Varshini Elangovan
Hi, I’m a Master’s student in Machine Learning at Carnegie Mellon University, graduating in December 2026.
I’m interested in foundation model behavior, post-training, and evaluation, particularly how training and data interventions shape model behavior and alignment. I’m also interested in multimodal models with a focus on reasoning and interaction with complex environments.
At CMU, I currently work with Prof. Virginia Smith and Aashiq Muhamed on post-training interventions for LLM alignment, and previously worked with Prof. Fernando de la Torre at the Human Sensing Lab on efficient multimodal reasoning for 3D VQA with 2D vision-language models. Before CMU, I spent three years at Google Hyderabad, as part of the Global Payments Platform Risk team, working on production risk systems for Google Pay and Google Wallet.
I’m currently seeking full-time roles in ML research and engineering starting January 2027.
Industry Experience
Jul 2022–Jul 2025
As part of Google's Global Payments Platform Risk team in Hyderabad, I worked on real-time payment risk systems for Google Pay and Google Wallet, across fraud mitigation, feature engineering pipelines, ML and rule-based decision policies, and the engineering risk strategy for Wallet’s Brazil PIX launch in 2024.
Research Experience
LLM post-training & alignment · CMU · May 2026–present
with Virginia Smith and Aashiq Muhamed
Suppressing Persona-Conditioned Behaviors in LLMs
How does post-training modify the accessibility of personas?
Investigating how persona-conditioned behaviors emerge across LLM post-training, how they can be selectively suppressed, and whether suppression generalizes across elicitation settings while preserving general capabilities.
Model character & misalignment · Sep 2026–present
Domain-Specific Entanglements and Persona Characterization in Emergent Misalignment
Research Fellow at SPAR (Supervised Program for Alignment Research), with Praxis Research Group
Multimodal reasoning & efficient inference · CMU · Oct 2025–Aug 2026
with Fernando de la Torre · in collaboration with Meta Reality Labs
CoVeR: Coverage-Based Token Pruning for Multi-View 3D Reasoning in VLMs
What visual information do VLMs need for multi-view 3D reasoning?
Multi-view reasoning can overwhelm 2D VLMs with redundant visual tokens. We developed a training-free, geometry-guided selection method that preserves spatial coverage across views, retaining 93.5% of full-token performance with about 8% of the visual tokens, reducing LLM compute by 13.3× and speeding up inference by 2.9×.
Human-centered AI safety · CMU · May–Sep 2026
with Virginia Smith
Safety Nudges: User-Facing Interventions for Real-Time AI Risk Awareness
Can real-time warnings help users recognize problematic AI behavior?
Conducted a two-week, 50-participant user field study of Safety Nudges, a Chrome extension that surfaces real-time warnings for behaviors such as sycophancy and overconfidence, examining how these warnings affect user trust and verification behavior.