Diagnosing LLM Judge Reliability: Conformal Prediction Sets and Transitivity Violations
2026-04-16 • Artificial Intelligence
Artificial IntelligenceComputation and LanguageMachine Learning
AI summaryⓘ
The authors studied how well large language models (LLMs) work as judges for evaluating natural language generation (NLG). They found that individual evaluation cases often show inconsistencies that are hidden when looking at overall numbers. To better understand reliability, they used a math method called split conformal prediction to measure uncertainty per document, discovering that some criteria like relevance are judged more reliably than others like fluency. Their methods consistently showed that the type of evaluation criterion matters more than which judge is used. The authors provide their tools and data for others to use and verify.
large language modelsnatural language generationautomatic evaluationtransitivity analysissplit conformal predictionLikert scoresper-instance reliabilityevaluation criteriaconsistencyrelevance
Authors
Manan Gupta, Dhruv Kumar
Abstract
LLM-as-judge frameworks are increasingly used for automatic NLG evaluation, yet their per-instance reliability remains poorly understood. We present a two-pronged diagnostic toolkit applied to SummEval: $\textbf{(1)}$ a transitivity analysis that reveals widespread per-input inconsistency masked by low aggregate violation rates ($\barρ = 0.8$-$4.1\%$), with $33$-$67\%$ of documents exhibiting at least one directed 3-cycle; and $\textbf{(2)}$ split conformal prediction sets over 1-5 Likert scores providing theoretically-guaranteed $\geq(1{-}α)$ coverage, with set width serving as a per-instance reliability indicator ($r_s = {+}0.576$, $N{=}1{,}918$, $p < 10^{-100}$, pooled across all judges). Critically, prediction set width shows consistent cross-judge agreement ($\bar{r} = 0.32$-$0.38$), demonstrating it captures document-level difficulty rather than judge-specific noise. Across four judges and four criteria, both diagnostics converge: criterion matters more than judge, with relevance judged most reliably (avg. set size $\approx 3.0$) and coherence moderately so (avg. set size $\approx 3.9$), while fluency and consistency remain unreliable (avg. set size $\approx 4.9$). We release all code, prompts, and cached results.