Enhanced Graph Transformer with Serialized Graph Tokens

Researchers have developed an Enhanced Graph Transformer that uses serialized graph tokens to overcome information bottlenecks in graph-level representation learning. The novel method aggregates node signals into multiple tokens and applies self-attention to capture complex dependencies, moving beyond the limitations of single-token approaches. Experimental results demonstrate state-of-the-art performance on several graph-level benchmarks, with ablation studies confirming the effectiveness of the proposed serialization modules. The paper has been accepted for ICASSP 2026.