Automatic Prompt Optimization with "Gradient Descent" and Beam Search

Reid Pryzant (Stanford), Dan Iter (Microsoft)

Abstract

We present a method for automatic prompt optimization using textual "gradient descent" and beam search. The method iteratively improves prompts by analyzing errors and generating improved versions.

Key Contributions

→Textual gradient descent for prompts
→Beam search prompt optimization
→Error-driven prompt improvement

Topics

prompt optimizationautomatic promptinggradient descentbeam search

Relevance Scores

Long-Horizon Score76

Enterprise Score74

Completeness70

Paper Info

Year2023

VenueEMNLP 2023

Typesystem architecture

ChapterCh. 5

Authors2

Frameworks

PASF