When Text and Topology Disagree: Bidirectional GNN–LLM Co-Evolution for Noisy Graph Learning

Text-attributed graphs combine relational structure with node-level natural language, making them a natural setting for integrating Graph Neural Networks (GNNs) and Large Language Models (LLMs). Existing Graph–LLM methods perform well on relatively clean benchmarks, but many rely on one way pipelines or confidence-blind alignment objectives that can degrade when text and topology disagree. This paper studies node classification on noisy text-attributed graphs, where textual descriptions may be incomplete or misleading and the observed graph may contain missing links or spurious cross-class edges. We propose Co-Evolve, a bidirectional Graph–LLM framework that treats graph topology and semantic embeddings as coupled latent variables and refines them through alternating optimization. In each cycle, the GNN summarizes high-order structural context and injects it into the LLM as structure-conditioned soft prompts. The LLM then produces updated semantic embeddings that construct a dynamic semantic graph and feeds it back into the GNN, and the cycle repeats. To stabilize this information exchange under noisy conditions, we incorporate three key innovations in Co-Evolve: adaptive node-level gating between static and learned graphs, a conflict-aware loss function that separates semantically similar but structurally inconsistent nodes while preserving high-order structural neighbors, and uncertainty-gated consistency so that one view guides the other only when it is confident. Experiments on Reddit, Instagram, and WikiCS show that in the presence of semantic corruption, missing links, and structural noise, Co-Evolve outperforms state-of-the-art methods by wide margins. Even on clean benchmarks, Co-Evolve is still highly competitive with very similar accuracy and F1-scores as the best performing existing methods.