CO-EVOLVE: Bidirectional Co-Evolution of Graph Structure and Semantics for Heterophilous Learning

Researchers Jinming Xing and Muhammad Shahzad have introduced CO-EVOLVE, a groundbreaking framework that fundamentally rethinks how Large Language Models (LLMs) and Graph Neural Networks (GNNs) work together. Traditional approaches treat these systems as static, unidirectional pipelines where errors from one model permanently corrupt the other—a problem the authors call "bidirectional error propagation." CO-EVOLVE solves this by treating graph topology and semantic embeddings as dynamic, mutually reinforcing variables that evolve together through a Gauss-Seidel alternating optimization strategy. This creates a continuous feedback loop where the GNN provides structural context to guide the LLM, while the LLM constructs dynamic semantic graphs to rewire the GNN.

The framework introduces three key innovations to stabilize this co-evolution. First, a Hard-Structure Conflict-Aware Contrastive Loss warps the semantic space to respect high-order topological boundaries. Second, an Adaptive Node Gating Mechanism dynamically fuses static and learnable structures to recover missing connections in the graph. Third, an Uncertainty-Gated Consistency strategy enables "meta-cognitive alignment," ensuring each model only learns from the other's confident predictions. During inference, an Entropy-Aware Adaptive Fusion layer integrates final predictions.

Extensive testing on public benchmarks demonstrates CO-EVOLVE's significant advantages over existing methods. The system achieved average improvements of 9.07% in accuracy and 7.19% in F1-score, particularly excelling in heterophilous settings where textual similarity contradicts topological reality. This represents a major advancement in creating AI systems that can reason about both language and complex network structures simultaneously.