When the Judge Changes, So Does the Measurement: Auditing LLM-as-Judge Reliability

A Judge Should Know What Changed:Construct Validity for LLM-as-a-Judge Evaluation

The paper introduces a two‑dimensional construct validity framework for evaluating large language models (LLMs) as judges, defining invariance (S) and sensitivity (R) to construct‑preserving and construct‑changing edits. Experiments across seven judges and four domains reveal high invariance (average S = 0.945) but low sensitivity (average R = 0.319), with sensitivity varying by edit type. Audits of public label sets show that surface‑only predictors can reproduce a substantial portion of labels, underscoring that high agreement does not guarantee construct validity.

Auditing and Repairing LLM-as-Judge Failures in a Production Text-to-SQL Pipeline

The paper evaluates a production text‑to‑SQL pipeline that uses an LLM as a judge, finding that the deployed gpt‑4o‑mini judge agrees with human annotators only weakly (Cohen’s kappa 0.04 on a disagreement‑enriched set and 0.42 on a random spot‑check). The authors identify a specific failure mode, GRADE‑HALLUCINATION, responsible for most over‑flags, and demonstrate that a self‑hosted Qwen3.6‑27B model achieves substantially higher agreement (kappa 0.72) at a much lower cost. They also show that ensembling judges does not improve performance, and that their audit method flags a significant portion of out‑of‑domain SQLs as potential issues.

MAWILE: Multi-Axis Workbench for Inspecting LLM Evaluators

arXiv:2609.22599v1 Announce Type: new Abstract: Large language model (LLM) judges provide a flexible and scalable method for evaluating model and agent outputs, but their verdicts can be sensitive to...

Benchmarking LLM-as-a-Judge for Long-Form Output Evaluation

The paper introduces LongJudgeBench, a benchmark designed to evaluate large language models (LLMs) acting as judges for long-form text generation. It highlights that long-form evaluation requires complex, document-level assessments beyond simple length, such as organization, coverage, depth, consistency, and scenario-specific quality. Experiments show a significant reliability gap among current LLM judges, indicating instability across scenarios and limited effectiveness of rubrics or references.

Judge, Retrieve, or Abstain: Uncertainty-Guarded LLM Judging with Provable Risk Guarantees

The paper introduces a risk‑controlled framework for using large language models (LLMs) as judges in tasks without reference answers. By calibrating uncertainty thresholds on a held‑out set, the method ensures that the false discovery rate of accepted verdicts stays below a user‑specified level α with high probability, using finite‑sample Clopper–Pearson intervals. When the parametric judge lacks confidence, the instance is routed to a retrieval‑augmented mode with a second calibrated threshold, preserving the error guarantee while achieving higher coverage than single‑mode baselines.

JudgeSense: A Benchmark for Prompt Sensitivity in LLM-as-a-Judge Systems

JudgeSense is a benchmark comprising 880 items from human‑labelled corpora, each presented under two differently worded instructions that ask the same question. The study evaluates 25 judges from six providers across four tasks, measuring how rewording affects agreement with the judge’s own verdicts. Results show that rewording reduces agreement on all tasks, with significant effects on two, and that stability varies across tasks and is not predicted by parameter count.

Judge Arena: Benchmarking LLMs as Evaluators

Can We Trust LLM Judges: A Study of Capability-Dependent Biases and Multi-Judge Ensemble for Bias Calibration

The paper investigates how large language models (LLMs) used as judges in absolute scoring tasks exhibit systematic biases that compromise reliability. It shows that a judge’s task accuracy strongly predicts both its judging accuracy and its directional bias, yet more capable examinee models consistently receive more lenient judgments. To mitigate these biases, the authors propose a calibrated weighted majority voting (WMV) ensemble that estimates judges’ error rates from inter-judge agreement patterns, achieving near-oracle performance without labeled data and improving both accuracy and fairness.

SESSE: Sketch, Expand, Sort, Summarize, Evaluate -- LLM-as-Judge Evaluation via Structured Decomposition

The paper introduces SESSE, a training‑free framework that breaks down LLM‑as‑judge evaluations into five steps—Sketch, Expand, Sort, Summarize, Evaluate—by mining sub‑questions from the judge’s own error cases. It requires no oracle responses, task‑specific rubrics, or fine‑tuning, yet on RewardBench it matches the performance of chain‑of‑thought baselines and rivals a fine‑tuned specialist (RISE‑Judge‑32B). SESSE provides per‑criterion vote evidence, offering an interpretable audit trail that can diagnose label ambiguity and judge failure modes that a single holistic output token cannot reveal.

JEV vs. LLMs as Rubric Judges: Cheaper, Faster, and Wrong in the Same Places

The study investigates whether Jev, a typed classifier that outputs probabilities over allowed answers without generating text, can replace large language model (LLM) rubric judges. Across nine panels from seven benchmarks, Jev’s accuracy differed significantly from LLM judges in only 8 of 27 paired comparisons, performing best on binary criteria and worse only on graded ones, while most other comparisons were inconclusive. In terms of cost and speed, Jev was 29 to 325 times cheaper and 30 to 220 times faster than the flash‑tier LLM judges, and a cascade approach that defers uncertain Jev verdicts to an LLM yielded only modest gains. whyItMatters":"The findings suggest that a lightweight classifier like Jev can serve as an efficient first‑stage evaluator, potentially reducing the reliance on expensive and slow LLM judges in automated grading pipelines."

The First Token Is Not the Verdict: Hidden Costs of Reading LLM Judges Without Generating

The paper demonstrates that reading a large language model (LLM) judge’s verdict from the logits of its first generated token—an approach used in constrained decoding and likelihood‑scoring evaluation—introduces a significant distortion in position bias. Because judges do not always start with a verdict token (12–49% of cases for Qwen3 judges and <3% for Llama‑3.1‑8B and Phi‑3.5‑mini), this readout often returns the first response rather than a true judgment, inflating position bias by up to 42 points while barely affecting judge accuracy. The authors recommend reporting the frequency with which a judge leads with a verdict token to provide a more accurate assessment of position bias.

Accounting for Bias Enables Sustainable LLM Evaluation

The paper argues that the current LLM-as-a-judge evaluation method, which compensates for systematic measurement bias by increasing the number of comparisons, is statistically unsound and computationally wasteful. It identifies that treating LLM judges as neutral ignores documented biases such as position bias, verbosity bias, judge severity, and self‑enhancement. The authors propose a unified latent variable framework that jointly models pairwise and ordinal data while explicitly correcting for these confounders, enabling reliable rankings with far fewer comparisons and negligible additional compute.