The question behind all my work: how does a system reach its answer, and is that process faithful to the evidence? A model can get the right answer through a broken process, and accuracy alone will never show it. I build the measurement that exposes the process: in LLM reasoning traces, in multimodal models that search images like people do, in retrieval-grounded systems, and in human gaze itself.
AI Research Scientist at F.initiatives · AIOps Architect at NeoPhi (AI for science, 100M+ indexed papers) · Ph.D. in Computer Science, Université d'Orléans (2024)
| peer-reviewed papers | Outstanding Paper ECCV 2026 Workshop |
LLM decisions analysed in KillBench |
reasoning traces turned into graphs |
eye-tracking participants across 10 museums |
The same idea, measuring the process and not only the outcome, applied across very different fields:
mindmap
root((How a system<br/>reaches its answer))
AI safety and fairness
KillBench bias analysis
Red-teaming RAG apps
LLM reasoning
ThinkProbe thought graphs
Reasoning distillation
Multimodal models
Foveated MLLM visual search
Agentic image editing
Health
Health misinformation
Mental-health assistant
Medical image quality
Clinicians' gaze on AI images
Cultural heritage
Attention on paintings
AVAtt dataset
Neurodiversity
Autistic gaze on art
Science
Citation hallucination detection
Literature at 100M+ scale
Efficient AI
Quantization and perception
Sub-1B distilled models
Creative and civic tools
Fashion LoRA for FLUX
Automated video essays
French civic exam coach
| Domain | What I did | Links |
|---|---|---|
| 🛡️ AI safety & fairness | Independent reasoning-level audit of White Circle's KillBench; red-team of my own RAG assistant | KillBench analysis · red-team report |
| 🧠 LLM reasoning | Non-generative profiling of chain-of-thought; reasoning distillation with merge correction | ThinkProb · distillation |
| 👁️ Multimodal models vs. humans | Do MLLMs search a scene the way people do? | MODG |
| 🔬 Scientific literature | Citation-verification cascade that flags hallucinated and phantom citations (NeoPhi, production) | private |
| 🏥 Public health | Health-misinformation triage grounded in medical sources (EU NGI Search) | HeReFaNMi |
| 💬 Mental health | Citation-grounded self-help assistant with crisis guardrails | Enlighten-AI |
| 🩻 Medical imaging | Image quality assessment, knee osteoarthritis grading, clinicians' gaze on real vs. AI-generated scans | papers |
| 🎨 Cultural heritage | Where people look in paintings, and models that predict it | AVAtt · SSLArtScanpath · SP_Gen |
| 🧩 Neurodiversity | Spatial and temporal gaze signatures of autistic adults, artists and neurotypical viewers | Artism |
| ⚡ Efficient AI | How quantization shifts what CNNs "look at"; sub-1B reasoning models | CAM_Quant |
| 👗 Generative AI | FLUX LoRA for garment images; identity-safe agentic portrait editing | Flux LoRA · APEX |
| 🎬 Media & education | Topic-to-video essay pipeline; coach for the French naturalisation civic exam (FR / EN / AR) | YT_essay_Gen · Examen civique |
|
White Circle's KillBench asks models to choose who dies in life-or-death dilemmas. I measured who they pick and how they reason first.
|
Turns a reasoning trace into a typed thought graph, with no generative model at any step, and scores a 22-metric profile across breadth, depth, structure, metacognition and efficiency.
|
|
🏆 Outstanding Paper, ECCV 2026 Human-Inspired Computer Vision Workshop Multimodal LLMs search a scene one foveated glimpse at a time and decide when to stop. They match or beat humans on the answer, yet their search is non-human and far more self-consistent than people are with each other.
|
Distils reasoning into Qwen3.5-0.8B. After every epoch, 7 merge methods (TIES, DARE, SLERP, DELLA…) compete against plain SFT on GSM8K / MMLU / ARC, and the winner continues.
|
|
Health-claim triage as Trustworthy / Doubtful / Fake, grounded in retrieved medical sources. EU NGI Search project with IULM University and ISASI-CNR; I led the technical team of 4 and was last author of the AIxIA paper on chunking strategies.
|
An MLLM plans each photo edit and judges the result; an ArcFace identity gate must agree before an edit is accepted, so the portrait stays recognisably you.
