Can LLMs Identify Advisor-Advisee Relationships in Coauthorship Network Based on Publication Metadata?

Identifying advisor–advisee relationships from bibliographic metadata has largely relied on coauthorship network signals. This paper evaluates whether large language models (LLMs) can infer advising relationships from publication metadata alone and whether they can improve on a locally trained probabilistic ranking model, Model M, evaluated at two levels: (1) paper-level advisor identification from a single publication and (2) profile-level ranking and reranking of advisor candidates generated by Model M using the advisee’s first ten publications. At the paper level, gpt-5.4-mini with chain-of-thought reasoning outperforms a last-author heuristic, achieving advisor precision and recall of 0.518 and 0.643, respectively, compared with 0.426 and 0.609 for the heuristic baseline. At the profile-level, LLM ranking alone underperforms Model M. Reranking with gpt-5.4-mini recovers 35 of 300 cases where Model M fails, mainly when its top-two ranked probabilities margin is narrow, but overturns 42 previously correct predictions of Model M, lowering Recall@1 from 0.749 to 0.743. Hence, how best to combine LLM reasoning with probabilistic model remains an open question.