On the Use of LLMs for Relevance Labelling

Large Language Models (LLMs) are increasingly used to replace human judges to assess the relevance of information objects, raising concerns about circularity, bias, and whether simulated preferences can substitute for human judgement. This work presents experiments using multiple LLMs to label passages for relevance. It examines their gullibility—how easily they are misled into labelling irrelevant passages as relevant. It also compares LLMs with human judges in ranking systems, analysing differences in discriminative power and whether some systems benefit under LLM-based evaluation. Results show that LLMs are influenced by the presence of query terms, even with irrelevant or random passages. Moreover, LLM-generated rankings are highly correlated with those of human judges, with strong agreement on which system is better in pairwise comparisons. However, LLMs may exhibit lower discriminative power, as seen in flatter ranking slopes and missed significance for meaningful improvements. Yet, there are no cases where capable LLMs and human judges reach opposing conclusions with significance. LLMs may boost traditional systems more than neural ones, adding a new concern of system bias. These findings highlight the strong potential of LLMs for relevance labelling, while also highlighting failure cases that call for careful adoption and further research to maintain evaluation integrity. 1