SEM: Sparse Embedding Modulation for Post-Hoc Debiasing of Vision-Language Models
Quentin Guimard*, Federico Bartsch*, Simone Caldarella, Rahaf Aljundi, Elisa Ricci, Massimiliano Mancini
CVPR Findings 2026
Postdoc @ UniTrento
Quentin Guimard
Currently in Denver: Presenting our work on post-hoc feature modulation (SEM) at the CVPR 2026 Findings track with my colleagues from Trento!
I am a postdoctoral research fellow at the University of Trento (DISI), where my research focuses on Trustworthy AI. I work within the Multimedia and Human Understanding Group (MHUG) in collaboration with Prof. Elisa Ricci.
During my PhD, supervised by Prof. Lucile Sassatelli, I explored the links between visual content, human attention, and emotions in immersive 360° environments, designing adaptive systems (such as DVMS) to ensure equitable streaming quality across diverse users.
This focus on fairness and reliability led to my current work on the robustness and capabilities of multi-modal foundation models. For example, my ongoing projects involve building frameworks to track and semantically map complex physical object transformations in video, and developing benchmarks to audit the perceptual and reasoning limits of video LLMs in safety-critical scenarios.
I am driven by a fundamental desire to understand exactly how AI models work under the hood. My current research interests span three axes:
- 01 Model Auditing (Black-Box): Developing automated benchmarks, bias detection frameworks, and red-teaming approaches to expose vulnerabilities and failure cases in multi-modal systems.
- 02 Mechanistic Interpretability & Interpretability-by-Design (White-Box): Exploring internal representations using linear probing and Sparse Autoencoders (SAEs), and investigating inherently interpretable architectures like Concept Bottleneck Models (CBMs) to map latent spaces to human-understandable features.
- 03 Bias & Fairness: Leveraging this understanding of models to characterize their biases and build fairer, more transparent systems.
Open Science & Reproducibility I strongly believe that trustworthy AI requires transparent science. I prioritize reproducibility and open-source engineering, an ethos reflected in the public release of frameworks like C2B and SEM, as well as the ACM reproducibility badges awarded to my doctoral research.
Selected Publications
Classifier-to-Bias: Toward Unsupervised Automatic Bias Detection for Visual Classifiers
Quentin Guimard, Moreno D'Incà, Massimiliano Mancini, Elisa Ricci
CVPR 2025
PEM360: A dataset of 360° videos with continuous Physiological measurements, subjective Emotional ratings and Motion traces
Quentin Guimard*, Florent Robert*, Camille Bauce, Aldric Ducreux, Lucile Sassatelli, Hui-Yin Wu, Marco Winckler, Auriane Gros
ACM MMSys 2022
Deep variational learning for multiple trajectory prediction of 360° head movements
Quentin Guimard, Lucile Sassatelli, Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, Alberto Del Bimbo
ACM MMSys 2022 Best Paper 
Open Source & Frameworks
SEM
A post-hoc, zero-shot debiasing framework that operates in a Sparse Autoencoder (SAE) latent space to pinpoint and modulate bias-relevant neurons in Vision-Language Models.
PyTorch Mechanistic Interpretability VLMs Debiasing