Project D01

Probing the limits of prompt variability in Large Language Models

PI(s): Prof. Dr. David Schlangen

This project investigates prompt (in)stability in Large Language Models. While this is typically treated as a technical problem, it is treated here as a means for exploring the limits of linguistic variability captured by these systems. By defining a range of controlled prompt manipulations and evaluating the induced functions on standard datasets, the project aims to elucidate the way these models go from analysis and interpretation of prompts to instruction following – and the extent to which it concurs with linguistic predictions.

Members