Bad evals, my own: five exercises from two LLM judges (digline.dev)

A recent exploration by a developer running two LLM (Large Language Model) judges revealed significant insights into their evaluation processes. The developer utilized these judges to sift through AI-related content, designed to identify articles worth reading and engaging with on platforms like Reddit. By conducting a series of exercises, the developer tested their judges' performance across various scenarios, highlighting issues related to accuracy and consistency. For example, evaluations showed fluctuating results over multiple runs despite identical configurations, indicating that LLM outputs may not be deterministic. This inconsistency raises concerns about relying solely on LLMs for assessments, as their outputs can vary significantly even on the same inputs. This examination is crucial for the AI/ML community as it emphasizes the need for robust evaluation protocols in LLMs. The developer noted that minor changes in instructions could drastically alter the LLM's decision-making, suggesting that model prompts require careful crafting to ensure they yield the desired outcomes. Furthermore, the findings underscore the importance of transparency in evaluation metrics, as hidden biases in judge outputs may obscure true performance assessments. Overall, this study contributes to ongoing conversations about enhancing the reliability and accuracy of LLMs, ultimately influencing their implementation in practical applications across industries.