Why traditional GNNs struggle

Graph Neural Networks power many graph tasks from citation analysis to recommendations. Most GNNs rely on homogeneous, static message passing where every node aggregates neighbor information using the same fixed rules. Two problems arise from this approach:

• Node informativeness imbalance. Some nodes hold rich, relevant data while others are sparse or noisy. Treating them identically can drown useful signals in noise or erase rare but important context.
• Locality limitations. Standard GNNs emphasize local neighborhoods, often missing semantically similar but distant nodes that could provide crucial evidence.

ReaGAN: nodes as autonomous agents

ReaGAN reframes each node as an agent that actively plans actions based on its private memory and current context. Instead of a uniform update rule, nodes consult a frozen large language model such as Qwen2-14B to decide what to do next. That decision can include gathering local neighbor info, retrieving semantically relevant global content from the graph, predicting a label, or doing nothing to avoid overfitting to noisy signals.

Key components of the agentic design:

• Agentic planning. Nodes build prompts from their memory and context, and a frozen LLM recommends actions like collect more info, predict label, or pause.
• Flexible actions. Local aggregation harvests neighbor features. Global aggregation uses retrieval-augmented generation to pull semantically related content anywhere in the graph. A NoOp option allows nodes to skip noisy updates.
• Private memory. Each node keeps a buffer with raw text features, aggregated context, and labeled examples to support tailored prompting and stepwise reasoning.

How ReaGAN works

The ReaGAN workflow is an iterative reasoning loop:

1. Perception. The node reads its state and memory buffer to form a local view.
2. Planning. A prompt summarizing memory, features, and neighbor info is sent to the frozen LLM, which suggests next actions.
3. Acting. The node performs local aggregation, global retrieval, label prediction, or NoOp, and stores outcomes back to memory.
4. Iterate. Several reasoning layers let nodes refine and integrate evidence.
5. Predict. In the last stage, nodes produce label predictions supported by combined local and global evidence.

Crucially, nodes operate asynchronously and decide independently rather than following a global synchronous clock or identical parameter updates.

Results and implications

ReaGAN performs strongly on classic benchmarks like Cora, Citeseer, and Chameleon. Remarkably, it achieves competitive accuracy without supervised training or fine-tuning, showing how agentic planning plus semantic retrieval can substitute for traditional end-to-end learning in certain graph settings.

| Model | Cora | Citeseer | Chameleon | |---|---:|---:|---:| | GCN | 84.71 | 72.56 | 28.18 | | GraphSAGE | 84.35 | 78.24 | 62.15 | | ReaGAN | 84.95 | 60.25 | 43.80 |

ReaGAN relies on a frozen LLM for planning and context retrieval, underscoring the power of prompt design and semantic retrieval strategies.

Key insights

• Prompt engineering matters. How nodes mix local and global memory in prompts affects performance, and the optimal strategy depends on graph sparsity and label locality.
• Label semantics influence bias. Revealing explicit label names can skew predictions, while anonymizing labels often improves results.
• Agentic flexibility helps. Decentralized, node-level reasoning is especially effective in sparse graphs or those with noisy neighborhoods.

ReaGAN demonstrates a new direction in graph learning where nodes are not passive feature containers but adaptive, context-aware reasoning agents that combine local structure with global semantic retrieval.