Abstract: We introduce a data-efficient approach using fine-tuned LLMs and few-shot Siamese networks for classifying flaky software tests. Our FlakyXbert model leverages knowledge distillation, data augmentation, and fairness-aware strategies to improve flaky test detection accuracy in enterprise environments by over 20%. This study establishes actionable LLM-based baselines for test reliability under realistic, resource-constrained conditions.

Abstract: This work presents ALIGNFREEZE, a framework and evaluation suite studying how fine-tuning and realignment affect the inner workings of large multilingual language models (LLMs) like XLM-R and mBERT. Probing over 15 languages, we reveal which layers and representations become robust or fragile under various alignment strategies, informing safe and fair deployment of multilingual LLMs. Accepted as a regular (full) paper at NAACL 2025, contributing new benchmarks to the community.

Abstract: This paper exposes and addresses major data leakage issues in the widely-adopted HumanEval benchmark, used for LLM code generation testing. By proposing a robust, combinatorial design for test splitting and coverage validation, we provide the first automatic system for data contamination discovery—improving reliability standards for LLM program synthesis evaluation by over 40%.

Abstract: Investigates how common data augmentation techniques can unintentionally bias both training and evaluation of ML models. Through controlled experiments on multiple datasets, we show that augmentation-driven class imbalance and complexity shifts affect fairness, robustness, and downstream generalization—highlighting best practices for equitable, unbiased ML pipelines.

Abstract: This thesis introduces a reinforcement learning framework for real-time collision avoidance in autonomous maritime vehicles. Using Unity simulation with lidar-derived sensor streams, the system trains agents to anticipate and prevent ship-whale collisions, paving the way for safer, AI-assisted navigation with significant implications for maritime environmental protection. The work demonstrates successful deployment of RL policies in noisy, real-world-inspired conditions.