|
Aditi
Master's Student in Intelligent Information Systems
Carnegie Mellon University, School of Computer Science I am interested in the evaluation, alignment, and deployment of intelligent systems (language models, agents, etc.). Previously, I was a Pre-doctoral Researcher at Google DeepMind, where I studied how language models can be linguistically and culturally adapted across different modalities. I was fortunate to work with Shachi Dave, Dr. Laura Rimell and Dr. Partha Talukdar. Prior to this, I was a Research Assistant at the Social AI Studio, Singapore University of Technology and Design (SUTD) advised by Dr. Roy Ka-Wei Lee, where I worked on multilingual and multimodal hate speech detection for low-resource languages and modalities. Email / Google Scholar / LinkedIn / GitHub |
|
News
|
Publications* denotes equal contribution / co-authorship. |
| ICML '26 |
ProEval: Proactive Failure Discovery and Efficient Performance Estimation for
Generative AI Evaluation
International Conference on Machine Learning (ICML), 2026
An active evaluation framework that estimates foundation model performance and proactively discovers
failures via Bayesian inference and transfer-learned Gaussian Process surrogates, reducing
evaluation cost by 65× while uncovering 5× more diverse failures across 16 LLMs/VLMs.
|
| AAAI '27 |
Steering for Multilingual Fact Recall in Vision-Language Models
Under Review, AAAI Conference on Artificial Intelligence (AAAI), 2027
Mechanistic interpretability analyses for VLMs using activation patching to identify cross-lingual
failure circuits. Shows that multilingual gaps stem from late-layer attention routing rather than
missing factual knowledge, and that text-derived steering vectors transfer across modalities to
recover up to 17% accuracy.
|
| TALLIP '26 |
SEAHateCheck: Functional Tests for Detecting Hate Speech in Low-Resource
Languages of Southeast Asia
ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP), 2026
The first functional hate speech benchmark for 8 Southeast Asian languages across 19 functional
tests. Includes a LoRA PEFT pipeline benchmarking 12 LLMs, exposing systematic model failures across
Age, Disability, and Gender categories.
|
| EMNLP '24 |
Bridging Modalities: Enhancing Cross-Modality Hate Speech Detection with
Few-Shot In-Context Learning
Empirical Methods in Natural Language Processing (EMNLP), 2024
Investigates cross-modal in-context learning strategies to transfer text-based hate detection
representations to multimodal inputs with few-shot demonstration selection.
|
| ACM MM '23 |
MATK: The Meme Analytical Tool Kit
ACM International Conference on Multimedia (ACM MM), 2023
A modular, extensible open-source toolkit unifying 8 datasets and 7 models for multimodal hateful
meme classification with YAML-driven experiment orchestration.
|
Honors & Awards
|