Rana Muhammad Shahroz Khan

I am a first second year PhD student at UNC Chapel Hill where I work on Efficient ML and NLP. I am grateful to be advised by Dr. Tianlong Chen. Before that I did my Bachelors in Computer Science and Mathematics from Vanderbilt University, where I was advised by Dr. Soheil Kolouri and worked in the intersection of Computer Vision and Optimal Transport.

During the summer of 2023, I was a Research Intern at Lawrence Livermore National Laboratory in the Machine Intelligence Group working under the guidance of Dr. Jay Thiagarajan, Dr. Shusen Liu and Dr. Rushil Anirudh on Bayesian Optimization for Uncertainty Quantification. In the Summer of 2024, I also had the pleasure of working as a ML Research Intern at HighArc under the supervision of Dr. Ardavan Bidgoli and Dr. Manuel Ladron de Guevara.

Email  /  GitHub  /  Google Scholar  /  LinkedIn  /  CV [03/2026]

Experience

• Anthropic — Research Fellow (May 2026 – Present)
• Cisco — Research Intern (May 2025 – Dec 2025)
• HighArc — ML Research Intern (May 2024 – Aug 2024)
• Lawrence Livermore National Laboratory — Research Intern (May 2023 – Aug 2023)

Updates

Research

I am broadly interested in efficient reasoning.

TMS: Trajectory-Mixed Supervision for Reward-Free, On-Policy SFT

We present a simple, drop-in SFT replacement that recovers RL-like retention by training on trajectory-mixed, near-policy targets instead of static labels—eliminating reward models while preventing forgetting.

The Quest for Efficient Reasoning: A Data-Centric Benchmark to CoT Distillation

We present a comprehensive benchmark (DC-CoT) that systematically shows data-centric strategies, especially augmentation like reverse reasoning, are the most effective way to distill strong reasoning abilities into smaller LLMs.

CAR-LoRA: Training Compression-Aware and Robust LoRA Adapters for Evolving LLMs

We present a unified training framework (CAR-LoRA) that learns a single compression-aware and temporally robust LoRA adapter, eliminating the need to retrain separate adapters across different hardware compressions and evolving LLM versions.

ORAL: Prompting Your Large-Scale LoRAs via Conditional Recurrent Diffusion

ORAL leverages conditional recurrent diffusion to instantly craft task‑tuned LoRA adapters that scale to billion‑parameter models, remain compatible with future model updates, and rival—or beat—full fine‑tuning accuracy.

Agents Under Siege: Breaking Pragmatic Multi-Agent LLM Systems with Optimized Prompt Attacks

Agents Under Siege introduces a flow‑optimized, order‑agnostic prompt attack that stealthily threads limited‑bandwidth, high‑latency multi‑agent LLM networks to jailbreak target models, overwhelming modern safety guards and boosting attack success up to seven‑fold.

PortLLM: Personalizing Evolving Large Language Models with Training-Free and Portable Model Patches

We present PORTLLM, a training-free framework that enables seamless knowledge transfer across evolving LLMs, achieving LoRA-level performance with significantly lower computational costs.

LLMGeo: Benchmarking Large Language Models on Image Geolocation In-the-wild

We evaluate multimodal LLMs for image geolocation using a new dataset, showing that fine-tuning helps open-source models approach closed-source performance.

PRANC: Pseudo RAndom Networks for Compacting Deep Models

We propose PRANC, a framework that reparametrizes deep models as a linear combination of frozen random networks, enabling extreme compression, efficient storage, and memory-efficient inference.

Linear Optimal Partial Transport Embedding

We propose the Linear Optimal Partial Transport (LOPT) embedding, enabling faster OPT distance computation and demonstrating its effectiveness in point-cloud interpolation and PCA analysis.

Design and source code from Jon Barron's website