Large Language Models (LLMs) are increasingly used as automated judges, yet decoding temperature is often set by intuition rather than evidence. We run over one million large scale evaluations, multiple open-source models, prompting strategies, and judging paradigms to examine how temperature affects judge reliability. Low temperatures yield consistent outputs with few errors, especially in smaller models, while higher temperatures reduce consistency and increase errors. Yet consistency alone does not guarantee sound judgments: low temperatures can narrow the range of reasoning explored. We characterize judge behavior through four reasoning personas and find that temperature shifts their distributions in model-dependent ways, most visibly in ambiguous cases where competing answers are hard to distinguish. Repeated high-temperature sampling reveals this hidden uncertainty, and combining it with majority voting or using high temperature for reasoning and low temperature for the final decision, improves agreement on ambiguous cases without hurting clear ones, at the cost of extra inference. Our findings suggest a simple rule: keep temperature low for stable outputs, but explore more when the judgment is hard.
Figures & tables
Figure 1: LLM-as-a-Judge Paradigm
Figure 2: Pearson correlation matrices between decoding temperature (T) and evaluation metrics—agreement (A), consistency (C), and error rate (E)—across four LLMs.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Figure 3: Comprehensive ATE Analysis Across Moderators, Temperature, and Feature Importance. Row 1: Temperature trends of different models (shared legend), showing mean tendencies across temperature ranges (lines indicate averages, not ATE values). Row 2: Boxplots of Average Treatment Effect (ATE) distributions across three moderators for each outcome metric (dashed line denotes ATE = 0). Row 3: SHAP beeswarm plots ranking feature importance by impact magnitude for each outcome metric. All panels share consistent bold styling and bright color palettes for clarity.
Large language model (LLM)-as-a-judge research uses language models to score, compare, and verify AI outputs as scalable alternatives to human evaluation.
LLM-as-judge ("grader") components are now standard in evaluation harnesses, including safety evaluations where a pass/fail verdict may gate downstream deployment decisions. A widespread assumption is that setting the grader's sampling temperature to 0 makes grading deterministic. We test this assumption against a real safety-evaluation codebase (Japan AISI's open-source aisev) and show it fails on two levels. First, the harness invokes its grader without setting temperature or seed; the underlying provider silently applies its default of 1.0, so items near the decision boundary flip pass/fail across identical runs (per-item disagreement up to ~50% over 20 runs). Second, pinning temperature=0 reduces but does not eliminate flips: across 690 API calls spanning two providers, three model tiers, and five sampling configurations, 1-2 of 7 borderline items remain non-reproducible even under forced greedy decoding (top_k=1). Claude Opus 4.7/4.8 has since deprecated temperature entirely, rendering the primary mitigation inapplicable to newer model generations. These findings expose a structural gap: evaluation harnesses that report single-run verdicts without variance or grader-disagreement metrics can present noise as a safety property. We release a reproduction harness (690 calls, 7 conditions) and recommend that harnesses treat grader disagreement as a first-class health metric alongside the scores themselves.
Researchers increasingly use Large Language Models as judges (LLM-as-a-judge) to evaluate model outputs. Yet there are no standards for how to design these judges. Typically, researchers choose the prompt, rating scale, and model intuitively. If these choices change the judge's verdicts, two studies can reach different conclusions about the same facts. To address this risk and to provide an empirical basis for judge designs, we evaluate 10 reasoning models across multiple designs on two tasks: a scalar rating of sentence sentiment and toxicity (over 500 items per category), as well as a binary accuracy classification of question-answer pairs (n=600). For the rating tasks, despite judges showing significant disagreements with the human ground truth, the practical size of differences is small enough to consider most judges reliable (mean absolute deviation of 0.11 points on a 1 - 7 scale); toxicity judges even outperform standard classifiers. Judges are also highly accurate on average (96.5%) for the accuracy classification task. However, design choices can produce shifts: changing the rating scale alone can shift measured bias by up to 0.93 points (rating task), and while accuracy levels are rarely impacted, design choices consistently impact judge leniency (classification task; leniency drop of 28.9 percentage points when using detailed prompts, and up to 56.1 percentage points when switching models). Counterintuitively, lower reasoning effort affects neither accuracy nor leniency. Across both tasks, model identity is the dominant source of variance. These findings suggest that while LLM judges are broadly trustworthy in aggregate, design choices can be meaningful sources of variance. Given the growing reliance on automated evaluation in LLM research, we intend this study as a methodological reference for designing more robust and replicable LLM-as-a-judge pipelines.
Modern LLM evaluation assumes that pinning a judge to a fixed model version and decoding at temperature zero yields reproducible verdicts. We show this assumption fails as a property of how LLM-as-Judge is operationalized on cloud serving infrastructure, not of any particular model family. Across four frontier judges served via a single major enterprise cloud platform and three standard benchmarks (Arena-Hard, AlpacaEval 2, MT-Bench), identical inputs to the same temperature-zero judge, at a constant serving-reported model version, produce different verdicts across re-runs: per-item flip rates of roughly 5% on average and about 40% on the close-call items that decide leaderboard margins, with a per-judge magnitude spanning a 40x range (0.13% to nearly 10%). We introduce metrics tailored to this instability: per-item flip rate, a two-part stability profile (waver fraction and conditional intensity), and adjacency separability. For the principal judge the aggregate ranking is stable (0% top-K instability, 0% pooled winner flip); what degrades is precision: under a paired hierarchical bootstrap, roughly one-fifth to three-quarters of adjacent leaderboard positions are statistically indistinguishable, a noise floor driven mainly by finite prompt sampling rather than the judge. Across judges, leaderboards agree on the coarse ordering but diverge in the middle (Kendall's tau of 0.42-0.64 between Gemini and Sonnet judges on Arena-Hard, values sensitive to answers truncated at the generation cap). Of 15 expected head-to-head orderings we re-judge, 12 survive every re-run of the principal judge but only 8 survive every judge. Leaderboards thus report unhedged point estimates that overstate their precision. We propose a minimal, low-cost reporting protocol: several judge re-runs, published stability profiles and adjacency intervals, and results under at least two judges from different families.