Research

I work on reinforcement learning for reasoning under verifiable rewards. I build executable verifiers that act as reward oracles, and study what reinforcement learning teaches a model that supervised imitation cannot, or actively undoes. I want to understand what a reward signal instils beyond the metric it scores, and to turn that into better training methods.

Papers

Findings of EMNLP 2026

Counterexample Generation via Per-Theorem Symbolic Verifiers: When Imitation Hurts and Reinforcement Repairs

Omar Farouk Zouak, Houssam Eddine Boukhalfa, Lakehal Soumaya, Shiv Katiyar, Samia Nefti-Meziani

Track: NLP and Symbolic Reasoning. First author.

Released SymCE: 4,707 false algebra and real-analysis conjectures, each with an executable verifier that is also the reward function.
Imitation alone collapses true-theorem recognition from 0.27 to 0.00. RL with a sparse outcome reward repairs it, to 0.66, above the base model.
Sparse and dense rewards match in-domain but differ by 33 points on a held-out calibration probe.
The 4B model beats every evaluated 7B open-weights math specialist and stays competitive with six frontier APIs.

Code and verifiers Project page

Show BibTeX

@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 Soumaya, Lakehal and Katiyar, Shiv and Nefti-Meziani, Samia},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026}
}

Under review, EACL 2027

What the Query Remembers: Query Abstraction Governs Lexical and Dense Retrieval of Classical Arabic Poetry

Omar Farouk Zouak, Selma Mani, Larbi Said Chikh, Raid Abderrezak Ouahioune

Under review, ACL Rolling Review, August 2026 cycle.

Nine dense encoders and BM25 on two corpora. Which retriever wins depends on how much of the poem’s wording the query still carries.

Preprint, arXiv

AgrI Challenge: A Data-Centric AI Competition for Cross-Team Validation in Agricultural Vision

Mohammed Brahimi, Karim Laabassi, …, Omar Farouk Zouak

Preprint, arXiv:2603.07356, March 2026.

Twelve teams collected 50,673 field images of six tree species, released publicly. Trained on one team’s data, models show validation-test gaps of up to 16.20%; training across teams cuts them below 3%.

arXiv Challenge and dataset

Show BibTeX

@misc{brahimi2026agrichallengedatacentricai,
  title         = {AgrI Challenge: A Data-Centric AI Competition for Cross-Team Validation in Agricultural Vision},
  author        = {Mohammed Brahimi and Karim Laabassi and Mohamed Seghir Hadj Ameur and Aicha Boutorh and Badia Siab-Farsi and Amin Khouani and Omar Farouk Zouak and Seif Eddine Bouziane and Kheira Lakhdari and Abdelkader Nabil Benghanem},
  year          = {2026},
  eprint        = {2603.07356},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2603.07356}
}

Doctoral work

My doctoral topic at ENSIA is physics-informed machine learning: ensemble methods for physics-informed models, building on surrogate models for solar magnetoconvection.

Also on Google Scholar and ORCID.