The paper investigates the reliability of LLM judges in the Arena-Hard benchmark, showing that aggregate scores can mask significant variability in individual verdicts. By running GPT‑4.1 across 500 prompts under eight different judge configurations and repeating each setup, the authors find that small changes—whether in temperature, criteria, or instruction framing—can shift win/tie/loss outcomes on up to 28% of prompts and alter aggregate scores by several points. The study concludes that stable aggregate scores do not guarantee repeatable individual judgments, and recommends that evaluations report both aggregate results and prompt‑level agreement.
arXiv:2606. 13685v1 Announce Type: cross Abstract: LLM-as-a-Judge is now widely used to rank model outputs, train reward models, and populate public leaderboards, but its run-to-run reliability remains under-characterized.
The paper investigates whether widely used evaluation frameworks for large language models (LLMs) implement a defense called commit‑first judging, which requires a judge to solve a task itself before accepting a candidate answer. Across 24 configurations in eight popular frameworks, none use the full commit‑first method; nine use a weaker variant that is ineffective. In controlled experiments, the weaker variant allowed systems to game the judge, while the full commit‑first approach eliminated this vulnerability but sometimes worsened evaluation when the judge’s own answer was incorrect.
LLM judges, models that score another system's output, can be gamed by the systems they score. Recent work identifies one defence that works: the judge solves the task itself first and commits to that...
The paper introduces TRACE, a protocol designed to diagnose whether changes in verifier scores for large language model agents reflect actual changes in agent behavior or merely alterations in the evaluation process. TRACE works by applying targeted changes to evaluation components, running paired experiments, and re‑scoring unchanged trajectories to isolate the source of score variation. Experiments on synthetic tasks and public benchmark tasks demonstrate that seemingly significant score shifts can often be attributed to evaluation artifacts, while TRACE can also detect genuine behavioral changes.
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.
arXiv:2606. 15474v1 Announce Type: new Abstract: Continuous evaluation of LLM products relies on a strong LLM judge treated as ground truth: a cheap monitor scores every interaction and a team is paged when the score drifts down.
The paper introduces the Wiggle Framework, a unified stress test for assessing epistemic stability in large language model (LLM) judges. It evaluates judge robustness across three dimensions—Mechanical Consistency, Single-turn Conviction, and Multi-turn Persistence—using 9 frontier models on 14 judging tasks related to safety, toxicity, AI writing detection, and political-response evaluation. Results show significant instability, with verdict flips ranging from 25–71% under static pushback and 62–91% when challenged by an adversarial LLM, and highlight that successful pressure often misaligns with ground truth.
The study analyzes 373,019 judgments from LLM‑scored benchmarks, decomposing variance into system, item, judge, and interaction components via generalizability theory. It finds that with a single judge, generalizability converges to a ceiling determined by the system‑by‑judge variance, which is substantially lower in pairwise preference settings, allowing one judge to suffice. The research also reveals significant biases in presentation order and highlights that many published win‑rate claims fall below the measured floor of the benchmarks.
arXiv:2608. 12645v1 Announce Type: new Abstract: LLM judges have become central infrastructure for model evaluations, online grading, and reward modeling.
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.
The paper introduces Conditional Accuracy Profiling (CAP), a diagnostic framework that breaks down LLM‑judge accuracy into eight deployment conditions covering content sensitivity, robustness, and rationale quality. CAP is benchmark‑agnostic and can be applied directly when annotations are available, via task‑subset proxies, or through controlled augmentation. Applied to seven judges across six benchmarks, CAP reveals that aggregate accuracy masks important differences—for example, judges highly sensitive to omitted qualifications rank poorly on a Hard‑Constructed subset, and Position Robustness is generally stable but can degrade sharply under adversarial stress.
arXiv:2610.18760v1 Announce Type: new Abstract: Test-time scaling for large language models (LLMs) can involve generating multiple candidate solutions for selection or revising an earlier attempt. To...