Skip to content
View CedricCaruzzo's full-sized avatar

Block or report CedricCaruzzo

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
CedricCaruzzo/README.md

Cédric Caruzzo

AI Research Scientist at Lunit. M.S. in Artificial Intelligence from KAIST.

I study how foundation models generalize, adapt to new information, and use evidence to guide reliable behavior. My work spans representation learning, post-training and controlled evaluation, with biomedical systems as a demanding application domain.

Selected research

Deceptive Grounding

Entity-attribution failures in clinical retrieval-augmented generation: genuine evidence about one medical entity can be used to answer a question about another. First-author study connecting controlled experiments with evaluation in a clinical system.Entity-attribution failures in clinical retrieval-augmented generation: genuine evidence about one medical entity can be used to answer a question about another. First-author study connecting controlled experiments with evaluation in a clinical system.

Paper · Project page

MoE routing and self-distillation

Separating routing divergence from its contribution to model behavior through exact decompositions and controlled analyses. The repository includes frozen research artifacts, deterministic reductions and CPU-based replay.

Paper · Code · Project page

CellPainTR

Cross-dataset representation learning for Cell Painting, combining source-context tokens with self-supervised and contrastive learning.

Paper · Code · Project page

MazeBench Control Center

An independent experimental platform built on the MazeBench engine, with trajectory replay, context-compaction logs and branchable continuations. I use it to study agent behavior through paired-history diagnostics and controlled changes to observations and memory.

Latest experimental write-up · Code

Current work

At Lunit, I lead medical language-model post-training and evaluation, including training-data synthesis, self-distillation, distributed execution, task-specific benchmarks and hospital collaborations. I also work on self-supervised Vision Transformers for whole-slide pathology.

Independently, I am investigating when an agent's correct explanation changes its next action, and what must be preserved when its context is compressed.

Website · Google Scholar · LinkedIn · Email

In the fields of observation, chance favors only the prepared mind.

Pinned Loading

  1. routing-divergence-moe routing-divergence-moe Public

    Code and data for Routing Divergence Is Not Evidence of Behavioral Influence in Same-Weight MoE Self-Distillation.

    Python

  2. mazebench-control-center mazebench-control-center Public

    Local control center for auditable MazeBench experiments, replay, analytics, and model-harness research

    Python

  3. pannuke-segmentation pannuke-segmentation Public

    An open-source UNet-based pipeline for nuclei segmentation in histopathology images using the PanNuke dataset. It features an interactive web app for easy data visualization and handling, making AI…

    Python 11 5

  4. Racial-Bias-Removal-for-Face-Age-Progression Racial-Bias-Removal-for-Face-Age-Progression Public

    This project introduces a method to mitigate demographic bias in generative face aging models. We adapt the StyleGAN-based SAM model by adding a race-consistency loss, enforced by the DeepFace clas…

    Python

  5. Neural-Style-Transfer Neural-Style-Transfer Public

    Implementation of Neural Style Transfer based on the original paper by Gatys et al., with adaptations from Deep Learning with Python by François Chollet. This notebook allows you to experiment with…

    Jupyter Notebook 4 2