Sudeshna Merugu

Research Assistant @Carnegie Mellon University - School Of Computer Science - Language Technologies Institute

Pittsburgh, PA, US

AI Text Detection ResearchAssisting with on going research of conducting extensive empirical evaluation of AI-generated text detection systems across multiple language models and domains. A comparative analysis of detection frameworks including RADAR, WILD, T5-Sentinel, LogRank, Binoculars, and Fast-DetectGPT.1. Evaluating detector performance across multilingual tasks including summarization, QA, and dialogue in EN, ES, FR, ZH2. Analyzing detection patterns in domains like code generation, abstract writing, translation, academic paper review writing3. Investigating model-specific detection characteristics across LLama-3, GPT-4, Mistral, and Phi-34. Exploring evaluation detection robustness against plain prompting, adversarial prompting, and rewriting strategies.5. Understanding the emphasis on TPR@FPR over AUROC for practical applications in academic settings.6. Tried adversarial attacking using different models like OpenReviewer and AuthorMist. Conducted extensive research on AuthorMist\'s capabilities against existing paraphrasing models/prompts like paraphrasing prompting and models like Qwen, Dipper and Supervised Fine-Tuning models.8. Observed evading capabilities in evading detector models using AuthorMist compared to the other three models for paraphrasing.9. Currently working on fine-tuning AuthorMist itself, using Grouped Reinforcement Policy Optimization (GRPO) on the original checkGPT data.

Indian Institute of Technology, Dharwad, India

Bachelor of Technology - BTech, Computer Science

Carnegie Mellon University School of Computer Science

Master of Science - MS, Artificial Intelligence