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
Research Scientist @ AMIAD
Quentin Guimard
Latest News (Sep. 2026): I have joined AMIAD – French Agency for AI in Defence at École polytechnique (Palaiseau) as a Research Scientist in Trustworthy AI!
I am a Research Scientist in the Research Division of AMIAD – French Agency for AI in Defence, located at the Drahi X-Novation Center (École polytechnique, Palaiseau), where I work within the Trustworthy AI group led by Gianni Franchi. Previously, I was a postdoctoral researcher at the University of Trento within the Multimedia and Human Understanding Group (MHUG) in collaboration with Prof. Elisa Ricci, where I continue to co-supervise Ph.D. students.
During my Ph.D. at Université Côte d'Azur / CNRS (i3S), supervised by Prof. Lucile Sassatelli, I explored the links between visual content, human attention, and emotions in immersive 360° environments—designing uncertainty-aware variational architectures (such as DVMS) to ensure equitable streaming quality across diverse users.
This focus on uncertainty, fairness, and reliability led directly to my current research on the trustworthiness of multimodal and language foundation models (LLMs, VLMs, and diffusion models). My ongoing work combines uncertainty estimation, mechanistic interpretability, and black-box auditing to understand when foundation models fail, why they fail, and how to steer them safely without costly retraining.
I am driven by a fundamental desire to understand how AI models work under the hood. My current research interests span three axes:
- 01 Uncertainty Quantification & Reliability (Knowing When Models Fail): Developing principled uncertainty estimation and hallucination detection methods—such as factoring in factual equivalence in semantic uncertainty—across autoregressive LLMs, Vision-Language Models, and diffusion language models.
- 02 Mechanistic Interpretability & XAI (Understanding Why Models Fail): Exploring internal representations using linear probing, Sparse Autoencoders (SAEs), Concept Bottleneck Models (CBMs), and visual concept decompositions to map latent spaces onto human-understandable concepts.
- 03 Model Auditing & Post-Hoc Steering (Fixing Models Without Retraining): Building automated black-box benchmarks and bias discovery frameworks (like C2B) to expose failure modes, and performing surgical latent-space interventions (like SEM) to mitigate biases and hallucinations at inference time.
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