Available for research collaborations

Giovanni Marraffini

PhD Candidate · LLMs, AI Safety & Computational Neuroscience

I work at the intersection of AI and cognition at INRIA & Université Paris-Saclay — spanning AI safety and interpretability, and bridging large language models, brain data, and the study of the mind.

Download CV View publications
Giovanni Marraffini
7
Papers & preprints
ACL · EMNLP · NeurIPS · ICML
Publication venues
Paris
Currently based in France

About

Research at the intersection of AI and the mind

I am a PhD candidate at INRIA and Université Paris-Saclay, working at the intersection of AI and cognition — spanning AI safety and interpretability, large language models, and computational neuroscience.

My current work pretrains and evaluates foundation models on large-scale brain data (fMRI and EEG) to understand how machine and biological representations relate. Across my research I study how models represent and reason about the world — from alignment and moral judgment to causal biases in LLMs, mechanistic interpretability, and the correspondence between AI representations and the human brain. I care about building rigorous, reproducible work that bridges large language models, brain data, and the study of the mind.

Research interests

AI safety Mechanistic interpretability LLM alignment Causality in LLMs Foundation models AI–brain representations Computational neuroscience

Publications

Selected papers

Full list on Google Scholar. Names in bold indicate my authorship.

Preprint
Under review
2026

The Variance Brain Foundation Models Forgot: Third-Order Statistics Predict Cognition Where Billion-Parameter Models Fail

Marraffini, G., Mahuas, G., Borrell, T., Shevchenko, V., Wassermann, D.

First comprehensive benchmarking of fMRI foundation models, showing that third-order statistics predict cognition where billion-parameter models fail.

arXiv
Play — what do billion-parameter models forget? Hide demo

Brain foundation models preserve a signal's variance (second-order structure) but largely destroy its co-skewness (third-order structure) — and it's that third-order signal that predicts cognition. Drag the skew and watch what a second-order model can't see.

← left tailSymmetricright tail →
Mean 0.00 Variance 1.00 Skewness +0.00

Drag the slider to reshape the distribution.

mean fixed  ·  variance fixed

A conceptual illustration of the paper's core result: mean and variance stay pinned while only the third-order structure changes — the "variance the models forgot." A model that captures a signal only up to second order cannot tell these distributions apart, yet the third-order signal is what tracks cognition.

ICML 2026
Mech. Interp. Workshop
2026

The Platonic Universe: Do Foundation Models See the Same Sky?

Borrell, T., Dillmann, S., Duraphe, K., Eris, F., Khederlarian, A., Kumar, A., Marraffini, G., Smith, M. J., Sourav, S., Di Tella, R., Wu, J. F.

Tests the Platonic Representation Hypothesis in astronomy — probing whether foundation models trained on different data converge toward a shared internal representation of the world.

OpenReview Code
Preprint
bioRxiv
2025

Shared Hierarchical Representations Explain Temporal Correspondence Between Brain Activity and Deep Neural Networks

Holm, E. L., Marraffini, G., Fernández Slezak, D., Tagliazucchi, E.

Shows that shared hierarchical representations explain the temporal correspondence between human EEG activity and deep neural networks during visual perception — early components track low-level features, later ones semantic content.

bioRxiv
NeurIPS 2025
CogInterp Workshop
2025

Do Large Language Models Show Biases in Causal Learning? Insights from Contingency Judgment

Carro, M., Mester, D., Gauna Selasco, F., Marraffini, G., Leiva, M., Simari, G., Martinez, V.

Shows that LLMs systematically infer unwarranted causal relationships in null-contingency scenarios, mirroring a human cognitive bias.

arXiv
ACL 2025
Main Conference
2025

Are Optimal Algorithms Still Optimal? Rethinking Sorting in LLM-Based Pairwise Ranking with Batching and Caching

Wisznia, J., Bolaños, C., Gianolini, A., Hsueh, N., Marraffini, G., Tollo, J., Del Corro, L.

Rethinks sorting algorithms for LLM-based pairwise ranking, using batching and caching to cut cost while preserving ranking quality.

ACL Anthology
EMNLP 2024
Main Conference
2024

The Greatest Good Benchmark: Measuring LLMs' Alignment with Utilitarian Moral Dilemmas

Marraffini, G., Cotton, A., Hsueh, N., Wisznia, J., Fridman, A., Del Corro, L.

A benchmark measuring how LLMs' moral judgments align with utilitarian dilemmas, revealing consistently encoded moral preferences across 15 models.

ACL Anthology arXiv Code
Play — are you more utilitarian than the models? Hide demo

Rate each statement from the Oxford Utilitarianism Scale — the same construct my benchmark probes in language models — and watch yourself move on the map.

Impartial beneficence We should give the well-being of a stranger on the other side of the world the same moral weight as someone close to us.

Strongly disagreeNeutralStrongly agree

Impartial beneficence It is wrong to keep money you don't really need if you could instead donate it to save lives.

