Lorenzo Sciarretta

I research how multiple large language models can work together: coordinate, agree, and reason as a team. I'm currently doing this at the DISCO Lab, ETH Zürich.

I'm finishing a joint Master's in Computer Science at ETH Zürich and the University of St. Gallen. My work sits between multi-agent LLM systems, weight-space learning, and post-training. I like systems that learn how to learn, models small enough to live on a chip, and clean abstractions that survive contact with reality.

I'm open to collaborations. If you work on post-training of LLMs, on the emergent behaviours that appear during pre-training and post-training and how they relate to model architecture, or on cooperative AI, I'd be glad to hear from you: write me an email.

Focus

2026–now Research on multi-LLM agentsDISCO Lab, ETH Zürich
2026 Compute grant SwissAI, student contributor16,000 GH200 GPU-hours on CSCS Alps
2025 Weight-space learningAI/ML Lab, HSG, with Prof. Damian Borth
2025 Research assistantETH AI Center / INI, with Prof. Shih-Chii Liu

Selected work

2026 AlphaGOmokuSelf-playing Gomoku agent in the style of AlphaZero, vibe-coded in a weekend with Gemini CLI.
2025 SANE-MoEExpert routing for reconstructing and generating neural network weights.
2024 NeuNetCNNEdge-deployable CNN for semantic segmentation on neuromorphic hardware, built at the ETH AI Center / INI.
2024 Simple Local RAGFully local retrieval-augmented generation: vector search, prompt chaining and document QA with open-weight LLMs.
2023 Kidney tumour segmentationCNN pipeline for kidney and tumour segmentation on volumetric NIfTI images, at NECSTLab.

Notes

Notes condensed from coursework and reading:

Reinforcement learning: from MDPs and POMDPs to tabular, approximate and model-based RL, in six chapters.
LLMs: notes and projects on large language models, starting with a fully local RAG system.
Statistical ML: classical machine learning in eight parts, from linear regression to neural networks.

Skills

Deep learning: PyTorch, Transformers, CNNs, mixture of experts, learning theory.
Generative AI: fine-tuning, RAG, prompt engineering, LangChain.
Cloud and deployment: Azure, Docker, Podman.
Databases: SQL, relational databases.
Languages: Python, Java, R, Julia, C.