🪲 multiplex: A Modular LLM-based Mutation Framework

A tool for prototyping and comparing LLM-based mutation testing approaches.

multiplex has a modular design to enable researchers to build, design and test LLM-based Mutation testing approaches in an easy to use framework.

Multiplex is currently designed to enable mutation at the method level, but prompting and further modularisation could support other context levels. Fork this project and add your own modules.

Getting Started

✅ Requirements

• UV Dependency Management: Follow installation instructions at uv.
• yq for parsing yaml config files with bash

👩‍💻 Installation

1. Clone the repository:

   git clone https://github.com/LLM-Mutation/multiplex.git

2. Install the dependencies and activate environment:

   cd multiplex
   uv sync 
   source .venv/bin/activate
   cd ..

🔧 Configuration

multiplex is configured using a yaml file. Users can use the example configuration file in the examples/ to set up their project. This section overviews the content required in the configuration file.

Project

This section includes the details about the project and method under test, including reference to the module for running and parsing the tests.

LLM

To configure the LLM, update the config.yml file. Below is an example for gpt-oss:20b, run locally using Ollama. To use gpt-oss:20b, download Ollama, and run ollama pull gpt-oss:20b. Ensure Ollama is running when using multiplex for this example.

llm:
    model: ollama/gpt-oss:20b
    endpoint: http://127.0.0.1:11434
    tokenenvvar: # env var name storing token (NOT TOKEN)

[IMPORTANT] Security Note: Never hardcode API keys in your config file. Set an environment variable and reference its name in the tokenenvvar field.

System Prompts

NOTE: Mutahunter Module prompts must be added by user due to licensing restrictions. - User Prompt Link / System Prompt Link.

The System Prompts are included in the configuration file for easy updating. The user prompts are constructed within multiplex so users should modify or create modules to alter these.

🏃 Running multiplex

Once configured and modules are set up, users can run multiplex using the following command:

uv run /path/to/multiplex ./path/to/config.yml

🧩 Existing Modules

multiplex currently includes five mutant generation modules and two execution and evaluation modules:

Mutant Generation Modules:

• HAZOP
• STPA
• basic (or unguided)
• LLMorpheus
• Mutahunter (note the user must collect Mutahunter's prompts themselves due to licensing restrictions).

Execution and Evaluation Modules

• Maven
• Defects4J (enables users to target specific bugs in the Defects4J dataset)

🏗️ Adding Modules

Adding LLM Modules

multiplex uses LiteLLM to interface with LLMs. Many local options will be configurable directly from the configuration file. However, specific LLM providers may require further information. In this instance, the user should create their own multiplex/model.py file and use LiteLLM to interact with their provider of choice.

Adding Mutant Generation (MG) Approach Modules

In order to create an MG approach, users should add an MG package to multiplex/approach containing a controller.py file that directs their approach. It may help users to use the existing modules in multiplex/approach as examples as there are single prompt (e.g., basic) and multi-prompt (e.g., HAZOP) examples.

Adding Execution and Evaluation Modules

Users can add a mutant execution and evaluation class to multiplex/execute/ that inserts the mutant into the source code file, runs the project's tests with the appropriate build and execution tools, and parses the results to identify which mutants are killed by which tests.

💬 Questions and Community

Have a question or found a bug? We’d love to hear from you. Please open an issue in the Issue Tracker and we'll get back to you!