Systems for generation of prompts for evaluation of language models
Published in United States Patent Office, 2025
Describes systems and methods to generate synthetic prompts for red-teaming large language models (LLMs). A first ML model (e.g., a Q-learning model) learns prompt modifications that increase the probability of eliciting responses that violate constraints, while additional models score prompts and responses and generate rationales to guide subsequent prompt generation and improve evaluation coverage.
Recommended citation: Lopez Martinez, Daniel. Systems for Generation of Prompts for Evaluation of Language Models. U.S. Patent Application Publication No. US 2025/0378274 A1, published Dec. 11, 2025 (filed Jun. 7, 2024). Applicant: Amazon Technologies, Inc.
View Patent