explainable-learning

X-LogSMask introduces a logarithmic structural mask for graph transformers, injecting graph topology directly into attention logits to achieve state-of-the-art performance on 13 out of 20 benchmarks while preserving interpretability and multi-hop information propagation.