AFlow
Object
AFlow is an automated framework for generating and optimizing agentic workflows tailored to large language models (LLMs). This innovation eliminates the need for manual workflow design while enhancing performance and reducing associated costs.
Features
- Monte Carlo Tree Search algorithm that efficiently explores code-represented workflow spaces
- Modular architecture with predefined operators (Generate, Format, Review, Revise, Ensemble, Test)
- Iterative workflow refinement through code modification and execution feedback
- Tree-structured experience backpropagation for continuous improvement
- Support for multiple benchmark datasets, including HumanEval, MBPP, GSM8K, MATH, HotpotQA, and DROP
- Cost-efficient performance optimization enabling smaller models to outperform larger ones
Outcome
Provides workflows that exceed the performance of manually engineered systems by an average of 5.7% across benchmark datasets, while facilitating smaller Large Language Models (LLMs) in attaining results comparable to those of larger models at merely 4.55% of the inference cost. This advancement significantly enhances both performance and cost-effectiveness for complex artificial intelligence tasks.