Questiones about graph-constrained decoding. (following the closed issue GCR Compare with RoG)

[liane886]

Hi,

Thank you for your valuable contribution. I have a question regarding how the LLM generates appropriate KG-grounded reasoning paths. Specifically, I noticed that the KG information does not appear to be explicitly provided as a prompt when the LLM is tasked with generating the reasoning paths.

How does the LLM manage to produce the relations and entities directly from the KG in this case? Is this because the model has been fine-tuned using ground-truth paths, or is there another mechanism involved?

Best.

[RManLuo]

Hi, Thanks for your interest in our work. In GCR, we adopt the graph-constrained decoding to ensure the truthfulness of LLM generations, which is realized by the KG-Trie. You can find more details in our paper.