PromptZ

Frontend - Step 1
Frontend - Real-time GA full-trace
Frontend - Interactive with the chromosomes (click in a node)
Algorithm overview
Chromosome and population diagram
Mixing algorithm and mutation

This study of developing a genetic algorithm for prompt generation within a Large Language Model (LLM) was inspired by the possibility of safeguarding language models against misuse by discovering prompts that can be potentially malicious.

What it does

Our solution is a ready to implement framework that generates and refines a list of prompts based on a target response. This functionality is designed to facilitate vulnerability testing for LLMs, offering a proactive approach to mitigate abuse and misuse.

How we built it

We constructed a Genetic Algorithm (GA) framework that optimizes prompts against specific target objectives. The definition of chromosomes, pivotal to the GA process, can be achieved through keywords or a more sophisticated approach involving natural language. The generation process adapts accordingly, utilizing mutations and crossovers, either through a KeywordsGenerator or NLGenerator, depending on the chosen chromosome definition.

The evaluator module, akin to a fitness function, plays a critical role in assessing whether the Large Language Model under Test (LLMUT) outputs align with predefined objectives. We implemented two evaluators: SemanticSimilarityEvaluator, which leverages semantic analysis to measure similarity, and ObjectiveSimilarityFunction, which evaluates based on defined objectives, such as generating harmful sentences or extracting privileged information.

Prioritizing usability, we incorporated both frontend and backend components into our system. The RestAPI with a hypercycle server allows efficient retrieval of optimal results, while the bidirectional real-time interaction between the frontend and backend provides a comprehensive trace of the algorithm's execution. This approach enables a deeper understanding of the LLM's behavior by identifying words or sentences prone to achieving better performance.