08.2026
Real-time Geometry Dash control from pixels using an 8x8 FSQ token grid, action-conditioned transformer dynamics, and an actor-critic refined with PPO entirely in frozen-model rollouts.
Florent Tariolle
I build deep neural networks that think, predict and act.
World models · Self-supervised learning · Reinforcement learning
Research that crosses representation and action.
I am an engineering student in Computer Science and Artificial Intelligence at INSA Rouen Normandy and an incoming Master's student in Machine Learning for Artificial Intelligence at the University of Rouen Normandy.
At InterDigital, I investigate an alternative to neural-network loop filtering for VVC (H.266) by predicting which decoder-side filters to apply to reduce compression and quantization artifacts. I have authored and filed more than 10 patent applications, and we are preparing a publication. At Enedis, I worked on generative modeling and large-scale pipelines for high-frequency electricity-grid time series.
Seeking a six-month research internship starting in April 2027
Selected Work.
Research is where I turn difficult questions into working systems.
Explore publications, independent research, and engineering projects.
06.2026
Selected as one of 100 participants from 650+ applicants for a Yann LeCun-sponsored 24-hour challenge bringing together students from France's leading engineering schools and universities; placed among the top 5 of 25 teams. Built a geometry-aware joint-embedding EEG SSL system with SIGReg frozen-transfer baselines and Riemannian analysis of SPD covariance representations.
05.2026
A lightweight wrapper for query-efficient black-box attacks that uses early class-confidence signals to reduce class drift and commit to promising adversarial targets.
2026
An official PyTorch Ecosystem project for visualizing neural-network architectures. I maintain the project and designed and implemented its MCP integration for generating editable diagrams from model definitions.
2025
Co-founded and lead a six-person team behind a real-time customization tool used by more than 30,000 people daily.
No matching work.
Try another filter or search term.