My research lies at the intersection of computational neuroscience and machine learning, where I explore how neural systems learn meaningful representations and abstractions that underlie reasoning and decision-making. --- ## Selected Publications \* *First author* contributions <div style="text-align: center; margin-left: -20px; margin-right: -20px;"> <strong style="font-size: 1.4em;">G1Bbon: a temporal reasoning benchmark for LLMs </strong> </div> <div style="text-align: center;"> Gonçalo Guiomar*, <strong>Elia Torre*</strong>, Mario Giulianelli, Victoria Shavina, Valerio Mante<br></div> <span> <table style="width:100%; border-collapse: collapse; border: none;"> <tr> <td style="border: none; width: 40%; vertical-align: middle; text-align: center;">![[gibbon.png]]</td> <td style="border: none; width: 60%; vertical-align: top; text-align: left; padding-left: 20px;"> <div style="margin-bottom: 10px; text-align: left;"> The <strong>G1Bbon benchmark</strong> is based on a set of temporal reasoning tasks, designed to test decision-making capabilities across human players, optimal agents and language models. The ultimate goal is to drive the understanding of the parallels between biological and artificial intelligence by uncovering interpretable mechanisms that underlie temporal reasoning and pattern recognition, while also providing a platform for evaluating the performance of language models and improving their agentic capabilities under stochastic environments. </div> [Learn more and play!](https://ai.trt-bench.org/introduction) </td> </tr> </table> </span> <div style="margin-bottom: 3px; text-align: left; font-size: 0.9em;"> <em><strong>Keywords</strong>: LLM Evaluations, Mechanistic Interpretability, Sequential Decision Making</em> <div style="margin-bottom: 3px; text-align: right;"> <em>Work in progress</em> </div> --- <div style="text-align: center; margin-left: -20px; margin-right: -20px;"> <strong style="font-size: 1.4em;">Mechanistic Interpretability of RNNs emulating Hidden Markov Models</strong> </div> <div style="text-align: center;"> <strong>Elia Torre</strong>, Michele Viscione, Lucas Pompe, Benjamin Grewe, Valerio Mante<br></div> <span> <table style="width:100%; border-collapse: collapse; border: none;"> <tr> <td style="border: none; width: 45%; vertical-align: middle; text-align: center;">![[rnn.png]]</td> <td style="border: none; width: 55%; vertical-align: top; text-align: left; padding-left: 20px;"> <div style="margin-bottom: 10px; text-align: left;"> We show how RNNs can learn to replicate the discrete, stochastic behavior of Hidden Markov Models (HMMs) through continuous dynamics, implementing a noise-sustained limit cycle with distinct slow regions corresponding to different output states. We uncover, via ablation studies, a key mechanism involving specialized <em>"kick neurons"</em> that trigger state transitions when activated by noise amplified through larger integrating neural populations, operating in a regime of self-induced stochastic resonance where structured connectivity converts input noise variance into quasi-periodic switching between behavioral states. </div> </td> </tr> </table> </span> <div style="margin-bottom: 3px; text-align: left; font-size: 0.9em;"> <em><strong>Keywords</strong>: RNNs, HHMs, Mechanistic Interpretability, Dynamical Systems</em> </div> <div style="margin-bottom: 3px; text-align: right;"> <em>Under review at NeurIPS 2025</em> </div> --- ## Projects & Hackathons as a Student Here you can find a list of the projects I worked on mostly during my studies or as extracurricular activities. They all have an associated Github repository and/or Project report which can be found in the Metadata inside the individual project page. - #### 2023 - [[Translational Neuromodeling - EEG in Auditory Mismatch Negativity]] - [[Solar Irradiance Prediction]] - #### 2022 - [[An Applied Evolutionary Analysis of Neural Machine Translation]] - [[Spotify Song Popularity Prediction & Playlist Content-Based Recommendation System]] - #### 2021 - [[Vodafone Data Science Hackathon 2021]] - #### 2020 - [[Telco-Customers Churn-Rate Analysis]]