Yahya Masri

Yahya Masri

CS @ GMU

I am an undergraduate majoring in Computer Science at George Mason University, advised by Prof. Chaowei Yang and Dr. Zifu Wang. My future research direction are pragmatic interpretability and the safety and value alignment of language models (i.e., scalable oversight, deception alignment and super alignment). My current research focuses on evaluating language models on domain-specific tasks, digital twins, and their intersection. You can find my cv here.

My answers to the Hamming question ("What are the most important problems [that you should probably work on]?"):

  • How can we rigorously evaluate alignment and safety properties of language models beyond benchmark performance?
  • How can interpretability methods be made actionable for safety, rather than purely descriptive?
  • How can digital twins serve as testbeds for alignment and safety research?

Feel free to reach out! Email: ymasri[at]gmu[dot]edu or yahya[dot]masri[at]yahoo[dot]com

Blog (Thoughts and Writings)

What is AI Alignment?

AI Allignment is the study and evaluation of a model's behavior remaining consistent with human intentions, ethical norms, and safety constraints

Red-Teaming: This approach generates adversarial scenarios that induce unaligned outputs or actions in AI systems in order to evaluate the robustness of their alignment under intentional pressure. Included human adversarial training and automated adversarial training.

Harmlessness & Misuse Testing:

Truthfulness and Honesty Evaluations:

Instruction-Following and Jailbreak Resistance:

High-Stakes β€œAgentic” Scenario Evaluations:

Other Evaluation Approaches and Considerations: Chain-of-thought and rationale inspection,

News

More news
  • 2024.12: πŸŽ™οΈ I co-authored a poster on automating map-making via RAG-enhanced geographic information extraction with LLMs presented at the AGU24.
  • 2024.11: πŸŽ™οΈ I presented updated findings on context-aware location extraction at the NSF Spatiotemporal Innovation Center (STC) Industry Advisory Board at George Mason University.
  • 2024.07: πŸŽ™οΈ I presented research on conflict incident classification using a BERT model at the International Symposium of Spatiotemporal Data Science 2024.
  • 2024.05: πŸŽ™οΈ I presented findings on context-aware location extraction at the STC Industry Advisory Board meeting hosted at Harvard University.

Publications

* denotes equal contribution, Ξ± denotes core contributors, and † denotes corresponding author

CCDS
scraper
Automating Data Collection to Support Conflict Analysis: Scraping the Internet for Monitoring Hourly Conflict in Sudan

Yahya Masri*, Anusha Srirenganathan Malarvizhi*, Samir Ahmed*, Tayven Stover*, Zifu Wang†, Daniel Rothbart, Mathieu Bere, David Wong, Dieter Pfoser, and Chaowei Yang

[Paper]

This paper introduces Data: an open-access, source-attributed Sudan conflict news corpus collected hourly, comprising 6,946 archived articles (as of Oct 25, 2024) from national, regional, and international outlets, with full text and URLs preserved for transparency and verification.

IJDE
ijde
Optimizing context-based location extraction by tuning open-source LLMs with RAGπŸ“š 3 citations

Zifu WangΞ±, Yahya MasriΞ±, Anusha Srirenganathan Malarvizhi, Tayven Stover, Samir Ahmed, David Wong, Yongyao Jiang, Yun Li, Mathieu Bere, Daniel Rothbart, Dieter Pfoser, David Marshall, and Chaowei Yang†

[Paper]

This paper makes the first systematic investigation of context-based location extraction with open-source LLMs, demonstrating that RAG substantially improves multi-incident, multi-location extraction accuracy.

MDPI
comparative
Comparative Analysis of BERT and GPT for Classifying Crisis News with Sudan Conflict as an ExampleπŸ“š 1 citations

Yahya MasriΞ±, Zifu WangΞ±, Anusha Srirenganathan Malarvizhi, Samir Ahmed, Tayven Stover, David WS Wong, Yongyao Jiang, Yun Li, Qian Liu, Mathieu Bere, Daniel Rothbart, Dieter Pfoser, and Chaowei Yang†

[Paper]

We introduce a systematic evaluation framework for Sudan conflict event classification that benchmarks GPT-style LLM prompting against BERT/BERT-large baselines.

Teaching & Mentoring

  • ASSIP Fellow mentoring (Summer 2025)

Miscellaneous

  • Projects: Glimpse (AI platform that converts product one-liners into cinematic marketing videos), SoccerBot (LLM-powered soccer prediction and analysis app with RAG), BlueTemp (AI platform for predicting sea water temperatures), Crushor (retro platform game)

Educations