Photo of Amr Hegazy

Amr Hegazy

MSc Informatics student · Technical University of Munich

Munich, Germany

I'm a first-year MSc Informatics student at the Technical University of Munich. My research is about making large models controllable and efficient. I work on inference-time activation steering for LLM safety, recurrent-depth transformers that recover expressivity without the memory cost of unique layers, and formal safety guarantees for LLM-controlled robots via reachability analysis.

I completed my B.Sc. in Computer Science and Engineering at the German University in Cairo, graduating with highest honors (top 5% of my class), and wrote my bachelor's thesis at TUM Heilbronn on safety-guaranteed use of foundation models in robotics (grade: A+). Before that, I spent two years as a machine learning researcher at Siemens DISW, and I worked as an AI & automations engineer at ResearchPreneurs.

LLM safety & steeringEfficient transformersRecurrent depthSafe roboticsReachability analysis

News

Publications

See Google Scholar for the latest list.

  1. 2026

    Gated Recurrent Transformers: Expressive Depth through Recurrent Modulation

    Amr Hegazy, Amr Alanwar, Mostafa Elhoushi

    Preprint, arXiv:2608.15062

    A recurrent-depth transformer: a single shared core iterated R times between fixed prelude and coda blocks, with a lightweight GRU-style update gate that keeps per-iteration specialization while reusing weights.

  2. 2026

    Guiding Giants: Lightweight Controllers for Weighted Activation Steering in LLMs

    Amr Hegazy, Mostafa Elhoushi, Amr Alanwar

    6th Workshop on Trustworthy NLP (TrustNLP @ ACL 2026)

    A small trainable controller reads intermediate activations and predicts a global scale plus per-layer weights for a refusal-direction steering patch. It raises refusal rates on harmful prompts without touching the LLM's weights.

  3. 2026

    Meta-heuristic Adaptive Exploration Strategy Through Routing Optimization

    A. Mostafa, Amr Hegazy, F. Metwally, M. Wael, Y. Elbrolosy, Z. Abdelrahman, et al.

    ICEENG 2026, International Conference for Innovations in Power and Control

  4. 2025

    Online Data-Driven Reachability Analysis using Zonotopic Recursive Least Squares

    Alireza Naderi Akhormeh, Amr Hegazy, Amr Alanwar

    Preprint, arXiv:2509.17058

  5. 2025

    Safe LLM-Controlled Robots with Formal Guarantees via Reachability Analysis

    Ahmad Hafez, Alireza Naderi Akhormeh, Amr Hegazy, Amr Alanwar

    Preprint, arXiv:2503.03911

    A data-driven reachability-analysis safety layer that formally verifies LLM-issued robot commands without needing an exact system model.

  6. 2025

    AI/ML Based Prediction of FPGA Synthesis Features From RTLs

    A. Adham, A. Gasser, Amr Hegazy, M. Wael, O. ElHarridy, S. Tarek, et al.

    37th International Conference on Microelectronics (IEEE ICM 2025)

    ML models that predict RAM, DSP, and FSM synthesis features directly from RTL, bypassing the traditional synthesis flow. This work came out of my time at Siemens DISW.

  7. 2025

    Keyphrase Trees: Enhancing LLM Prompts with Hierarchical Topic Extraction

    Amr Hegazy, Mohamed Abdelkarim, Reem ElAdawi

    3rd Cognitive Models and Artificial Intelligence Conference (AICCONF 2025)

    Documents are split into segments and clustered semantically into a hierarchy, which then guides an LLM's keyphrase extraction.

Experience

Education

Projects

More on GitHub, including Haskell, Prolog, and systems projects.

Academic Service

  • Reviewer, ICML 2026 (AI4GOOD & CompLearn Workshops)
  • Reviewer, IROS 2026
  • Reviewer, AAAI 2025 (Main Conference)

Leadership

  • Machine Learning Head, QuBit GUC (2023–2024)
  • ICPC Trainee, ICPC GUC Community (2022–2026)
  • AWS Solutions Architect Associate training (2024)

Skills

Languages
Python, C/C++, CUDA, Java, JavaScript/TypeScript, SQL, Go, Haskell, Prolog
ML
PyTorch, HuggingFace, vLLM, scikit-learn, OpenCV
Infra
Docker, AWS, Kafka, Spark, Airflow, RabbitMQ, Git
Spoken
Arabic (native), English (fluent), French (B1), German (A2)