Amherst, Massachusetts, United States
• Developed a SELF-DISCOVER method for long-form text style transfer, integrating contrastive style guides with LLMs, achieving a 44% improvement in target style accuracy and enabling human-editable guidelines for publication-grade content.
• Created a benchmark dataset of 600,000 semantically-paired documents across 5 domains (Tech, Entertainment, Finance, Food, Games), leveraging IoU scoring over Google keyword searches and TF-IDF for precise style differentiation.
• Designed and tested zero-shot, few-shot, and SELF-DISCOVER prompting strategies, incorporating BERT embeddings and Longformer models, boosting contextual precision by 30% in AI-driven writing tools.
• Trained classifiers on formality, politeness, humor, and simplicity, reducing stylistic deviation by 25% compared to baselines, as validated on a 100-pair benchmark subset.