Portrait of Cédric Caruzzo

Cédric Caruzzo

AI Research Scientist at Lunit

Learning, adaptation, and reliable model behavior

Post-training · Safety and alignment · Biomedical AI

Email Google Scholar GitHub LinkedIn
Selected research Download CV

Research Profile

I study how models learn, use evidence, and act on what they know. My interests include AI safety and alignment, post-training, and continual learning—especially how a model can revise what it has learned, use that revision in new situations, and preserve what remains valid.

At Lunit, I lead medical language-model post-training and evaluation, and model-side safety work within Lunit's contribution to AI4ALL. My first-author paper Deceptive Grounding, accepted to the NeurIPS 2026 main track, studies a failure where real evidence is attributed to the wrong medical entity.

My background also includes representation learning: at KAIST, I developed CellPainTR for cross-dataset biological imaging, and I now develop self-supervised models for whole-slide pathology at Lunit. Earlier, at Institut Pasteur, I worked on phenotypic drug discovery and infectious-disease research.

Independently, I build open-source tools and run MazeBench experiments to investigate how an agent's observations and memory affect its decisions, and when a correct explanation fails to guide its next action.

Publications and Research Outputs

NeurIPS 2026
Deceptive Grounding: Entity Attribution Failure in Clinical Retrieval-Augmented Generation
Cédric Caruzzo, Donggeun Yoo, Tae Soo Kim
Advances in Neural Information Processing Systems (NeurIPS), 2026

Studies how clinical RAG assigns real evidence to the wrong drug, measures the failure in a 740-pair deployment sample, and evaluates a claim-level entity-attribution check.

Preprint PDF Project Code
arXiv 2025
CellPainTR: Generalizable Representation Learning for Cross-Dataset Cell Painting Analysis
Cédric Caruzzo, Jong Chul Ye
arXiv preprint arXiv:2509.06986, 2025

Learns batch-robust cellular representations across heterogeneous studies and generalizes to an unseen dataset without target-dataset fine-tuning.

PDF Code Project
SLAS Europe 2024 Poster
Cellular Phenotypic Profiling: Combining Live Imaging and Cell Painting Techniques
Soonju Park, Cédric Caruzzo, Nakyung Lee,
Poster presented at SLAS Europe 2024, Barcelona, Spain

Combines longitudinal live-cell imaging with endpoint Cell Painting for richer phenotypic profiling in high-content drug discovery.

Poster
Scientific Reports 2021
Wolbachia detection in Aedes aegypti using MALDI-TOF MS coupled to artificial intelligence
Antsa Rakotonirina, Cédric Caruzzo, Valentine Ballan,
Scientific Reports 11, 21355, 2021

Uses machine learning with MALDI-TOF mass spectrometry to detect Wolbachia infection in Aedes aegypti mosquitoes for vector-control research.

Paper

Selected Experience

Lunit, Seoul
Mar 2025 - Present
AI Research Scientist
  • Led medical-model post-training and evaluation for Chain of Evidence, including the comparative evaluation and adoption of on-policy self-distillation.
  • Lead model-side safety within Lunit's contribution to AI4ALL (ModuAI), defining agent behavior and testing unsafe responses, unnecessary refusals, and faulty tool outputs.
  • Conceived and led Deceptive Grounding, accepted to the NeurIPS 2026 main track.
  • Develop self-supervised Vision Transformers for whole-slide pathology.
KAIST, Kim Jaechul Graduate School of AI
Feb 2023 - Feb 2025
M.S. Researcher, advisor: Prof. Jong Chul Ye

Full research experience

Selected Research

Deceptive Grounding failure structure
NeurIPS 2026 · First author

Deceptive Grounding

Clinical RAG can cite a real source while assigning its evidence to the wrong drug. We measured the failure in one deployment and tested an explicit entity-attribution check.

Clinical RAG Entity attribution Evaluation blind spot
Preprint Project page Code
CellPainTR cross-dataset representation-learning concept
First-author preprint and code · 2025

CellPainTR

A Transformer for Cell Painting representations that preserves biological signal across batch and feature shifts, including generalization to an entirely unseen dataset without fine-tuning.

Cross-dataset OOD No target fine-tuning Open source
Paper Code Project page
MazeBench Control Center showing an agent's reasoning, actions, and exploration history
Independent experiments & open-source software · 2026

MazeBench: From Explanation to Action

Exploratory studies of how observations, memory, and context compression affect agent decisions, including cases where a correct explanation fails to guide the next action. Built a Control Center to replay runs and branch from earlier decisions.

Agent behavior Memory and context Research notebook
Read the experiments Code

Selected Technical Writing

All research and writing