GLOW: Graph-Language Co-Encoding for Agentic Workflow Performance Prediction

GLOW proposes a unified framework that combines Graph Neural Networks (GNN) with Large Language Models (LLMs), aiming to predict the performance of intelligent agent workflows. This framework constructs graph-oriented LLMs through instruction fine-tuning to extract topologically-aware semantic representations, while using GNNs to model structural information and optimizing fusion and contrast learning in a shared latent space. In the FLORA-Bench benchmark, GLOW outperforms existing state-of-the-art baselines in both prediction accuracy and ranking utility. When integrated into the automatic workflow generation framework AFLOW, system optimization time is reduced by 98.7%, and the average score only decreases by 0.031, demonstrating efficient agent evaluator performance.