Judge, Generator, Executioner: Utilizing an LLM for KBC

In this work, we introduce our submission to the 2025 LM-KBC Challenge. In keeping with the spirit of the challenge, we avoided using additional tools or information; to explore the degree to which the LLM was able to holistically handle the task of triple completion. We found that candidate filtering is crucial to refining the quality of the generated triples, as all filters we introduced improved on our baseline. This is further improved upon with the addition of more initial tail candidates per triple, which highlights the LLM’s potential to act as both the generator of candidates and judge of quality in triple completion tasks.