Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution

Source: arXiv Authors: Chrisantha Fernando, Dylan Banarse, Henryk Michalewski, Simon Osindero, Tim Rocktäschel Date: 2023-09-28 Primary category: cs.CL All categories: cs.CL, cs.AI, cs.LG, cs.NE

Abstract / key passage

Promptbreeder evolves a population of task prompts and, crucially, also evolves the mutation prompts used to edit them. Its central move is self-referential improvement: the mutator itself becomes part of the search object rather than a fixed hidden heuristic.

Harness takeaway

Promptbreeder matters because it treats optimization procedures as writable artifacts, not only the task prompt. For long-lived harnesses that is a suggestive model: mutation heuristics, repair prompts, and evaluator-facing instructions may all be worth versioning rather than treating as ambient magic.