Autoregressive graph generators overfit to specific linearizations rather than learning graph structure, as evidenced by high expected calibration error under permutation and LU correlating better (AUC 0.85) with molecular stability than NLL (AUC 0.43) on QM9.
Note that while biased strategies exhibit lower Novelty compared to Random Order, we do not consider this a degradation in generative quality
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Same Graph, Different Likelihoods: Calibration of Autoregressive Graph Generators via Permutation-Equivalent Encodings
Autoregressive graph generators overfit to specific linearizations rather than learning graph structure, as evidenced by high expected calibration error under permutation and LU correlating better (AUC 0.85) with molecular stability than NLL (AUC 0.43) on QM9.