Conference2026
TamperBench: Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering
A systematic framework for stress-testing LLM safety under fine-tuning and tampering.
University of Waterloo, Critical ML Lab
Nayeema Nonta
ML systems, safe adaptation, and efficient training.
I'm a PhD student at the Critical ML Lab researching LLM systems, with an emphasis on post-training, efficient adaptation, tamper resistance, and reliable model behavior.
Publications / Selected
Recent work.
Selected papers and preprints. Full records include links and BibTeX blocks on the publications page.
Conference2025
SubTrack++: Gradient Subspace Tracking for Scalable LLM Training
A gradient-subspace tracking method for scalable LLM training that reduces wall-time while preserving the reduced memory footprint.
Research / Featured
Research threads at the intersection of safety and efficiency.
A compact view of the questions I am most focused on: how models adapt, how safety boundaries shift, and how training can become cheaper without becoming less reliable.
OngoingLLM adaptationPEFT
Safe and efficient LLM adaptation
Post-training methods that preserve model utility while limiting unsafe drift.
I am interested in adaptation pipelines that make LLMs more useful without making them brittle: parameter-efficient updates, safety-preserving fine-tuning, and evaluation protocols that expose failure modes early.
Post-trainingEvaluation
OngoingRefusalsRepresentation space
Refusal geometry and model safety
Understanding how refusal behavior shifts across model internals and attacks.
This theme studies refusal behavior as a geometric and representational object: where safety boundaries live, how they move during tampering, and which interventions keep them stable.
SafetyMechanistic analysis
PublishedOptimizationSubspace methods
Efficient training and optimization
Gradient-subspace methods for reducing memory and wall-time costs in LLM training.
My work in this area focuses on optimization methods that lower training cost without losing convergence quality, especially low-rank and subspace-aware approaches for large models.
LLM trainingScalability