TREX: Multi-Agent Collaboration Enables End-to-End Automation of LLM Fine-Tuning Process

TREX is an innovative multi-agent system proposed to address the complex and time-consuming issues of LLM fine-tuning. Through collaboration between two core modules—Researcher and Executor—it实现 end-to-end automation of the LLM training process. The system uses a tree-structured search framework to manage multi-round experiments and introduces the FT-Bench benchmark to validate its effectiveness, aiming to solve challenges in traditional training workflows such as multi-stage dependencies and low iteration efficiency.

The traditional LLM training process requires significant effort in literature research, data preparation, strategy formulation, and experimental iteration, involving multi-dependent stages such as requirement analysis, data collection, and training evaluation. Existing AI agents can only perform isolated scientific tasks and struggle to handle complete training workflows; experimental iteration requires extracting insights from historical results to plan directions, which a single agent or simple script cannot handle.

The core of TREX is a multi-agent collaboration model:

• Researcher Module: Responsible for high-level planning and decision-making such as requirement analysis, literature research, data study, and strategy formulation;
• Executor Module: Receives strategies and data recipes, executes model training and evaluation, and feeds back results. The division of labor ensures rational decision-making and efficient execution.

TREX models multi-round experiments as a tree-structured search framework, with advantages including:

1. The system plans exploration paths to avoid redundant work;
2. Reuses historical results to enhance efficiency through incremental exploration;
3. Analyzes tree-structured paths to extract high-level insights and guide future exploration directions.

The research team built the FT-Bench benchmark, which includes 10 real-scenario tasks covering general language abilities (reasoning, code generation) and specific domains (mathematical problem-solving, professional Q&A). The task design considers the diversity of practical applications, making the evaluation results more representative.

Experiments show that TREX can continuously optimize the performance of models on target tasks. Its significance lies in:

• For researchers: Shortens the cycle from idea to validation, allowing focus on innovative problems;
• For industry: Reduces the threshold for LLM fine-tuning, helping organizations customize models; It also demonstrates the potential of multi-agent collaboration to solve complex AI engineering problems.

Insights from TREX:

1. Modular design: Decomposes tasks into planning and execution to ensure decision quality and efficiency;
2. Structured experiment management: Tree-structured search is an effective way to organize and reuse experimental knowledge;
3. Multi-agent collaboration: Role division enhances the overall capability of the system. Outlook: In the future, more automated systems will accelerate the development and application of AI models.