Content-type: text/html Downes.ca ~ Stephen's Web ~ When "A Helpful Assistant" Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models

Stephen Downes

Knowledge, Learning, Community

As this article (29 page PDF) states, we often think of each other in terms of personas. Also, the use of personas is a staple in user interface design. A persona is a prototypical person, defined in terms of that person's role, and often associated with demographic factors, such as age and gender. This paper asks whether we get better results from a large language model (LLM) if we define a persona in our prompt. The answer is a bt complex, because the use of a persona does change the output, and can even be better in some cases, but the distribution of cases is random, and there's no real net benefit to using personas in prompts. Obviously this finding should inform people using personas in LLMs in learning support applications.

Today: 750 Total: 991 [Direct link] [Share]


Stephen Downes Stephen Downes, Casselman, Canada
stephen@downes.ca

Copyright 2024
Last Updated: Dec 04, 2024 11:26 p.m.

Canadian Flag Creative Commons License.

Force:yes