Multimodal Reasoning & Integration
How do multimodal models decide which evidence to trust? I study whether MLLMs truly integrate complementary modalities — or over-rely on the most dominant or misleading signal.
01 — About
A question of evidence
I'm a Ph.D. student in Computer Science at Boston University, advised by Prof. Deepti Ghadiyaram in the Image & Video Computing Group. My research focuses on multimodal large language models, vision-language reasoning, robustness, uncertainty quantification, interpretability, and generative AI.
My recent work studies how multimodal models integrate conflicting information across different modalities — revealing modality biases and vulnerabilities such as cross-modal attacks. More broadly, I want to understand when models rely on the right evidence, how their attention and uncertainty reflect reasoning, and how training and evaluation can make these systems more faithful and reliable.
Previously I earned B.S. degrees in Mathematics and Computer Science from The Ohio State University, where I worked on weakly supervised segmentation, long-tailed recognition, LiDAR-based 3D detection, wildfire segmentation, and multimodal video synthesis. My work has appeared at ECCV, ICLR, CVPR, and NeurIPS venues. This summer I'll be a Research Intern at Google Research.
Building multimodal AI that knows what evidence to trust.
02 — Research
What I work on
Diagnosing and improving how multimodal models allocate attention, estimate uncertainty, and integrate evidence across modalities.
Robustness & Cross-Modal Attacks
Seemingly natural cues — on-screen text, captions, or textual prompts — can mislead multimodal models even when the correct evidence is present. I study these attacks and their safety implications.
Interpretable Multimodal AI
Using attention shifts, attribution, and log-probability diagnostics, I ask whether a model's internal evidence allocation explains when it succeeds, fails, or becomes overconfident.
Uncertainty Quantification for VLMs
Multimodal models should know not only what answer to give, but also when the evidence is insufficient, conflicting, or unreliable.
Generative AI & Object States
I improve text-to-image models' ability to represent fine-grained physical states — open, closed, full, broken, folded, or melted — via synthetic data and targeted fine-tuning.
Computer Vision Foundations
Earlier work building technical breadth: weakly supervised segmentation, long-tailed recognition, LiDAR 3D detection, wildfire segmentation, and topological losses.
A·V·T
CVPR 2026 Findings
Multimodal
Some Modalities Are More Equal Than Others: Decoding and Architecting Multimodal Integration in MLLMs
Controlled audio-video-text conflict benchmarks reveal that MLLMs exhibit strong modality biases rather than balanced multimodal reasoning.
Aa
Under Review
Robustness
A Systematic Study of Cross-Modal Typographic Attacks on Audio-Visual Reasoning
Semantic distractors injected through speech, on-screen text, and prompts can mislead audio-visual models even when correct evidence is present.
±
In Submission
Uncertainty
VLM-UQBench: A Benchmark for Modality-Specific and Cross-Modality Uncertainties in Vision-Language Models
A benchmark of human-labeled samples and synthetic perturbations to evaluate modality-specific uncertainty, calibration, and reliability in VLMs.
◐
CVPR 2025 Workshop
Generative
Improving Physical Object State Representation in Text-to-Image Generative Systems
An automatic synthetic-data pipeline and fine-tuning improve GenAI-Bench by 8.2% and two new object-state datasets by 17% / 24%.
⚡
CVPR 2025
Multimodal
EEE-Bench: A Comprehensive Multimodal Electrical and Electronics Engineering Benchmark
A multimodal benchmark probing reasoning over electrical and electronics engineering problems.
ICLR 2024
3D Vision
Pre-training LiDAR-based 3D Object Detectors through Colorization
Colorization as a pretext task for self-supervised pre-training of LiDAR 3D detectors in autonomous driving.
ECCV 2022
Segmentation
Learning with Free Object Segments for Long-Tailed Instance Segmentation
Mines free object segments from web images and ranks high-quality masks to improve long-tailed instance segmentation.
🔥
AIAA SciTech 2026
Segmentation
Centralized Copy-Paste: Enhanced Data Augmentation for Wildland Fire Semantic Segmentation
Extracts, centralizes, and pastes refined fire clusters to combat label scarcity and class imbalance in drone-based fire segmentation.
No publications in this category yet.
04 — Experience
Where I've worked
-
May – Aug 2026
Research Intern · Google Research
San Francisco Bay Area, CA
Robust multimodal LLMs — multimodal reasoning, evidence allocation, counterfactual evidence construction, and model behavior under conflicting or misleading cues. Exploring attention-aware supervision and self-distillation for better multimodal grounding.
-
Sep 2024 – Present
Research Assistant · BU Image & Video Computing
Boston, MA · Advisor: Dr. Deepti Ghadiyaram
Multimodal LLMs, vision-language reasoning, robustness, uncertainty, and generative AI: cross-modal conflict benchmarks, adversarial attacks, object-state generation, and modality-specific uncertainty.
-
Sep 2021 – May 2024
Research Assistant · OSU — ML as the Basis Lab
Columbus, OH · Advisor: Dr. Wei-Lun Chao
Weakly supervised segmentation with SAM, LiDAR-based 3D detection pre-training, long-tailed instance segmentation, and sensor-adaptation pipelines for 3D detection.
-
Aug 2023 – May 2024
Research Assistant · OSU Autonomy & Complex Systems Lab
Columbus, OH · Advisor: Dr. Mrinal Kumar
Semantic segmentation and data augmentation for wildland fire imagery, improving drone-based fire-monitoring perception.
-
Aug 2023 – May 2024
Researcher · OSU CS — ReelMaker & Talk-to-TikTok
Columbus, OH · Advisor: Dr. Arnab Nandi
Multimodal short-form video synthesis from papers and talks using PDF parsing, vision, speech-to-text, and LLMs. Most Innovative Project, HackOHI/O.
05 — Teaching
Teaching & mentoring
Teaching Fellow — CS 455: Computer Networks
Boston University · Spring 2025
Delivered lectures, led labs, and mentored students on routing, transport protocols, latency, and congestion control.
Teaching Assistant — CSE 1223 & CSE 2321
The Ohio State University · 2020–2024
Intro to Programming in Java; Foundations I: Discrete Structures.
I enjoy connecting formal concepts with intuitive examples, visual explanations, and hands-on implementation.
06 — Service & Honors
Community
Reviewer
CVPRICCV ICLRAAAI WACVCOLM TCSVTIJCV
Student Academic Chair — BU AI Research Initiative
Oct 2024 – Aug 2025
Organized AIR Seminar programming and academic events.
Honors & Awards
Dean's List (all attended semesters) · HackOHI/O 11 Most Innovative Project Award.
07 — Toolkit
Skills
Research Areas
Multimodal LLMs · Vision-Language Reasoning · Robustness & Safety · Uncertainty Quantification · Interpretability · Generative AI · Semantic Segmentation · 3D Detection
Tools
PyTorch · TensorFlow · Hugging Face Transformers · OpenCLIP/CLIP · Qwen-VL/Qwen-Omni · LLaVA-style models · Gemini APIs · WandB · Docker · Git · AWS S3
Languages
Python · MATLAB · Java · C · R · SQL · LaTeX
Get in touch
Let's talk about multimodal AI.
Open to research collaborations, reviewing, and conversations about trustworthy multimodal systems.