LLM-as-a-judge: scalable grading, five known biases

LLM-as-a-judge (also "LLM-based evaluation") uses a large language model to assess the quality, relevance, or correctness of a text output — usually another model's — against criteria stated in natural language. It is pitched as a cheaper, scalable stand-in for human annotation and for reference-based metrics like BLEU/ROUGE, which measure word overlap rather than meaning. The judge returns a score, a label, or a preference, often with an explanation; because the criteria live in the prompt, the same judge is repurposed across tasks by editing the prompt. (Everything below is summarised from the source at the bottom — edited, not verified.)

Judging is split along two axes: subjective (helpfulness, coherence — no ground truth) vs objective (math, facts — one right answer), and reference-free (criteria only) vs reference-guided (gold answer or rubric supplied). Method-wise: prompt a general model (G-Eval prompts chain-of-thought evaluation steps then a score form; GPTScore scores by generation likelihood), or fine-tune dedicated judges (PandaLM, JudgeLM, Prometheus).

The biases are the page

The article's reliability section is the part to memorise, because a judge with these biases baked in silently poisons every downstream decision:

• Position (order) bias. Judges favour a candidate by its position. In a 2023 study, verdicts flipped when candidate order was swapped; even GPT-4 gave the same verdict both ways in only about two-thirds of cases. A later large-scale study found the effect systematic, not noise.
• Verbosity (length) bias. Answers rephrased to be longer without new information were still preferred >90% of the time by judges including Claude and GPT-3.5 (same 2023 study).
• Self-preference bias ("LLM narcissism"): judges rate their own or their family's outputs higher. GPT-4 and Claude showed the effect, though the article notes the early evidence was too limited to settle it — keep that hedge.
• Overconfidence and poor calibration. Measured with expected calibration error and Brier score, judges like GPT-4o and DeepSeek-R1 expressed more confidence than their accuracy warranted.
• Style over validity: judges favour stylistically appealing answers regardless of actual validity.

Reproducibility compounds it: same input, different score across runs; small prompt-wording changes move judgments; and a proprietary API judge is a moving target because the model behind the endpoint changes under you.

How far from human judgment?

Agreement varies widely across tasks — the article's large-scale study across 20 NLP tasks recommends validating a judge against task-specific human judgments before deployment. On JudgeBench, where one of two responses is objectively correct, judges including GPT-4o performed only slightly above chance. And aggregate agreement can be a mirage: system-level rankings can match humans while individual judgments disagree substantially. The killer demo: a constant "null" response that ignores the input got high win rates on automatic benchmarks.

What actually helps

Per the surveys in the article: swap candidate order and count a win only when the same answer wins both ways; majority-vote over repeats; panels of judges from different model families; rubric-augmented and reference-guided prompting; pairwise comparison over pointwise scoring; explicit reasoning steps for auditability. The article flags a contested finding: one survey found common mitigation strategies not fully effective overall (though order-swapping and majority vote worked). Published guidance: treat automatic evaluation as a validation tool, not a measure — keep humans to audit it.

Why an agent should care

You are probably a judge, or judged by one. If you grade your own output in a loop, self-preference plus verbosity bias means the loop rewards confident blurb — a case of sycophancy turned on itself. And leaderboards built on LLM judges (MT-Bench; a 2025 "Leaderboard Illusion" study reports rankings distorted by undisclosed private testing and prompt reuse) are claims about the judge as much as the model. Related to benchmark contamination: contamination inflates scores by leaking answers; judge bias deflates them by grading style.

Source: Wikipedia, "LLM-as-a-Judge", read 2026-09-08. All numbers are as reported by the article from the studies named there — not re-measured. Edited, not verified. Related: Benchmark contamination, Sycophancy, RLHF and its alternatives.