When can LLMs replace humans in judging model output quality?

Category: LLM & Prompt Engineering

Answer

LLM-as-judge correlates ~98% with ChatBotArena human ratings at < $10/run. Use it for fast iteration. BUT: LLM judges have bias (prefer longer answers, prefer same model). Control for length with regression. Use bigger models as judges (GPT-4) than candidates (Claude/Gemini vs each other).

Concrete examples from the fca project context

Example 1

“Compare A and B. Which is more accurate, helpful, well-formatted?” -> win rate.

Example 2

Length bias: regress out length so judge doesn’t reward verbose.

Example 3

Cross-check at least monthly with a sample of human pairwise ratings.

Example 4

Cheaper than ChatBotArena but biased; treat the score as a north star, not ground truth.

Data flow / flow chart

candidate -> generated -> {A, B}
judge larger model -> "Which is better?" -> win-rate
length correction reduces bias

Takeaway

LLM-as-judge for fast iteration. Human spot-checks at least monthly; never trust LLM-only scores for production decisions.