CV
Education
- 2025 – 2028 (expected): PhD in Computer Science
- Sorbonne Université, ISIR (MLIA team), Paris
- Interpretability, steering and hallucination mitigation in vision–language(–action) models
- Advisors: Matthieu Cord, Alasdair Newson
- 2022 – 2023: Engineering Program (GPA: 4.0/4.0)
- Télécom Paris, Palaiseau
- 2019 – 2023: Engineering Program (Ingénieur Polytechnicien)
- École Polytechnique, Palaiseau
- 2017 – 2019: B.Sc. Electrical Engineering (ranked 2nd/120) & Mathematics (second major)
- Amirkabir University of Technology (Tehran Polytechnic), Tehran
Publications
(* equal contribution)
- P. Khayatan, J. Parekh, A. Dapogny, M. Shukor, A. Newson, M. Cord. When Prompts Override Vision: Prompt-Induced Hallucinations in LVLMs. NeurIPS 2026; also at NeurIPS 2026 Workshops (VLM4RWD). [paper] [code] [project]
- P. Khayatan*, S. Meziane*, J. Parekh*, M. Cord. DiMaS: Distribution Matching for Steering Vision-Language-Action Models. ECCV 2026 Workshops (eXCV). [paper] [code] [project]
- J. Parekh*, P. Khayatan*, M. Shukor, A. Dapogny, A. Newson, M. Cord. Learning to Steer: Input-dependent Steering for Multimodal LLMs. NeurIPS 2025. [paper] [code]
- P. Khayatan*, M. Shukor*, J. Parekh*, A. Dapogny, M. Cord. Analyzing Finetuning Representation Shift for Multimodal LLMs Steering. ICCV 2025. [paper] [code]
- J. Parekh, P. Khayatan, M. Shukor, A. Newson, M. Cord. A Concept-Based Explainability Framework for Large Multimodal Models. NeurIPS 2024. [paper] [code]
- Y. Yin, P. Khayatan, É. Zablocki, A. Boulch, M. Cord. ReGentS: Real-World Safety-Critical Driving Scenario Generation Made Stable. ECCV 2024 Workshops (W-CODA). [paper] [code]
Research experience
- Mar 2024 – Sep 2025: Research Engineer
- Sorbonne Université, ISIR (MLIA team), Paris — supervisor: Matthieu Cord
- Concept-based explainability of large multimodal models via dictionary learning (NeurIPS 2024)
- Analysis of fine-tuning representation shift and training-free concept steering of MLLMs (ICCV 2025)
- Input-dependent steering with a lightweight learned steerer for safety and hallucination mitigation (NeurIPS 2025)
- Jul 2023 – Jan 2024: Research Intern
- Valeo.ai, Paris — supervisors: Éloi Zablocki, Alexandre Boulch
- Adversarial generation of safety-critical driving scenarios for motion-planner stress testing, with a differentiable simulator (ECCV-W 2024). [code]
- Mar 2022 – Aug 2022: Research Intern
- EPFL, VITA Lab, Lausanne — supervisors: Alexandre Alahi, Taylor Mordan
- Clothing keypoint localization and category classification; released an OpenPifPaf plugin for DeepFashion. [code]
- Jun 2021 – Aug 2021: Research Intern
- Amadeus
- ML-based intrusion detection for cloud security
Skills & languages
- Programming: Python, PyTorch, Hugging Face Transformers, C++, Git, Linux, SLURM / HPC clusters
- Research: Vision–language models, interpretability (dictionary learning, concepts), activation steering, preference optimization (DPO)
- Languages: Persian (native), English (fluent, TOEFL 105), French (fluent, TCF C1)
Honors, teaching & outreach
- 2023: Represented Télécom Paris at SWERC (ICPC regional programming contest), Milan.
- 2019 – 2022: Eiffel Excellence Scholarship (French Ministry for Europe and Foreign Affairs); Fulgence Bienvenüe Scholarship (École Polytechnique).
- 2019 – 2021: École Polytechnique: TA in Probability Theory (2020); Fun MOOC math content for high-schoolers.
- 2016 – 2017: Bronze Medal, IMC Global Competition (Iran’s team, 2016); top 0.002% in Iranian University Entrance Exam (2017).