Federico Villa

Reinforcement learning-based approach for training large language models to generate examples that are difficult to translate. Using Group Relative Policy Optimization (GRPO), we modify small LLMs to improve how they generate challenging examples, outperforming base models and alternative methods.

Study on the impact of staleness in Asynchronous Stochastic Gradient Descent (ASGD) on model generalization. Our research demonstrated that ASGD variants achieve comparable convergence and equal or better generalization than standard SGD despite asynchronous updates.

Developed a novel PyTorch monocular depth estimation model by applying transfer learning from a transformer-based architecture (SegFormer - image segmentation), achieving high accuracy with minimal fine-tuning and outperforming traditional approaches.

Studying how LLM-based agents access personal information, compose their decision making process and how to improve the human consent in the loop. The project was a joint collaboration between Microsoft Research and EPFL cryptography research lab.

Tested the robustness and security of the codebase of the popular open source terminal multiplexer tmux searching for memory bugs. Code testing was performed using state of the art fuzzing techniques with OSS-Fuzz and libDebug harness.

Semester Research project aimed at adapting an open-source C++ simulator to emulate side-channel attacks. Conducted under the supervision of faculty members at Politecnico di Milano.

Digital implementation of the board game Codex Naturalis. Developed a Java version of the board game enabling remote playing via RMI and Socket protocols. Clients can play via a GUI (JavaFX) or a CLI. Features include chat, multiple game hosting, and resilience to disconnections.