|
👁️ Visual attention & gaze toolkit (click to expand)
| Project | What it is | |
|---|---|---|
| FDISS | Foveal Disc IoU Scanpath Score, a perceptually grounded scanpath metric | pip install fdiss · ETRA 2026 |
| scanpath_nlp_metrics | Compares scanpaths by what was seen, through VLM descriptions and NLP metrics | pip install scanpath_nlp_metrics · ETRA 2026 |
| Artism | Autistic, artist and neurotypical gaze on 30 paintings: 2,091 scanpaths, 95,282 fixations | preprint |
| AVAtt | Visual-attention dataset for paintings, scanpaths and saliency maps | ETRA 2024 |
| SSLArtScanpath | Barlow Twins self-supervised scanpath prediction | CVPR Workshops 2022 |
| SP_Gen | Stochastic, domain-adaptive scanpath generation for paintings | CBMI 2022 |
| CAM_Quant | How int16 / int8 quantization shifts the class activation maps of 6 CNNs | IPTA 2024 |
| GBVS_Pytorch | PyTorch port of Graph-Based Visual Saliency (Harel, Koch & Perona, 2006) |
🛠️ Tools & side builds (click to expand)
| Project | What it is |
|---|---|
| Enlighten-AI | Citation-grounded RAG self-help assistant with timestamped sources, crisis guardrails and a red-team report (OWASP LLM Top 10, MITRE ATLAS) |
| llm-tuning-lab | Claude Code plugin: SFT, LoRA / QLoRA, continual pretraining, DPO / ORPO / KTO and merging, run with controls, error bars and significance tests |
| YT_essay_Gen | Topic in, narrated and captioned video essay out: Wikipedia research, LLM thesis and script, TTS, LLM-judged visuals |
| FR-Examen-civique | Full-stack coach for the French civic exam: mock exams, spaced repetition, mind maps, FR / EN / AR explanations |
| Flux1-Shnell-LoRA-JLSCOM | FLUX.1-schnell LoRA for garment images · model on 🤗 |
| mlflow_pytorch_exp | One command to scaffold an MLflow + PyTorch experiment project · pip install mlflow-pytorch-exp |
- 🏆 M. A. Kerkouri, M. Tliba, A. Chetouani, U. Bagci, A. Bruno. Matched Outcomes, Divergent Gaze: How Foveated MLLMs Search Compared to Humans. ECCV 2026 Workshops (HCV), Outstanding Paper Award. arXiv
- M. A. Kerkouri et al. ThinkProbe: Beyond Accuracy, Structural Profiling of Open-Ended LLM Reasoning Traces via Non-Generative Thought Graphs. Preprint, 2026. arXiv
- M. A. Kerkouri, M. Tliba, B. Wang, A. Chetouani, U. Bagci, A. Bruno. What They Saw, Not Just Where They Looked: Semantic Scanpath Similarity via VLMs and NLP Metrics. ETRA 2026.
- M. A. Kerkouri, M. Tliba, Z. Sellam, C. Distante, A. Bruno, A. Chetouani. Closing the Foveal Gap: Perceptually Grounded Scanpath Comparison with Disc IoU. ETRA 2026.
- A. Bruno, P. L. Mazzeo, A. Chetouani, M. Tliba, M. A. Kerkouri. Insights into Classifying and Mitigating LLMs' Hallucinations. Preprint, 2023. arXiv
- M. A. Kerkouri, M. Tliba, A. Chetouani, A. Bruno. Quantization Effects on Neural Networks Perception. IPTA 2024. arXiv
- M. Tliba, M. A. Kerkouri, et al. SATSal: A Multi-Level Self-Attention Based Architecture for Visual Saliency Prediction. IEEE Access, 2022.
- M. Tliba, M. A. Kerkouri, A. Chetouani, A. Bruno. Self-Supervised Scanpath Prediction Framework for Painting Images. CVPR Workshops 2022 (equal contribution). paper
Full list on Google Scholar. Collaborations with the Bagci Lab (Northwestern University) and the Merabet Lab (Harvard Medical School).
Evaluation & statistics: benchmark and metric design · LLM-as-judge · calibration · mixed-effects models · signal detection theory · bootstrap CIs · multiple-testing correction · human annotation studies