Nick Henderson

Toolformer

During a semester abroad in Barcelona at Universitat Pompeu Fabra, I collaborated with a Meta-backed AI research team led by Roberto Dessi, focusing on Toolformer, a large language model designed to enhance its capabilities by making external API calls for tasks it cannot independently complete. My contributions included assisting in the development of a benchmark dataset, building APIs, selecting relevant domains for evaluation, and generating demonstrations to guide Toolformer’s functionality. I specifically focused on optimizing the timing, argument structures, and integration of API call results into the model’s predictions, facilitating more accurate token generation in future tasks.

Exemplar Optimization for In-Context Learning with Large Language Models

While Pursuing my Undergraduate degree at the University of Maryland, College Park, I collaborated with Professor Tianyi Zhou and PhD student Shwai He on an in-context learning project. Our goal was to create an exemplar optimization pipeline by training a white-box large language model (LLM) to select relevant exemplars for few-shot prompting on black-box LLMs. I tested a range of white-box LLMs to identify the most effective ranking system for demonstration selection. Additionally, I worked with LLMs hosted on the UMD cluster, managing GPU resources and conducting experiments to fine-tune the models for optimal performance in few-shot learning tasks