ScoreFlow: Automated Multi-Agent Workflow Generation and Optimization 191

小编 · 2025-03-26

ScoreFlow is an automated method for generating and optimizing multi-agent workflows. Using gradient-based continuous-space optimization combined with Score-DPO, it iteratively refines collaborative processes to boost model performance. It outperforms existing methods on benchmarks for QA, coding, and mathematical reasoning, and enables smaller models to surpass larger ones at lower inference cost. Developed by teams from the University of Chicago, Princeton, and Oxford.

ScoreFlow: Automated Multi-Agent Workflow Generation and Optimization

What is ScoreFlow?

ScoreFlow is an automated multi-agent workflow generation and optimization method. It leverages gradient-based continuous space optimization techniques combined with Score-DPO to continuously improve collaborative workflows, thereby enhancing model performance. It outperforms existing methods across multiple benchmarks, including question answering, coding, and mathematical reasoning, and enables smaller models to surpass larger models at lower inference costs. ScoreFlow was developed by teams from the University of Chicago, Princeton University, and Oxford University.

ScoreFlow Features

Efficient Optimization: Leverages continuous space gradient optimization techniques to enhance model performance.
Preference Optimization: Combined with Score-DPO, integrating Score-DPO—a variant of Direct Preference Optimization that incorporates quantitative feedback to optimize model outputs.
Small Model Advantage: Helps smaller models outperform larger models, reducing inference costs.
Multi-Task Support: Covers various application scenarios including question answering, coding, and mathematical reasoning.

Experimental Results

Across six different benchmarks (including question answering, coding, and mathematical reasoning), ScoreFlow achieves an 8.2% improvement over existing baselines.
ScoreFlow also enables smaller models to outperform larger models at lower inference costs.

ScoreFlow: An Automated Multi-Agent Workflow Generation and Optimization Method.webp

ScoreFlow Use Cases

Question Answering Systems: Improves the accuracy and efficiency of QA models.
Programming Assistance: Optimizes code generation models to improve programming efficiency.
Mathematical Reasoning: Enhances model performance in solving math problems.
Multi-Agent Collaboration: Optimizes multi-agent workflows to improve overall task efficiency.

How to Use ScoreFlow

Hardware Preparation: Requires 80-90GB of VRAM.

Environment Setup:

Use conda to create a new virtual environment and install the necessary dependencies.
Download the dataset and unzip it into the project directory.

Optimization Process:

Iteratively optimize the model to progressively improve performance.
Use the generate.py and evaluate.py scripts for optimization.
The optimization process can be adjusted by increasing the number of iterations.

Inference: