MATRIX-Gen - A multi-agent simulation system jointly developed by Shanghai Jiao Tong University and Oxford University

What is MATRIX-Gen?

MATRIX-Gen, a multi-agent simulation system developed by research teams from Shanghai Jiao Tong University and Oxford University, simulates a society composed of over 1000 AI agents with independent identities and personalities, generating diverse and high-quality training instruction data. This training instruction data is used for post-training of large-scale language models (LLMs), improving the models' ability to follow human instructions and demonstrating superior performance across multiple tasks. MATRIX-Gen synthesizes instructions according to different needs, including general and domain-specific datasets, driving the self-evolution and performance improvement of large models.

Main functions of MATRIX-Gen

• Data SynthesisMATRIX-Gen synthesizes diverse and high-quality training instruction data to meet specific needs for post-training of large language models (LLMs).
• Scenario simulationBased on simulating the social interactions of more than 1,000 AI agents, MATRIX-Gen generates realistic and rich scenarios, covering a wide range of fields from software development to business activities.
• Instruction generationBased on the simulated scenario, MATRIX-Gen can generate instructions that conform to human intentions, ensuring the realism and controllability of the synthesized instructions.
• Performance improvementMATRIX-Gen synthesized data can improve the performance of LLMs in multiple domains, including code generation, multi-turn dialogue, and security tasks.
• Self-evolutionUsing data synthesized with MATRIX-Gen, LLMs can achieve self-evolution and outperform traditional training methods even with limited data.

MATRIX-Gen's technical principles

• Multi-agent simulationMATRIX-Gen is an AI social simulator (MATRIX) based on multi-agent simulation technology, which creates more than 1,000 agents, each with a unique identity and personality based on real human profiles.
• Real-world archive initializationThe agent is initialized based on anonymized real human profiles, and its personality and life goals are generated by a large language model (LLM). The goals are broken down into actionable steps to form the agent's action plan.
• Structured communication mechanismMATRIX-Gen is based on a structured communication mechanism, using packets and modulators to manage communication between agents, thereby improving the scalability and realism of the simulation.
• Scene generationBased on agent interaction, MATRIX-Gen generates large-scale real-world scenes, which are used as the basis for post-training data synthesis.
• Instruction generatorMATRIX-Gen, as a scene-driven instruction generator, synthesizes training data based on simulated scenes and specific user needs, including supervised fine-tuning (SFT) datasets, preference tuning (DPO) datasets, and domain-specific SFT datasets.

MATRIX-Gen project address

• GitHub repository:https://github.com/ShuoTang123/MATRIX-Gen
• arXiv technical paper:https://arxiv.org/pdf/2410.14251

Application scenarios of MATRIX-Gen

• Software developmentGenerate instruction data for code generation, code review, debugging, and testing; train and optimize the performance of LLMs in software development tasks.
• Business activitiesSimulate scenarios such as business decision-making, market analysis, and customer service to generate data that enhances the application capabilities of LLMs in business intelligence and strategic planning.
• Medical diagnosisCreate instructional data related to medical diagnosis, case analysis, and treatment plan discussions to improve the decision support capabilities of LLMs in the medical field.
• Education and trainingGenerate instructional data for teaching content, curriculum design, and learning path planning, thereby improving the application of LLMs in personalized education and online learning platforms.
• Customer ServiceSimulate scenarios of customer consultation, problem solving, and service process optimization to generate data that improves the effectiveness of LLMs in automated customer service systems.