Qualitative Diagnosis of LLMs as Judges Using LevelEval

Recent studies focus on evaluating LLMs as judges to see if they match human-level capabilities. LLMs are used as judges in methods like pairwise comparison, single answer grading, and reference-guided grading. While traditional methods provide overall statistics of their performance, they may miss specific task weaknesses. LevelEval introduces an automated pipeline using rank-based evaluation to analyze LLMs across different input quality levels. This paper extends LevelEval by integrating a qualitative diagnosis report that assesses LLM strengths and weaknesses based on document types. This report includes difficulty scores for tasks and uses topic modeling to identify where LLMs excel or struggle during evaluation. Results show that this qualitative approach gives more insight than traditional metrics like BLEU and ROUGE. For instance, LLMs performed well on some topics but struggled with others. Improving open-source LLMs like Mistral on specific topics could enhance their evaluation performance in tasks like summarization and question answering, potentially matching GPT-4. The study found that Mistral underperformed on land-related law summarization and difficult math problems in retail and asset management. It also highlighted the need for better bias detection and classification in all open-source models to match GPT-4.