Isotta Magistrali

Subliminal Signals in Preference Labels

Isotta Magistrali, Frédéric Berdoz, Sam Dauncey, Roger Wattenhofer

2026, ICLR @ Agents in the Wild Workshop

Does the Model Know Which Steps Matter? Probing Causal Importance in Chain-of-Thought Reasoning

Probing whether large language models can identify which intermediate steps causally drive their final answer in chain-of-thought reasoning.

Improved Detection of Statistical Entities

Oana Balalau, Théo Galizzi, Isotta Magistrali, Ioana Manolescu, Gabriele Mura

2024, Infox sur Seine - short paper

OpenICE-SLAM: Open-Set Neural Implicit Encoding for Semantic SLAM

Open-set neural implicit encoding for semantic SLAM, building dense scene representations that jointly capture geometry and open-vocabulary semantics.

2025

Chaos and convergence in models of random recurrent neural networks

Study of the dynamics of random recurrent neural networks, focusing on numerical analysis of dynamical properties such as chaos, stability, and convergence. Developed biologically motivated extensions to explore the intersection of AI and neuroscience for more sustainable and efficient architectures.

Work completed as my Bachelor’s thesis

Deep Jigsaw: neural approaches to visual puzzle solving

Designed and trained neural networks to solve visual jigsaw puzzles by reordering shuffled image patches. Compared CNN-based spatial prediction with pre-trained AlexNet and Vision Transformer encoders combined with Sinkhorn-based permutation optimization.

Mapping neurodevelopmental signatures in human brain organoids

Fine-tuning transformer-based models for single-cell RNA sequencing data to predict cellular exposure to endocrine-disrupting chemicals and hormonal pathways, linked to neurodevelopmental disorders.

Correlation and inference within and across brain areas

Analysis of the Allen Brain Observatory to model neural dynamics in the mouse visual cortex using correlation analysis, maximum entropy modeling, and classical machine learning for firing rate prediction.

Machine Learning approaches for predicting treatment resistance in cancer

An applied machine learning and deep learning study on gene expression data from breast cancer cells to predict hypoxic and normoxic states, aiming to identify mechanisms linked to treatment resistance.

Network science applied to the Venture Capital industry

Analysis of large-scale datasets containing millions of entries to uncover structural patterns and leader–follower dynamics within the Venture Capital ecosystem, quantify and interpret complex network relationships.

3D Ising model simulation

Developed Python simulations of the 3D Ising model to analyze magnetization–temperature relationships across varying spin configurations.

The paradox of well-being: how prosperity relates to mental health

A statistical study examining how socio-economic factors and the COVID-19 pandemic relate to the prevalence of mental health disorders across countries.