Strongly disagreeNeutralStrongly agree

Instrumental harm It can be morally right to harm one innocent person if that is the only way to save several others.

Strongly disagreeNeutralStrongly agree

Instrumental harm Torturing one person could be justified if it were the only way to stop an attack that would kill hundreds.

Strongly disagreeNeutralStrongly agree
LLMs tended to land here Instrumental harm → Impartial beneficence →
Impartial beneficence 50% Instrumental harm 50%

Move the sliders to place yourself on the map.

This mirrors what the benchmark probes in language models. The highlighted region illustrates the paper's headline finding — most models cluster toward strong impartial beneficence while rejecting instrumental harm — rather than exact per-model coordinates.

Experience

Research, industry & teaching

PhD CandidateINRIA · Université Paris-SaclayMar 2026 — Present
Paris, France

Foundation-model pretraining and evaluation for understanding cognition. First comprehensive benchmarking of fMRI foundation models, probing whether internal representations retain cognitively relevant variance, then fine-tuning to recover a reconstruction-vs-representation trade-off.

AI ConsultantSigma NovaMar 2026 — Present
Paris, France

Designed and maintained training and evaluation pipelines for large-scale fMRI foundation models on Scaleway clusters, managing experiment tracking, compute provisioning, and reproducibility tooling.

Lead AI InstructorAnyone AIMar 2026 — Present
Remote

Helping software engineers transition into AI-centered careers by teaching the latest advances and industry best practices.

Research EngineerParis Brain InstituteMay 2025 — Mar 2026
Paris, France

Applied state-of-the-art time-series foundation models to analyze large-scale EEG data from patients with neurodegenerative diseases.

AI DeveloperData VoicesApr 2025 — Dec 2025
Remote

Developed and deployed AI solutions for international companies, from real-time conversational responses to large-scale text-document reorganization.

NLP EngineerLumina LabsDec 2023 — May 2025
Remote

Built a multi-agent RAG pipeline for financial presentations, involving LLM fine-tuning, retrieval reranking, prompt engineering, and MLOps for deployment on Azure.

Teaching FellowUniversidad de Buenos AiresJun 2024 — Feb 2025
Buenos Aires, Argentina

Taught introductory and advanced algorithms courses for Data Science and Computer Science students.

Research AssistantUniversidad de Buenos AiresFeb 2024 — Dec 2024
Buenos Aires, Argentina

EEG–ViT comparison study: analyzed EEG data from subjects viewing images, comparing the timing of human visual processing with activation layers in Vision Transformers.

Founder · Paper Reading ClubUniversidad de Buenos AiresSep 2023 — Dec 2024
Buenos Aires, Argentina

Founded and led a research paper reading club for data science and computer science students to stay current with the latest research.

NLP ResearcherUniversidad de Buenos AiresJun 2023 — Dec 2023
Buenos Aires, Argentina

Researched the impact of large language models on job roles and employment for the Ministry of Labour.

Teaching AssistantUniversidad Torcuato Di TellaMar 2023 — Aug 2024
Buenos Aires, Argentina

Taught Graph Theory and Computational Business Applications.

Education

Academic background

2026 — Present

PhD in Computational Neuroscience

Université Paris-Saclay
Paris, France
2022 — 2024

MSc in Data Science

Universidad de Buenos Aires
Buenos Aires, Argentina
GPA 9.30 / 10
2019 — 2022

BSc in Data Science

Universidad de Buenos Aires
Buenos Aires, Argentina
GPA 8.75 / 10

Conferences

Talks & presentations

Jul 2026

ICML 2026

Mechanistic Interpretability Workshop
Seoul, South Korea
Workshop paper
Jun 2026

OHBM 2026

Organization for Human Brain Mapping
Bordeaux, France
Poster
Dec 2025

NeurIPS 2025

Neural Information Processing Systems
San Diego, California
2 accepted papers
Jul 2025

ACL 2025

Assoc. for Computational Linguistics
Vienna, Austria
Main conference
Mar 2025

KHIPU 2025

Latin American Meeting in AI
Santiago, Chile
Poster
Nov 2024

EMNLP 2024

Empirical Methods in NLP
Miami, Florida
Main conference
Oct 2024

SAN 2024

Sociedad Argentina de Neurociencias
Buenos Aires, Argentina
Poster

Skills

Technologies & tools

Machine Learning

Foundation-model pretraining LLM fine-tuning Mechanistic interpretability AI safety & alignment RAG Prompt engineering Transformers Vision Transformers Representation learning Optimization

Languages & Frameworks

Python PyTorch HuggingFace LangChain LangSmith MNE · Nilearn fmriprep C++ R SQL Azure

Languages (spoken)

Spanish — Native English — Proficient French — Intermediate

Contact

Let's get in touch

Open to research collaborations, talks, and interesting problems at the intersection of AI and cognition. The fastest way to reach me is by email.

giovanni.marraffini@gmail.com Google Scholar GitHub LinkedIn