Carlo Nicolini – AI Research Scientist

About Me

Research focus: Mechanistic interpretability, multi-agent AI systems, and machine learning for language and finance. I’m also interested in open-source ML tools and the statistical physics of complex systems.

Hi, I’m Carlo Nicolini—a physicist turned computational scientist, now Senior AI Research Scientist at Ipazia SpA in Milan. I build multi-agent AI systems, work on deep learning and interpretability, and maintain skfolio for portfolio optimization.

My research bridges statistical physics, complex networks, and artificial intelligence. I publish at venues like COLM and ICAIF, and my work appears on Google Scholar.

Senior AI Research Scientist, Ipazia SpA (2022–present)
Maintainer of skfolio (portfolio optimization in Python)
Research: interpretability, multi-agent systems, NLP, complex systems, statistical physics

Selected publications

Recent work on LLM interpretability, NLP, and statistical physics. Full list →

Sokobench: Evaluating Long-Horizon Planning and Reasoning in Large Language Models — Sebastiano Monti, Carlo Nicolini, Gianni Pellegrini, Jacopo Staiano and , Bruno Lepri, Transaction of Machine Learning Research (2026)
Unveiling LLMs: The evolution of latent representations in a dynamic knowledge graph — Marco Bronzini, Carlo Nicolini, Bruno Lepri, Jacopo Staiano, Andrea Passerini, COLM 2025: Conference On Language Modeling (2025)
Glitter or gold? Deriving structured insights from sustainability reports via large language models — Marco Bronzini, Carlo Nicolini, Bruno Lepri, Jacopo Staiano, Andrea Passerini, EPJ Data Science (2024)
Spectral Entropy of Random Spanning Forests — Carlo Nicolini, arxiv preprint (2025)

Software & code

Open-source projects I develop or maintain. Full list →

xyz — Information-theoretic estimators for continuous and time-series data
skfolio-online — Online Convex Optimization branch of the skfolio library
autoresearch-skfolio — Autonomous portfolio optimization research using AI agents
scikit-xml — Evaluation metrics for extreme multilabel classification
scikit-omikuji — Scikit-learn compatible wrapper for Omikuji extreme multi-label classification
loop-erased-thermodynamics — Wilson's algorithm for thermodynamic analysis of graphs

Latest blog posts

PhD studies

In my PhD I tackled the problem of modular structure identification in brain functional networks, from the point of view of complex networks. Complex networks theory offers a framework for the analysis of brain functional connectivity as measured by magnetic resonance imaging. Within this approach the brain is represented as a graph comprising nodes connected by links, with nodes corresponding to brain regions and the links to measures of inter-regional interaction. A number of graph theoretical methods have been proposed to analyze the modular structure of these networks. The most widely used metric is Newman's Modularity, which identifies modules within which links are more abundant than expected on the basis of a random network. However, Modularity is limited in its ability to detect relatively small communities, a problem known as resolution limit.
To read more, here is my PhD thesis.

Contact

Find my contact on LinkedIn, then write me!