Omar Khursheed

MATS Research Scholar

I study the behavior and alignment of large language models by training model organisms that demonstrate various forms of misalignment and developing training and inference-time mitigations.

I grew up in Saudi Arabia and studied computer engineering at Aligarh Muslim University. I completed an MS in computer science at UMass Amherst, where I worked with Mohit Iyyer and Brendan O'Connor on gender bias in narrative tropes.

From 2021 to 2025, I was an Applied Scientist on Amazon's Alexa Edge AI team, working on wakeword detection and device-directedness. During the first half of 2026, I was an Anthropic AI Safety Fellow, working with Fabien Roger on the alignment properties of helpful-only models. I am currently a research scholar at MATS.

Publications

(Mis)generalization of Helpful-only Fine-tuning

Khursheed*, Sosis*, Roger · arXiv

2026

Two Centuries of Women’s Voices: Computational Analysis of Gender Participation and Suffrage Debates in British Parliament

Khursheed*, Sawkar*, KhudaBukhsh · Under review

Small Footprint Slimmable Networks for Keyword Spotting

Akhtar, Khursheed, Du, Liu · IEEE ICASSP

Latency Control for Keyword Spotting

Jose, Wang, Strimel, Khursheed, et al. · INTERSPEECH

Tiny-CRNN: Streaming Wakeword Detection

Khursheed, Jose, Kumar, et al. · IEEE ASRU

Analyzing Gender Bias within Narrative Tropes

Gala*, Khursheed*, Lerner, O'Connor, Iyyer · EMNLP Workshop

Writing

Practical Failure Modes in RL Training for LLMs

Notes on rollouts, masking, rewards, optimization, and infrastructure.

Analysis of Whisper-Tiny Using Sparse Autoencoders

A sparse autoencoder analysis of Whisper-Tiny's encoder representations.

Datasets

UK Parliamentary Debates

Speeches from UK Parliament, 1803–2005, with speaker and gender metadata.

TVTropes

Examples of 30K narrative tropes across film, television, and literature.