A Survey on LLM-as-a-Judge

3 min read

The survey defines an LLM judge as an evaluation produced by an LLM from the item to be judged combined with a context (prompt, history), and then asks how to make that evaluation reliable. It sorts methods into in-context learning, model choice, post-processing and pipeline design, reviews tricks for improving judges, and organises meta-evaluation into agreement with humans, bias and adversarial robustness. It also runs its own check of common tricks on LLMEval2 (2,553 human-labelled response pairs) and EvalBiasBench (80 samples over six bias types), with position bias measured by swapping the two candidates.

Findings

  • Bias catalogue. Task-agnostic: diversity, cultural, self-enhancement. Judgment-specific: position, compassion-fade (swayed by model names), style, length or verbosity, concreteness or authority. A cited example: Vicuna-13B beats ChatGPT when ChatGPT is the judge, simply by being placed second.
  • Attacks cited: learned phrases that inflate scores, a “null model” with a constant answer that still wins often, and tokens like ”:” or “Thought process:” that fool judges.
  • In the authors’ experiment, agreement with human labels is close across models: GPT-4-turbo 61.54%, GPT-3.5-turbo 54.72%, Qwen2.5-7B 56.54%, LLaMA3-8B 50.72%. Position consistency is 80.31% for GPT-4-turbo and 38.85% for LLaMA3-8B. On the length-bias set, GPT-3.5-turbo and LLaMA3-8B both score 20.59%.
  • Tricks on GPT-3.5-turbo: asking for an explanation drops position consistency from 68.78% to 48.97%; self-validation changes little; majority vote over five runs helps; mean or best of five does not.
  • A three-LLM vote depends on who sits on the panel. Swapping LLaMA3-8B for Qwen2.5-7B raises length-bias accuracy from 26.47% to 64.71% and position consistency from 32.28% to 70.98%.
  • Reasoning models (o1-mini, o3-mini, Gemini-2.0-thinking, DeepSeek-R1) are not reliably better than GPT-4-turbo at matching humans. When the human label is a tie, judge accuracy ranges from 13.20% to 42.31%.
  • Their practical advice for pairwise judging: a strong model, swapped positions and a majority over repeated runs.

Limits

  • The bias sets in their experiment are small (24 to 34 samples per type), run at temperature 0, mostly with GPT-3.5-turbo as the base judge.
  • The authors note that judges used to label large datasets get far less meta-evaluation than judges ranking models.
  • Controlled studies are hard: lengthening an answer also changes its style, and a judge that prefers its own answers may just be right.
  • Judging embodied agents gets one paragraph of future work and no experiments.

Relevance

For robots, an LLM judge would score plans, logs or a teammate’s report, and each bias has a physical twin: trusting the first report to arrive, the longest log or the most confident tone. Voting only helps when the voters fail differently, which is the byzantine-fault-tolerance question again. Feeds robust robotics and llm-evaluation.