Publications / Papers

Papers, preprints, and research artifacts.

Research outputs across LLM safety, tamper-resistance evaluation, and efficient training methods.

VenueSort

Showing 8 of 8 publications.

ConferenceOralKDD2026

TamperBench: Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering

Saad Hossain, Tom Tseng, Punya Syon Pandey, Samanvay Vajpayee, Matthew Kowal, Nayeema Nonta, Samuel Simko, Stephen Casper, Zhijing Jin, Kellin Pelrine, Sirisha Rambhatla

OpenReview Code

A systematic framework for stress-testing LLM safety under fine-tuning and tampering.

LLM safetyTamper resistanceBenchmarkingEvaluationKDD

BibTeX

@inproceedings{
hossain2026tamperbench,
title={TamperBench: Systematically Stress-Testing {LLM} Safety Under Fine-Tuning and Tampering},
author={Saad Hossain and Tom Tseng and Punya Syon Pandey and Samanvay Vajpayee and Matthew Kowal and Nayeema Nonta and Samuel Simko and Stephen Casper and Zhijing Jin and Kellin Pelrine and Sirisha Rambhatla},
booktitle={KDD 2026 Datasets and Benchmarks Track (Cycle 2)},
year={2026},
url={https://openreview.net/forum?id=R5TNXfdPn8}
}

ConferenceNeurIPS2025

SubTrack++: Gradient Subspace Tracking for Scalable LLM Training

Sahar Rajabi, Nayeema Nonta, Sirisha Rambhatla

Paper OpenReview Slides Code

A gradient-subspace tracking method for scalable LLM training that reduces wall-time while preserving the reduced memory footprint.

OptimizationLLM trainingGradient subspacesEfficiency

BibTeX

@inproceedings{rajabi2025subtrack++,
  title={SubTrack++: Gradient Subspace Tracking for Scalable LLM Training},
  author={Rajabi, Sahar and Nonta, Nayeema and Rambhatla, Sirisha},
  booktitle={39th Conference on Neural Information Processing Systems (NeurIPS 2025)},
  year={2025}
}

WorkshopICLR Workshop on Deep Generative Model in Machine Learning2026

Principled Randomized Exploration of Gradient Subspaces for Efficient LLM Training

Sahar Rajabi, Nayeema Nonta, Sirisha Rambhatla

Workshop paper on randomized exploration of gradient subspaces for efficient LLM training.

Gradient subspacesEfficient trainingGenerative modelsICLR

BibTeX

@inproceedings{rajabi2026principled,
  title={Principled Randomized Exploration of Gradient Subspaces for Efficient LLM Training},
  author={Rajabi, Sahar and Nonta, Nayeema and Rambhatla, Sirisha},
  booktitle={ICLR 2026 2nd Workshop on Deep Generative Model in Machine Learning: Theory, Principle and Efficacy},
  year={2026}
}

WorkshopICLR Workshop on Geometry-grounded Representation Learning and Generative Modeling2026

Random but Right: A Geometric Explanation for Efficient LLM Training

Sahar Rajabi, Nayeema Nonta, Sirisha Rambhatla

Workshop paper connecting randomized gradient subspaces with geometric structure in efficient LLM training.

LLM trainingGeometryOptimizationICLR

BibTeX

@inproceedings{rajabi2026random,
  title={Random but Right: A Geometric Explanation for Efficient LLM Training},
  author={Rajabi, Sahar and Nonta, Nayeema and Rambhatla, Sirisha},
  booktitle={ICLR 2026 Workshop on Geometry-grounded Representation Learning and Generative Modeling},
  year={2026}
}

WorkshopICLR Workshop on Test-Time Updates2026

Refusal-Orthogonal Gated Editing for Safer Localized LLM Adaptation

Nayeema Nonta, Sirisha Rambhatla

Workshop paper on safer localized LLM adaptation using refusal-orthogonal gated editing.

AI safeguard protectionRefusal geometryICLRActivation editingEfficient fine-tuning

BibTeX

@inproceedings{nonta2026refusal,
  title={Refusal-Orthogonal Gated Editing for Safer Localized LLM Adaptation},
  author={Nonta, Nayeema and Rambhatla, Sirisha},
  booktitle={ICLR 2026 Third Workshop on Test-Time Updates (Main Track)},
  year={2026}
}

WorkshopICLR Workshop on Representational Alignment2026

Safe Downstream Adaptation of LLMs via Refusal-Orthogonal Gated Editing

Nayeema Nonta, Sirisha Rambhatla

Workshop paper on downstream LLM adaptation with refusal-orthogonal gated editing.

AI safeguard protectionRefusal geometryICLRActivation editingEfficient fine-tuning

BibTeX

@inproceedings{nonta2026safe,
  title={Safe Downstream Adaptation of LLMs via Refusal-Orthogonal Gated Editing},
  author={Nonta, Nayeema and Rambhatla, Sirisha},
  booktitle={ICLR 2026 Workshop on Representational Alignment (Re $\{$$\backslash$textasciicircum$\}$ 4-Align)},
  year={2026}
}

WorkshopICLR Workshops2026

TamperBench: A Systematic Framework to Stress-Test LLM Safety Under Fine-Tuning and Tampering

Saad Hossain, Tom Tseng, Punya Syon Pandey, Samanvay Vajpayee, Matthew Kowal, Nayeema Nonta, Samuel Simko, Stephen Casper, Zhijing Jin, Kellin Pelrine, Sirisha Rambhatla

OpenReview

Presented across ICLR 2026 workshops on monitoring ML models under drift, agents in the wild, and recursive self-improvement.

LLM safetyTamper resistanceBenchmarkingEvaluationKDD

PreprintarXiv2025

Randomized Gradient Subspaces for Efficient Large Language Model Training

Sahar Rajabi, Nayeema Nonta, Samanvay Vajpayee, Sirisha Rambhatla

Paper

An analysis of gradient-space dynamics with randomized algorithms for memory-efficient LLM pretraining.

LLM trainingMemory efficiencyRandomized algorithmsOptimization

BibTeX

@article{rajabi2025randomized,
  title={Randomized Gradient Subspaces for Efficient Large Language Model Training},
  author={Rajabi, Sahar and Nonta, Nayeema and Vajpayee, Samanvay and Rambhatla, Sirisha},
  journal={arXiv preprint arXiv:2510.01878},
  year={2025}
}