Judging LLMs on a Simplex

Patrick Vossler, Fan Xia, Yifan Mai, Jean Feng

Introduction

The rapid advancement of Large Language Models (LLMs) has led to a shift in evaluation methodologies. Traditional metrics like BLEU or exact-match accuracy are increasingly inadequate for complex, open-ended tasks such as mathematical reasoning or creative writing. Consequently, the industry has adopted the "LLM-as-a-judge" approach, where a high-capability model (e.g., GPT-4o) evaluates the outputs of other models based on a predefined rubric.

The LLM-as-a-judge paradigm: candidates are evaluated against a rubric by multiple judges.

However, this paradigm relies on the assumption that the judge is either perfectly accurate or that its biases are negligible. Standard uncertainty quantification methods, such as bootstrapping, typically only address aleatoric uncertainty—the random noise stemming from finite sample sizes. They fail to account for epistemic uncertainty, which arises from our lack of knowledge regarding the judge's inherent quality and systematic biases. If a judge is consistently biased, rankings may be fundamentally flawed regardless of the number of samples collected.

The Simplex Perspective

To address these foundational issues, we represent the LLM-as-a-judge problem through the geometry of an $(M-1)$-dimensional probability simplex. For an $M$-level scoring system (e.g., a 3-point scale of "Incorrect," "Partial," "Correct"), every candidate model and judge can be mapped as a point or a set of vertices within this space.

Mapping scores to a probability simplex for 2-level (a) and 3-level (b, c) scoring systems.

This geometric mapping allows for a formal analysis of identifiability theory. We demonstrate a "phase transition" in the mathematical recoverability of model rankings. In 2-level (binary) scoring systems, the true ranking of candidates is often identifiable even with a mediocre judge. However, in systems with 3 or more levels, the geometry reveals that multiple, equally plausible judge configurations can lead to diametrically opposed candidate rankings. This provides a rigorous mathematical basis for the "folk wisdom" that researchers often prefer binary rubrics for robustness.

A Bayesian Framework for Robust Evaluation

We introduce a hierarchical Bayesian framework designed to quantify both aleatoric and epistemic uncertainty. The framework incorporates several key components:

  1. Random Effects: We relax the assumption that a judge's performance is constant across all candidates, allowing for variability in judge behavior.
  2. Weight Propagation Graph: This structure encodes prior beliefs about judge quality, allowing the model to leverage existing knowledge while remaining flexible.
  3. Sensitivity Analysis: By systematically varying hyperparameters, we can determine if candidate rankings are stable or if they "flip" under different assumptions about judge bias.

Experimental Results

The framework was evaluated across five diverse benchmarks, including GPQA, MMLU Pro, MTBench, SummEval, and Omni-MATH. These datasets represent varying degrees of verifiability and scoring complexity.

Our Bayesian approach demonstrated significantly superior coverage compared to traditional methods. On the GPQA dataset, the Bayesian method achieved a 77.8% coverage rate (the probability that the true ranking lies within the estimated interval), whereas bootstrapping and Bradley-Terry models achieved only 50% and 44.4%, respectively. This suggests that accounting for epistemic uncertainty provides more "honest" and reliable confidence intervals.

Furthermore, sensitivity analyses highlighted that while rankings are stable for many datasets (like MTBench), they are highly "fragile" for others (like Omni-MATH). In the latter case, models cluster closely in the simplex, meaning minor shifts in judge bias can reorder the leaderboard. The model also successfully identified and corrected for "self-preference bias," where LLM judges favor models from their own developer family.

Conclusion

By formalizing LLM evaluation within a geometric simplex, we provide a new lens through which to view judge-based metrics. While multi-level scoring offers nuance, it introduces significant mathematical non-identifiability that must be managed through robust Bayesian priors and sensitivity analysis. Future work will explore extending this framework to pairwise comparisons and automating prior tuning using small sets of human-verified "gold standard" data.

Summary updated

BibTeX
@inproceedings{vossler2026judging,
  author = {Patrick Vossler and Fan Xia and Yifan Mai and Jean Feng},
  title = {{Judging LLMs on a Simplex}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2026},
  url = {https://openreview.net/forum?id=HxmC9vLZE3},
}
Figure