How reliable is LLM-as-a-judge and how do you validate it?
hardAnswer
- Frontier LLMs correlate 0.6-0.8 with human raters on many pairwise-preference tasks — usable but biased.
- Known biases: (1) position bias (prefer first response); (2) length bias (prefer longer answers); (3) self-preference (GPT-4 favors GPT-4-style outputs); (4) domain gaps (biology / law require domain-expert judges).
- Validation: label a small human-annotated set (100-500 examples), compute correlation, adjust prompt / model.
- Mitigations: randomize position, use pairwise + swap, ensemble across judges, calibrate against humans.
Check yourself — multiple choice
- LLM judges are perfect
- 0.6-0.8 correlation with humans; known biases (position, length, self-preference); validate on human-labeled subset + swap-and-average
- Never correlate
- Only for classification
LLM-judge: 0.6-0.8 human correlation; biased on position/length/self — always validate and de-bias.
#evaluation#reliability
Practise LLMs & GenAI
214 interview questions in this topic.
Related questions
- How do you detect hallucinations in production?
- How well-calibrated are LLM confidences and how do you fix them?
- Why do LLMs hallucinate and how do you reduce hallucinations?
- How do you evaluate an LLM system?
- What is benchmark contamination and how do you detect it?
- What is sycophancy in LLMs and how do you reduce it?