Cédric Caruzzo

• Deceptive Grounding: conceived and led a first-author study of entity-attribution failure in clinical RAG, from the research question and controlled benchmark to analysis, deployment evaluation, and writing. Accepted to the NeurIPS 2026 main track.
• Medical-model post-training: owned post-training and evaluation for Chain of Evidence, contributing to the public L1-16B-A3B and L1-30B-A5B models. Built synthetic-data and tool-use SFT pipelines; implemented and comparatively evaluated on-policy self-distillation, leading its adoption in the production training pipeline.
• Model-side safety: lead this work within Lunit's contribution to AI4ALL (ModuAI), a Kakao-led national consumer-AI initiative. Authored a priority-ordered behavior specification for a medical-information agent and designed tests of unsafe responses, unnecessary refusals, and faulty tool outputs.
• Training and behavior analysis: conducted GRPO, SDPO, and reward-modeling experiments. Diagnosed reward hacking and checkpoint regressions through model outputs, reasoning traces, and evaluation signals, revising data, objectives, templates, or serving configurations accordingly.
• External evaluation: led acceptance-evaluation delivery, from task data and post-training to a checkpoint with a reproducible evaluation harness. Designed rubric-based agent benchmarks and blinded clinician evaluation with hospital partners.
• Pathology models: develop self-supervised Vision Transformers for whole-slide pathology, studying positional encoding, optimization, and transfer. Built fault-tolerant distributed inference for clinical-scale image collections.
• Evaluation lead, Conquer Health: designed HealthBench-based scoring against frontier baselines and a blinded clinician evaluation protocol for the August 2026 hackathon.
• Research culture: lead Lunit's recurring AI paper-study program and write for its official technical publication, including You Could Have Invented Dr.GRPO Yourself.

Deceptive Grounding: studied entity-attribution failure across 13 models and a 740-pair sample from one clinical RAG deployment. Project · Paper · Code

• Developed CellPainTR, a Transformer with source-context tokens and staged self-supervised and contrastive training for batch-robust Cell Painting representations.
• Evaluated out-of-distribution generalization under dataset, batch, and feature shift, including transfer to an unseen dataset without target-dataset fine-tuning.
• Developed ChemPainTR, a dual-encoder framework for bidirectional translation between chemical structures and cellular phenotypes.
• Completed the thesis Development of Foundation Models in the Context of Cell Painting as a KAIST Global Presidential Scholar.

• Built machine-learning classifiers and interpretable feature-extraction methods for MALDI-TOF mass spectrometry.
• Supported Wolbachia detection, mosquito-species identification, and leptospirosis studies with public-health and hospital collaborators.
• Contributed data analysis to the Scientific Reports paper on Wolbachia detection.