Skip to content

Mahmoud Zaher

I'm currently an AI research engineer building reliable, trustworthy and explainable AI systems. My recent work spans Arabic NLP, speech, and model interpretability, with three papers accepted at ArabicNLP 2026. I'm always looking for more challenging problems and exploring new research directions.

CV (PDF) Email GitHub Hugging Face LinkedIn

Highlights

  • 1st place on unseen targets (and 2nd on seen targets) in the StanceEval-2026 Arabic stance-detection shared task, out of 22 teams (average F1 0.940 and 0.896).
  • 2nd place in both sentence-level tracks of the BAREC 2026 Arabic readability shared task, as sole author (quadratic weighted kappa 85.0 and 85.2 against an 82.5 baseline).
  • Three papers at ArabicNLP 2026, co-located with EMNLP 2026 in Budapest (ACL Anthology links).
  • 3rd place, 2024 Arab IoT & AI Challenge — qualified to present at GITEX Global 2024, Dubai (details).

Selected publications

  1. ZHR at StanceEval-2026: Surface-Matched Synthetic Demonstrations and Metric-Aware Decoding for Arabic Cross-Target Stance Detection

    M. Zaher, M. Alshrief, M. Bahgat

    ArabicNLP 2026 @ EMNLP 2026, Budapest, Hungary · Oct 2026

    1st place on unseen targets · 2nd on seen targets (22 teams)

    Paper (opens in new tab) Code (opens in new tab) Models (opens in new tab) Results (opens in new tab)

  2. ZHR at BAREC Shared Task 2026: Ordinal-Objective Ensembles with Threshold Calibration and Silver-Label Pretraining for Arabic Readability Assessment

    M. Zaher

    ArabicNLP 2026 @ EMNLP 2026, Budapest, Hungary · Oct 2026

    2nd place in both sentence-level tracks · sole author

    Paper (opens in new tab) Code (opens in new tab) Results (opens in new tab)

  3. The Nile at NADI 2026 Subtask 1.1: A Two-Phase Acoustic Fine-Tuning Strategy for Robust Dialectal Arabic Speech Recognition

    M. Alshrief, M. Bahgat, M. Zaher, S. Selim

    ArabicNLP 2026 @ EMNLP 2026, Budapest, Hungary · Oct 2026

    Paper (opens in new tab)

All publications →

Open source

  • DLA Suite (author; document layout analysis benchmark) — benchmarked 58 layout configurations across 23 repositories and built a label-free validator for missing text, duplicate regions, and reading-order errors.
  • AgentScope (agent framework for LLMs) — fixed failures on empty streamed Anthropic tool calls and duplicate tool-call IDs in Ollama.
  • SynapseKit (application framework for LLMs) — repaired the Mistral chat and embedding integrations after breaking changes in the mistralai client 1.x and 2.x.
  • OpenProof (evidence-review system) — exposed a text-recognition failure that attached footnotes to figures by proximity alone, and added a review rule that checks the footnote marker and text first, which the project adopted.

Get in touch

I'm open to graduate research opportunities and collaborations — especially in interpretability and AI safety. Reach me through the contact form, by email at maredazaher@gmail.com, or via my CV.