Systematic Evaluation of Structural Knowledge Prompting Generalization in Language Models
A framework for judging whether structural knowledge prompting -- injecting knowledge-graph structure into an LLM through a structural encoder and adapter -- generalizes beyond the single task it was tuned for. Generalization is assessed along four dimensions (granularity, transferability, scalability, universality) using the SUBARU benchmark of nine graph-derived reasoning tasks, revealing what actually drives SKP's success and where its capability boundaries lie.
2501.00244
Structural knowledge prompting (SKP) improves the factual accuracy of large language models by encoding a knowledge graph's entities and relations with a structural encoder and bridging them into the…