Contribution
A re-run of the benchmark used to argue that graph transformers are needed for long-range tasks, with baselines tuned as carefully as the proposal.
Key insight
The baselines were handicapped by configuration, not architecture. A linear readout cannot compute a target that depends non-linearly on pooled graph features. A transformer layer’s feed-forward block partly compensates, and an MLP head fixes it more cheaply.
Key results
- Peptides-func AP: GCN 0.593 to 0.686, GPS 0.654 to 0.653. Peptides-struct MAE: GCN 0.350 to 0.246, GPS 0.250 to 0.251
- The MLP head alone accounts for most of the MPGNN gain
- PascalVOC-SP F1 after tuning: GatedGCN 0.388, GPS 0.444. Normalization is at least half of most models’ gain
- Filtered vs raw MRR on PCQM-Contact changes the scores substantially
Limitations
- A short preprint. Positional encoding is tuned as a hyperparameter, with no separate ablation
- GPS stays ahead on the vision datasets. The closure is dataset-specific
- Same parameter budget (500k) as the benchmark, and larger models are not tested