Learning How Much to Think: Difficulty-Aware Dynamic MoEs for Graph Node Classification

D2MoE addresses a critical efficiency gap in Graph Neural Networks (GNNs) by applying the 'dynamic compute' logic currently popular in LLMs (e.g., Mixture-of-Experts) to node-level classification. While static MoEs for graphs exist, the 'difficulty-aware' aspect—allocating more experts to 'hard' nodes and fewer to 'easy' ones—is a clever optimization for sparse and heterogeneous graph data. The project currently has 0 stars but 7 forks within 24 hours of release, a signal usually indicating a research lab release where multiple contributors are active or the paper is being prepared for a major conference (ICLR/NeurIPS). Its defensibility is low (3) because it is primarily an algorithmic contribution; once the paper is published, the technique can be easily reimplemented in standard libraries like PyTorch Geometric or DGL. Frontier labs (Google DeepMind, Meta AI) are high-probability 'competitors' here, as they maintain the primary GNN frameworks and are actively seeking ways to scale GNNs to trillion-edge graphs. The displacement horizon is 1-2 years, as dynamic routing is likely to become a standard toggle in graph-learning frameworks rather than a standalone product.