Paris-Saclay, FranceMachine learning researcher · M2 Data Science, Université Paris-Saclay
I train language models against checkers that can prove an answer right or wrong, and study what the models actually learn from them.
Looking for a 6-month research internship from late March 2027Research
I'm finishing an M2 in Data Science (TRIED) at Université Paris-Saclay, co-accredited with Institut Polytechnique de Paris. Before that I completed a five-year engineering degree and a Master 2 in Artificial Intelligence and Data Science at ENSIA in Algiers.
With the University of Birmingham, we showed that supervised fine-tuning can teach a language model the format of correct answers without the reasoning behind them, and that reinforcement learning against symbolic verifiers repairs it. I also built LLM world models that predict how a planning environment changes after each action, and evaluated how they generalise to domains they have never seen.
On the engineering side, I led the AI team at Nojoom.AI that trained Arabic language models up to 42 billion parameters, and I care about making training and inference efficient, from data selection down to GPU kernels.
News
Results
Supervised fine-tuning learned the answer format but lost the logic; RL with verifier rewards brought it back. The 4B model beat every evaluated 7B open math specialist. Paper
Unseen domains
Seen domains
The gain came from penalising repeated errors with a memory of past counterexamples, not from denser reward shaping. Thesis
Mixture-of-experts model depth-extended from Qwen3-30B-A3B (48 → 66 layers, 4.5B active), with an Arabic tokenizer using 28% fewer tokens and 275B tokens of continued pretraining.
Publications
Counterexample Generation via Per-Theorem Symbolic Verifiers: When Imitation Hurts and Reinforcement Repairs
EMNLP 2026 (Findings), Budapest
My role: set up the training pipeline and trained the models (SFT, then GRPO-based RLVR).
@inproceedings{zouak2026counterexample,
title = {Counterexample Generation via Per-Theorem Symbolic Verifiers:
When Imitation Hurts and Reinforcement Repairs},
author = {Zouak, Omar Farouk and Boukhalfa, Houssam Eddine and
Lakehal, Soumaya and Katiyar, Shiv and Nefti-Meziani, Samia},
booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
year = {2026},
note = {arXiv:2610.02444}
}Crafting Culturally Aligned Narratives: Large Language Models for Arabic Children's Story Generation
Muslims in ML Workshop, NeurIPS 2025, San Diego
@inproceedings{boukhalfa2025crafting,
title = {Crafting Culturally Aligned Narratives: Large Language Models
for Arabic Children's Story Generation},
author = {Boukhalfa, Houssam Eddine and Mani, S. and Said Chikh, L.},
booktitle = {Muslims in ML Workshop at NeurIPS 2025},
year = {2025}
}In preparation: a paper on detecting information influence operations on TikTok, from my work at LIX, École Polytechnique.
Thesis
Final-year thesis on counterexample-guided reinforcement learning with verifiable rewards, at ENSIA in collaboration with the University of Birmingham, supervised by Prof. Samia Nefti-Meziani, Dr Shiv Ashutosh Katiyar and Dr Soumaya Lakehal.
Experience
Education
Projects
Quality Takeoffs
AI agents and computer vision read technical drawings, extract and measure their elements (legend tables, fixtures, measurements) and write the takeoff report.
Roman mosaic restoration with GANs
A GAN with a Vision Transformer and contextual attention that restores damaged Roman mosaics.
Reinforcement learning for autonomous racing
A DQN driving agent in simulation, from environment modelling and sensor inputs to reward design and robustness analysis.
Skills
Also: co-founder of the ENSIA Tech Community Club, winner of three national AI hackathons, and supervisor of the student teams that organised the Algerian AI Council (Dec 2024) and the AI for Africa conference series (Dec 2023).
Contact