Thomas Katraouras

Machine Learning Engineer & PhD Candidate building the next generation of safe, scalable, and resource-efficient AI. My research focuses on the continual adaptation and compression of Large Language Models. My published work optimizes deep learning models, resulting in significantly improved efficiency and state-of-the-art performance across vision and language tasks.

tkatraouras [at] uth [dot] gr

Thomas Katraouras's GitHub Stats (Light Mode) Thomas Katraouras's GitHub Stats (Dark Mode)

Experience

Education

Publications

Continual Alignment of Large Language Models with Entropic Gradient Refraction

Thomas Katraouras, Dimitrios Rafailidis

Under Review

Proposed EGR, a continual alignment strategy that treats the parameter space as a heterogeneous medium and the optimization path as a light ray propagating through it, where the medium's density is defined by the model's confidence on past tasks. Gradient updates are refracted away from previously learned structural knowledge, balancing stability and plasticity during LLM continual learning. Improved performance by 170% (BLEU-4), 5.4% (ROUGE-L), and 8.2% (LLM-as-a-Judge) over baselines.

Code

Memory Bank Compression for Continual Adaptation of Large Language Models

Thomas Katraouras, Dimitrios Rafailidis

The 41st ACM/SIGAPP Symposium on Applied Computing, 2026

Proposed MBC, a memory-augmented continual learning model that reduces the memory bank footprint by 99.7% by compressing stored representations through codebook optimization and online resetting mechanisms. Improved QA accuracy by 11.84% (EM) and 12.99% (F1) over state-of-the-art baselines, enabling efficient knowledge updates in LLMs without catastrophic forgetting.

PDF/arXiv Code

Pruning Overparameterized Multi-Task Networks for Degraded Web Image Restoration

Thomas Katraouras, Dimitrios Rafailidis

IEEE/WIC International Conference on Web Intelligence and Intelligent Agent Technology, 2025

Proposed MIR-L, a compressed multi-task image restoration model that achieves 90% reduction in trainable parameters through lottery tickets while maintaining or exceeding state-of-the-art performance on deraining, dehazing, and denoising tasks.

PDF/arXiv Code

Academic & Professional Service

Program Committee Member

  • ACM/SIGAPP Symposium on Applied Computing (SAC) 2026

Projects

Bitcoin-SDL

Developed a LSTM model for Bitcoin price forecasting that integrates X (Twitter) sentiment analysis, achieving 15.88% reduction in RMSE, 15.49% reduction in MAE, and 11.71% reduction in MAPE compared to price-only baseline.

News Classification with Neural Nets

Developed a text classification system for news articles using CNNs with custom word embeddings and fine-tuned Transformer models (DistilBERT) to improve accuracy to 78.9%.

Steam Games Recommender System

Built a recommender system for Steam games using content-based and collaborative filtering (Cosine Similarity, SVD), featuring a custom review-weighted algorithm, an experimental Neural Network with embeddings, and an interactive GUI.

Certifications

Tech Stack

▲
×