RainbowPlus: Enhancing Adversarial Prompt Generation via Evolutionary Quality-Diversity Search

Quy-Anh Dang, Chris Ngo, Truong-Son Hy · 2025

Read RainbowPlus: Enhancing Adversarial Prompt Generation via Evolutionary Quality-Diversity Search by Quy-Anh Dang, Chris Ngo, Truong-Son Hy (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-15047.txt.

Description

Quy-Anh Dang, Chris Ngo, and Truong-Son Hy build a faster, broader generator of attack prompts for language models, in a 2025 paper. Red teaming hunts for prompts that trick models into unsafe or biased output, but older methods are slow or narrow. Their framework, RainbowPlus, borrows from evolution: it keeps an archive of varied, high-quality attack prompts and breeds new ones from them.

Unlike earlier quality-diversity methods, it stores many prompts per niche and scores several at once. Across six data sets and four open models, it produced up to a hundred times more distinct prompts than a rival method. On a standard harm benchmark it beat the best rival's success rate and ran about nine times faster.

Catalog record

CREATORQuy-Anh Dang, Chris Ngo, Truong-Son Hy
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE99 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-15047

Browse the full library · About the Tower of Babel