LLM-as-a-Judge (also LLM-based evaluation or language model-based evaluation) is a technique in natural language processing in which a large language model (LLM) is used to assess the quality, relevance, or correctness of a text output, typically one generated by another model, against predefined criteria. It is used as a scalable, lower-cost alternative to human annotation and to reference-based metrics such as BLEU and ROUGE, which measure word overlap rather than meaning.
Definition In LLM-as-a-judge, a large language model acts as an evaluator: it receives an evaluation prompt specifying the target text, the criteria, and the required output format, and returns a judgment such as a numerical score, a categorical label, or a preference between candidates. Because the criteria are expressed in natural language, the same judge can be repurposed across tasks by changing the prompt rather than by retraining. Instead of relying solely on human annotators, the approach uses the general language capabilities of large language models to serve as automated judges. The judge usually returns a numerical score or a verdict, and in many implementations also produces a natural-language explanation for its decision. Sources also call the technique "LLM-based evaluation", or "language model-based evaluation". Judging tasks are commonly distinguished along two axes. The first contrasts subjective evaluation, in which there is no single ground-truth answer, such as rating the helpfulness or coherence of an open-ended response, with objective evaluation of outputs whose correctness can be verified, such as answers to mathematical or factual questions. The second contrasts reference-free judging, in which the model assesses an output using only the stated criteria, with reference-based (or reference-guided) judging, in which a gold answer or scoring rubric is supplied.
Methods Surveys organize the methods for building LLM judges into prompting a general-purpose model and tuning a model specifically for judging. The simplest approach uses a general-purpose model as a judge with no additional training, so that the prompt alone specifies the task, the criteria, and the output format. Representative prompting-based methods include G-Eval, which prompts the model to produce chain-of-thought evaluation steps and then fill in a scoring form, and GPTScore, which instead scores a text by the generation likelihood the model assigns to it. A common pattern is to use a more capable model to evaluate the outputs of a weaker one, for example GPT-4 grading a 13-billion-parameter LLaMA model. A second approach fine-tunes openly available models specifically for judging, which is intended to be cheaper and more reproducible than proprietary judges accessed through an API: PandaLM is trained on judgments produced by GPT-3.5, and JudgeLM on judgments produced by GPT-4, while the Prometheus models are trained to score responses against explicit rubrics and were reported to approach GPT-4 when supplied with a reference answer and a rubric. Judging need not rely on a single text-only model. Multiple judges drawn from different model families can be combined into a panel or jury so that their votes offset individual biases, as in the panel of LLM evaluators, which is also discussed as a means of mitigating bias. The approach has also been extended to multimodal judges that assess text together with images or video, including the MLLM-as-a-Judge benchmark and the open Prometheus-Vision.
Reliability The reliability of an LLM judge is the extent to which its verdicts agree with human judgment and remain stable, unbiased, and reproducible. LLM judges exhibit a range of systematic biases and vulnerabilities to manipulation that complicate their use as a substitute for human evaluation.
Biases Surveys of LLM judges describe several recurring biases, among them position (order) bias, verbosity (length) bias, and self-preference (self-enhancement) bias. Position, or order, bias is the tendency to favor a response according to its position in the prompt. An early 2023 study found that judges frequently favored the answer shown first and often reversed their verdict when the order was swapped; among the models it tested, even the strongest model tested, GPT-4, reached the same verdict both ways in only about two-thirds of cases. A later large-scale study found this effect to be systematic rather than a matter of chance, indicating it persists across models. Verbosity, or length, bias is a tendency to prefer longer answers independently of quality. The same 2023 study that documented position bias found that, when answers were rephrased to be longer without adding new information, judges such as Claude and GPT-3.5 still preferred the longer version more than 90% of the time. Self-preference, or self-enhancement, bias is the tendency of a judge to rate its own outputs, or those of its model family, more highly, an effect sometimes called "LLM narcissism". GPT-4 and Claude were observed to rate their own outputs more highly, although this early evidence was described as too limited to establish the effect firmly. Judges also tend to favor stylistically appealing answers regardless of their actual validity, and one study has proposed a taxonomy of twelve distinct bias types.
Calibration and reproducibility One survey lists overconfidence among the biases of LLM judges, describing a tendency to exhibit inflated confidence in their judgments. One study reported that judges such as GPT-4o and DeepSeek-R1 are poorly calibrated, expressing more confidence in their verdicts than their accuracy warrants; it measured this with the expected calibration error and the Brier score. Reproducibility is a further concern. Because language model generation is inherently random, a judge can return different scores for the same input on repeated runs, and its judgments are further destabilized by minor changes in prompt wording. Separately, the black-box nature and version dependency of proprietary judges accessed through frequently updated APIs hinder reproducibility, since the model behind the API may change over time.
… excerpt ends here. Continue reading the full article.
