Lesson 2 of 5Intermediate 9 min

LLM-as-a-judge, and its biases

Strong models grade open-ended answers about as consistently as people do. They also prefer longer answers, the first answer, and their own writing.

In this lesson you'll

Know when to use code, model and human graders
Recognise the main judge biases
Calibrate a judge against human labels

When answers are open-ended, exact match doesn't work. Zheng et al. found that a strong LLM judge agreed with human preferences over 80% of the time, about the same rate at which humans agree with each other. They also documented its biases: position bias (favouring the first answer), verbosity bias (favouring longer answers) and self-enhancement bias (favouring its own outputs) ^[1].

Choose the cheapest grader that works

Code graders: exact match, regex, schema validation, unit tests. Fast, cheap, deterministic, but brittle on open-ended output.
Model graders: rubrics scored by an LLM. Flexible, but they need calibrating.
Human graders: the gold standard, used sparingly to calibrate the others ^[2].

Calibrate the judge

Shankar et al. observed criteria drift: people refine what they mean by "good" only after grading real outputs, so you can't fully specify a rubric up front ^[3]. Grade a sample by hand, compare with the judge, rewrite the rubric where they disagree, and repeat. Swap answer order when comparing pairs to cancel position bias.

In Kojen

Every evaluation answer is graded by an LLM judge and gets a verdict. Custom evaluations take your own rubric, so write criteria that are specific and checkable.

Key takeaways

Use code graders where you can, LLM judges where you must, and humans to calibrate.
Correct for position, verbosity and self-enhancement bias.
Expect your rubric to change after you read real outputs.

Practise in the dojo

Create a Custom evaluation with three specific, checkable criteria, then spot-check ten verdicts by hand.

Write a custom rubric

Check your understanding

Score 70 or more to complete the lesson. Progress is saved in this browser.

1. Which bias makes a judge prefer the first of two answers?Verbosity biasPosition biasSelf-enhancement biasRecency bias

2. What is criteria drift?Models changing over timeEvaluators refining their criteria after seeing real outputsDatasets growingLatency increasing

Sources

Checked against the original sources in November 2026.

[1]Zheng, L. et al. (2023). Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. NeurIPS 2023 · arXiv:2306.05685.
[2]Anthropic (2026). Demystifying evals for AI agents. Anthropic Engineering, January 2026.
[3]Shankar, S. et al. (2024). Who Validates the Validators? Aligning LLM-Assisted Evaluation of LLM Outputs with Human Preferences. UIST 2024 · arXiv:2404.12272.