Alexander George Padula

We implemented a new board-game DSL in JAX: games compile to XLA, enabling RL training entirely on the GPU. Offers up to 10x speedup against previous SoTA DSL and lays groundwork for training general game-playing models.

We designed a system to generate novel board games in a DSL using a LLM as a mutation operator. Automated evaluation through play-through sampling between general game-playing agents, with heuristics to score runs.

In my BSc thesis, I explored novel PPO variations to stabilize LLM training with explicit rewards (now known as RLVR). Restored KL-divergence to the classic RL formulation and introduced a novel batch-entropy exploration bonus.

I'll work on reward-based post-training for code at Mistral and could choose to stay for an industry PhD.

Contributed to the online RL pipeline for Apertus, a fully open-source 70B LLM. Improved code sandbox stability and reduced training time 6x.

Authored an internal paper on code generation with LLMs. Presented research to 60+ employees, gaining traction across departments with the potential to impact the next product cycle. Learned to share complex ideas concisely.

Developed a coding assistant for the Ludii DSL. Trained a LLM and overcame data scarcity using grammar constrained decoding; had to develop a parser for incomplete code that could mask illegal tokens during inference.

Covered tuition and rent by serving as a university teaching assistant and tutoring privately in computer science.

Built web interface currently used by glaucoma patients and their doctors to manage data and view test results. Worked with founders and dev team. Limited technical debt and managed state across complex React applications.

Built and deployed a content-based image search engine. Mentored by Dr. Bolioli’s team, I trained BERT and Inception V3 to embed images and descriptions in a shared latent space. This project underpinned my high school thesis.

Worked in the dev team of a mobile digital payment application. Debugged Node.js API and built a prototype in React. Operated in a large codebase and loved working in a team. Received an offer to return and a recommendation letter.

In collaboration with IBM, I'm developing a post-training method to make LLMs more robust to the noise introduced by analog computations. This is a critical step toward commercializing IBM's in-memory-compute devices.

We derived Muon's preconditioner and found that Muon operates beyond the classical edge of stability. We also empirically validated that the flat minima hypothesis holds for Muon’s solutions in a realistic setting.

Exploring RL for massively multi-agent, mixed-motive games. Developing open-source gym API for Open Front.

Only course project selected by IVIA lab for publication (under review). We built a LLM-assisted study environment that syncs recordings and exercises into a visual knowledge graph to quickly navigate between multimodal-sources.

Designed novel data-passthrough joints and contributed to a ML model for inferring a modular robot’s configuration.

Built 3D printer and transformed it into a DIY laser cutter. Developed the open-source Python library svg-to-gcode. The tool’s GUI has garnered over 20,000 downloads and a community that helps maintain it.