threat Model and Poisoning Attack for Federated Test-Time Personalization

Federated Test-Time Adaptation (FTTA) studies how decentralized models can adapt to client-specific distribution shifts at test time using unlabeled local data, without sharing raw data. This project…

View project
FedRAPP: Federated Test-Time Prompt Tuning for Vision-Language Models with Reliability-Aware Prompt Personalization

FedRAPP: Federated Test-Time Prompt Tuning for Vision-Language Models with Reliability-Aware Prompt Personalization

Source-Free Federated Test-Time Prompt Tuning (SF-FTTPT) studies how vision-language models can collaboratively adapt prompts at test time when clients have private, heterogeneous data with no access…

View project
Cross-Lingual Concept Steering and Interpretability in Language Models

Cross-Lingual Concept Steering and Interpretability in Language Models

Large language models serve speakers of hundreds of languages, yet most of what is known about their inner workings comes from English. Recent interpretability research shows that abstract concepts,…

View project
Semantic and Contextual Conditioning for Flow-Matching Speech Tokenization

Semantic and Contextual Conditioning for Flow-Matching Speech Tokenization

Modern spoken language systems convert raw audio into compact discrete representations that downstream models generate from and reason over. The central constraint is compression: as few tokens per…

View project
Speech-to-Motion: From Next-Token to Next-Gesture Prediction

Speech-to-Motion: From Next-Token to Next-Gesture Prediction

Large language models learn to predict sequences from nothing but text, yet human communication was never just words. Every sentence is accompanied by gesture: the hands, the shoulders, the tilt of a…

View project
Reliable Multimodal Document Intelligence for Low-Resource Languages

Reliable Multimodal Document Intelligence for Low-Resource Languages

This project develops a unified framework for reliable multimodal document understanding in low-resource and multilingual settings It consists of four complementary parts that progressively test…

View project
PACER - Personalized Adaptive Coaching for Engagement and Reflection

PACER - Personalized Adaptive Coaching for Engagement and Reflection

PACER (Personalized Adaptive Coaching for Engagement and Reflection) is an agentic AI tutoring system for IUB courses in Computer Programming |, Computer Programming II and Data Structure. While LLMs…

View project
LangTutor: An Agentic AI Tutor for Adaptive Language Learning

LangTutor: An Agentic AI Tutor for Adaptive Language Learning

LangTutor is an agentic AI tutor that decides what to teach next after each learner response, combining pedagogical methods such as scaffolding, hint-based guidance, retrieval practice, spaced…

View project