When is an LLM-as-a-Judge Good Enough?

Generative AI models are capable of generating human-like content, such as large language models (LLMs) generate texts. To automate processes, certain requirements regarding trustworthiness and correctness are fulfilled. To relieve people from controlling the created content, a concept known as LLM-as-a-judge is available. In this scenario, another instance of an LLM is prompted to act as a judge, and check the generated content for certain quality requirements. However, such a judgement might not always align with a human judgement. Besides noise and errors in the human judgement, another reason is just the fact that different human evaluators can have different opinions about the evaluation. One reason could be that there might be no universal or unequivocal assessment since it is challenging to formulate objective rules. This natural existence of variance in human judgement defines the best possible benchmark level of alignment for a model judge. In this work, a method is presented to calculate this best possible benchmark under variance in human judgement. By comparing the distribution of the deviations between human judgement with the distribution of the deviation between model and human judgement with statistical means allows defining a model judge as good enough. Thus, a pragmatic way is provided to decide whether human judgement for a specific task can reliably be automated by a model judge in the quality that we can expect from the corresponding human judges. The code and prompts of the analysis performed are available under https://github.com/Ocean-code-1995/llm-judge-good-enough .