Question: Clarification on `LogitsProcessor` usage for labeled graphs vs. Appendix C.3

[lauritsf]

Hi, I'm digging into the codebase to better understand the paper and found something I was hoping you could clarify.

In Appendix C.3 of the paper (arXiv:2502.02216v2), it's mentioned:

Surprisingly, our experiments demonstrate that the constraint-free variant could always generate semantically correct tokenized SENTs... Therefore, we did not use any transition constraints during the inference in our experiments.

However, in the HFSequenceModel.generate method, it looks like a LabeledGraph LogitsProcessor is instantiated and used when self.tokenizer.labeled_graph is True:

AutoGraph/autograph/models/seq_models.py

Lines 282 to 285 in 3dd8170

	if self.tokenizer.labeled_graph:
	logits_processor = LogitsProcessorList([
	LabeledGraph(self.tokenizer, batch_size, input_ids.device),
	])

Since the configs for molecular datasets (like QM9, MOSES) set labeled_graph: true, it seems this processor would be active for those experiments.

I was just wondering if you could clarify this? Was this LogitsProcessor indeed used for the labeled graph experiments, with the statement in the appendix referring only to the simpler, unattributed graph experiments?

Thanks for the great work!

[claying]

Hi @lauritsf

Thank you for your interest in our work and for your careful remark. You are totally correct that our statement in Appendix C.3 refers to the unattributed graph experiments.

For attributed graphs, we found that the model might generate semantically incorrect sequences (e.g., misplacing certain tokens, thus non-convertible to a graph). Therefore, we recommend using constrained decoding for complex graphs.

We will include this clarification in our paper.