Research

Publications

Quantifying and Mitigating the Impact of Obfuscations on Machine-Learning-Based Decompilation Improvement

Dramko L., Bölöni-Turgut D., Le Goues C., Schwartz E.

In Proceedings of Detection of Intrusions and Malware, and Vulnerability Assessment (DIMVA), 2025

Decompilation Malware Software Engineering

Team cornell-1 at PAN: Ensembling Fine-Tuned Transformer Models for Writing Style Analysis

Bölöni-Turgut D., Verma D., Cardie C.

In Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, CEUR-WS.org

Natural Language Processing Machine Learning Style Analysis

Experience

Undergraduate Researcher

Jan 2024 - Present

Cornell University NLP Group, Advisor: Dr. Claire Cardie

🤗

Multi-Author Sentence Pair Style Analysis (PAN 2025 Shared Task)

Design and train an ensemble model from multiple fine-tuned transformer models for multi-author writing style analysis. Given pairs of sentences, identify whether the sentences were written by the same author or different authors (binary classification). Submit system and first author paper to PAN Labs 2025 shared task, achieving average macro F1 score of 0.8.

Paper Code

commit0

Research LLM coding agents ability to implement Python packages given only textual specifications and a unit test suite, i.e. starting from commit0 of a GitHub repository.

README Code

Open Ended Question Evaluation

Evaluated length bias and self bias in open-ended essay responses by GPT and Llama3 LLMs. Utilized prompt engineering techniques to generate concise and verbose responses, while controlling and measuring semantic overlap. Work acknowledged in NAACL 2025 Findings paper.

Research Poster

Research Assistant

May 2023 - May 2024

Carnegie Mellon University, REU in Software Engineering, Advisor: Dr. Claire Le Goues

Developed neural approaches to decompilation (converting machine code back into source code) of C programs with applications in malware analysis. Compiled a 800GB+ novel dataset of obfuscated binary code, starting from open-source C-language repositories. Designed the data pipeline to automate GitHub scraping, obfuscation, and compilation with g++ and clang. Repurposed and refactored a large-scale Python codebase.

C/C++ Obfuscation

Outcome

Published paper in Detection of Intrusions and Malware & Vulnerability Assessment (DIMVA) 2025

Research Poster README