Who can we trust? LLM-as-a-jury for Comparative Assessment

A Large Language Model (LLM) as judge evaluates the quality of victim Machine Learning (ML) models, specifically LLMs, by analyzing their outputs. An LLM as judge is the combination of one model and one specifically engineered judge prompt that contains the criteria for the analysis. The resulting automation of the analysis scales up the complex evaluation of the victim models' free-form text outputs by faster and more consistent judgments compared to human reviewers. Thus, quality and security assessments of LLMs can cover a wide range of the victim models' use cases. Being a comparably new technique, LLMs as judges lack a thorough investigation for their reliability and agreement to human judgment. Our work evaluates the applicability of LLMs as automated quality assessors of victim LLMs. We test the efficacy of 37 differently sized conversational LLMs in combination with 5 different judge prompts, the concept of a second-level judge, and 5 models fine-tuned for the task as assessors. As assessment objective, we curate datasets for eight different categories of judgment tasks and the corresponding ground-truth labels based on human assessments. Our empirical results show a high correlation of LLMs as judges with human assessments, when combined with a suitable prompt, in particular for GPT-4o, several open-source models with $\geqslant$ 32B parameters, and a few smaller models like Qwen2.5 14B.

Large language models (LLMs) are increasingly used as judges to predict which of two candidate answers a human evaluator would prefer, providing a scalable alternative when human evaluation is too costly. Unfortunately, LLM judgments can exhibit systematic biases. Prior work has documented position, verbosity or model family of the candidate answers influencing the judge decision rather than their quality, while additional biases may remain undocumented. $\textit{Multicalibration}$ provides a natural framework to address this problem: each bias corresponds to miscalibration on an identifiable group, $\textit{e.g.}$, family bias manifests as overconfidence in answers generated by models belonging to the same family of the judge. The challenge is to fix miscalibration simultaneously across many, potentially unknown groups. To this end, we introduce $\mathrm{CaLLM}$, a post hoc framework that multicalibrates an LLM judge, thereby bounding all such biases simultaneously. We prove that $\mathrm{CaLLM}$ bounds the bias of every group whose error structure is captured by an available representation, and that augmenting it with explicit bias groups further reduces the residual bias. We then derive theoretical conditions under which $\mathrm{CaLLM}$ improves the judge's accuracy, rather than trading it for calibration. Finally, we evaluate $\mathrm{CaLLM}$ across pairwise preference benchmarks, showing that it reduces the multicalibration error on both specified and unspecified bias groups, while improving predictive performance of the judge.

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.