LLM Enhancers for GNNs: An Analysis from the Perspective of Causal Mechanism Identification

Hang Gao, Huang Wenxuan, Fengge Wu, Zhao Junsuo, Changwen Zheng, Huaping Liu

TL;DR

This paper pulls back the curtain on the "LLM-enhancer-plus-GNN" paradigm, using causal interchange interventions to reveal how information actually flows between these two complex models. The authors discover that Large Language Models (LLMs) primarily contribute raw, node-level semantic data rather than complex causal reasoning. To address the bottleneck between the two systems, they introduce a novel Attention-based Transmission (AT) module that significantly boosts performance by optimizing the feature interface between the LLM and the Graph Neural Network (GNN).

What Problem Does This Paper Address?

In recent years, researchers have found a winning formula for graph tasks: use an LLM to generate rich text features for nodes, then feed those features into a GNN. While this "LLM-enhancer" approach works remarkably well, it has largely been treated as a black box.

We don't fully understand the internal logic of these hybrid systems. Does the GNN actually "understand" the causal relationships the LLM provides? Is the bottleneck in the size of the GNN, the power of the LLM, or the way they talk to each other? This paper aims to answer these questions by applying causality theory to identify the underlying mechanisms of the LLM-GNN framework.

Methods

The authors employ a sophisticated analytical framework based on Interchange Intervention to map the neural network's behavior to understandable causal logic.

1. Causal Modeling: They treat the neural network as a "low-level" model and compare it against a "high-level" causal model representing the ground truth.

Aligning a neural network's internal representations with a high-level causal model.

2. Interchange Intervention: To test causal alignment, they take two different graph inputs ($G^{\text{orig}}$ and $G^{\text{diff}}$). They then "swap" specific internal representations (neurons) from the first input's forward pass into the second input's forward pass. If the final output changes in a way that matches the ground-truth causal logic, it proves that those specific neurons are successfully modeling that causal variable.

The interchange intervention process, swapping variables between original and different graphs to calculate intervention loss.

3. The AT Module: Based on their findings that the "interface" is a major bottleneck, they propose the Attention-based Transmission (AT) module. Instead of just taking the last-layer features of an LLM, the AT module uses a Transformer encoder to generate attention scores across multiple prompts and token positions. This allows the system to dynamically select the most relevant semantic information to pass to the GNN.

Datasets

  • CCSG (Controlled Causal-Semantic Graph): A novel synthetic dataset created by the authors. It uses 5,660 Wikipedia entries across 15 subcategories. Unlike standard datasets, CCSG allows researchers to precisely control the "causal knobs"—manipulating semantic relationships and topological structures to see exactly how models respond.
  • Standard Benchmarks: To validate their AT module, they used the well-known Cora, Pubmed, and Instagram datasets.
  • Model Backbones: They tested across multiple LLMs (Llama 2, Llama 3, Qwen 2) and GNN architectures (GCN, GAT, GraphSAGE).

Key Insights and Ideas

  • LLMs as Raw Data Processors: The analysis reveals that in this paradigm, LLMs primarily function as high-quality encoders for node-level and raw data-level information. They don't necessarily "solve" the graph's causal structure; they just provide better ingredients for the GNN to work with.
  • GNN Logic Invariance: Surprisingly, the "logical pattern" within a GNN remains relatively consistent even as you scale the model (adding more layers or hidden dimensions). Scaling up a GNN might make it larger, but it doesn't fundamentally change its capacity for causal relation modeling.
  • The Interface Bottleneck: The authors found that the specific token positions chosen from the LLM output have a massive impact on performance—often more than the choice of the LLM backbone itself. This suggests that how we transmit information from the LLM to the GNN is the most critical design choice.

Key Results

The proposed AT module delivered consistent performance gains across all tested backbones by optimizing this critical interface:

  • Pubmed: Using Llama 3 and GraphSAGE, the AT module provided a +3.09% accuracy boost.
  • Cora: Using Qwen 2 and GCN, it achieved a +2.03% improvement.
  • Instagram: Across various LLMs, the module consistently added between +1.3% and +2.2% to the baseline.
  • Causal Alignment: The experiments showed that lower "Interchange Intervention Loss" ($L_{II}$) correlates strongly with higher model accuracy, providing a new metric for evaluating model "health" beyond just final labels.

Limitations and Future Work

The authors acknowledge that their study primarily focuses on fixed-parameter LLMs (where the LLM isn't fine-tuned alongside the GNN) to keep computational costs manageable. This leaves the door open for future research into "end-to-end" causal mechanism identification where both models are updated simultaneously.

Additionally, the current Structural Causal Model (SCM) treats nodes as the smallest units. Future work could dive deeper into the internal interactions within a single node's text description to see how sub-token features contribute to graph-level reasoning.

Summary updated

Abstract

The use of large language models (LLMs) as feature enhancers to optimize node representations, which are then used as inputs for graph neural networks (GNNs), has shown significant potential in graph representation learning. However, the fundamental properties of this approach remain underexplored. To address this issue, we propose conducting a more in-depth analysis of this issue based on the interchange intervention method. First, we construct a synthetic graph dataset with controllable causal relationships, enabling precise manipulation of semantic relationships and causal modeling to provide data for analysis. Using this dataset, we conduct interchange interventions to examine the deeper properties of LLM enhancers and GNNs, uncovering their underlying logic and internal mechanisms. Building on the analytical results, we design a plug-and-play optimization module to improve the information transfer between LLM enhancers and GNNs. Experiments across multiple datasets and models validate the proposed module.
BibTeX
@inproceedings{gao2025llm,
  author = {Hang Gao and Huang Wenxuan and Fengge Wu and Zhao Junsuo and Changwen Zheng and Huaping Liu},
  title = {{LLM Enhancers for GNNs: An Analysis from the Perspective of Causal Mechanism Identification}},
  booktitle = {International Conference on Machine Learning},
  year = {2025},
  url = {https://openreview.net/forum?id=UWTz4ai3FZ},
}
Figure