Lorenzo Steno

Founding Researcher at Seldon Technologies

Email · GitHub · LinkedIn · Google Scholar · Blog

About

I am a Founding Researcher at Seldon Technologies, where I focus on post-training and evaluation for large language models.

I hold a Master's degree in Computer Science and completed my thesis at ETH Zurich's Agentic Systems Lab. Previously, I worked as a Solutions Architect Intern at AWS and as a Research Assistant at the AI & IoT Lab at the University of Twente.

Research Interests

Post-training and evaluation of large language models
Safety and trustworthiness of large language models
Reinforcement learning for complex decision-making

Selected Projects

CADBench: How Good Are Agents Actually at CAD?

CADBench evaluates whether frontier computer-use agents can execute 105 expert-authored, long-horizon mechanical-design tasks in Autodesk Fusion and produce reliable, editable CAD models that pass deterministic verification.

Computer-Use Agents CAD Evaluation

Cost-Aware Recursive Language Models with RLVR

Built an RL environment to train Qwen3-4B Recursive Language Models. Released LoRA adapters and checkpoints that raised pass@1 from 22.5% to 64.8% while reducing cost from 139k tokens to 433.

Recursive Language Models RLVR LoRA

Clash Royale Reinforcement Learning Agent

Developed a model-based RL agent (DreamerV3) to master complex real-time strategy gameplay. Engineered a visual perception pipeline using YOLO for real-time game state extraction.

Reinforcement Learning Computer Vision DreamerV3