Hello, my name is Boglárka Ecsedi /ˈboɡlaːrkɒ ˈɛt͡ʃɛdi/
but everyone calls me Bogi /ˈboɡi/
CS PhD Candidate @ University of Toronto & Graduate Researcher @ Vector Institute
My research asks how machine learning systems can remain reliable when deployment conditions are harder than training conditions, and how modular AI systems can be adapted and deployed safely in real-world, open-world environments. Longer term, I am interested in the intersection of modular AI architectures, safety mechanisms, and governance frameworks that support robust, accountable, and socially beneficial frontier AI development.
Background & Introduction
At the Medical University of Vienna, I worked on clinically grounded cancer diagnostics problems with small, noisy, and heterogeneous datasets. I contributed to the development of DEBI-NN, a parameter-efficient neural architecture for small-data settings, and later led work on its regularization properties. These projects shaped my interest in how model design, training strategy, and evaluation interact when reliable generalization matters more than raw benchmark performance.
During my undergraduate studies at Georgia Tech (Hoffman Group), I broadened this work to transfer learning, uncertainty-aware adaptation, and model merging across vision and language. Through AUGCAL, KnOTS, and my thesis on LoRA-merge scaling behavior, I studied robustness under shift and efficient composition of independently trained experts. Through internships at Intel, TandemAI, University of Pennsylvania, and Carnegie Mellon University, I further developed methods under real-world constraints.
I am now a first-year direct-entry PhD student in Computer Science at the University of Toronto, supervised by Prof. Colin Raffel and affiliated with the Vector Institute. One of my current projects investigates model merging as an alternative to standard multitask supervised fine-tuning for combining instruction-following capabilities in large language models. Another one of my projects focuses on evaluating large language models under adaptive adversarial pressure, including pressure-conditioned safety metrics that capture how model behavior changes as attacker effort increases.
Selected Publications
- Impact of Regularization in Optimizing Distance-Encoding Biomorphic-Informational Neural Networks for Small Nuclear Medicine Datasets
EANM Innovation (2025), first author.
Role: led study design, literature review, formal analysis, validation, figures, and manuscript writing/revision. - [ICLR'25] Model Merging with SVD to Tie the Knots
The Thirteenth International Conference on Learning Representations (2025), co-author.
Role: conceptual development, experiment design/implementation, ablations, figures, and manuscript preparation. - [ICLR'24] AUGCAL: Improving Sim2Real Adaptation by Uncertainty Calibration on Augmented Synthetic Images
The Twelfth International Conference on Learning Representations (2024), co-author.
Role: literature review, experiment design/implementation, ablations, figures, and manuscript preparation. - DEBI-NN: Distance-Encoding Biomorphic-Informational Neural Networks for Minimizing the Number of Trainable Parameters
Neural Networks (2023), co-author.
Role: implemented visual rendering tool and parameter-count analysis, contributed to literature review, figures, and manuscript revision.
- Scaling Laws in Model Merging: Investigating the Scaling Behavior of Merging LoRA Models
Undergraduate Research Option Thesis, Georgia Institute of Technology (2025).
Role: sole author; designed and conducted the study, implemented experiments, analyzed results, and prepared manuscript/figures.
To see the full, up-to-date list of publications, visit my Google Scholar profile.
Beyond Research
Beyond research, I have been active in teaching, mentoring, science communication, and initiatives supporting diversity and international collaboration in science. I have served as a teaching assistant for Neural Networks and Deep Learning at the University of Toronto and graduate Computer Vision at Georgia Tech, mentored students over multiple years, and participated in public-facing science communication, including TEDx, FameLab, and the John von Neumann Series.
I have also contributed to initiatives such as Women in AI Research Canada, the Association of Hungarian Women in Science, EducationUSA's Cutting Edge Fields campaign, and the European Commission's #SHEU LEADS campaign.