The article examines how classical test theory statistics—Kuder‑Richardson coefficient, dependability index, and Livingston‑Lewis accuracy—can mislead when applied to large language model (LLM) judges that are evaluated with a single prompt and no gold labels. Using Claude Haiku 4.5 on 210 short‑answer items, the authors show that these metrics fail to isolate the judge’s performance because the judge’s single administration provides no variance component. They argue that reliable statements about an LLM judge require gold labels or varied scorer facets, and that bank design heavily influences reliability estimates.
By Louis Yiven Zhu
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.
By Atul Anand
arXiv:2607. 28685v1 Announce Type: new Abstract: Agent-safety benchmarks measure different behaviors, and their scores get quoted interchangeably as an agent's safety.
By Youting Wang, Xiao Han, Dingyan Shang, Yuan Tang, Bowen Liu
The paper investigates whether large language model (LLM) safety judges are truly policy‑invariant, meaning their verdicts should depend only on the agent’s behavior and not on how the evaluation policy is phrased. It introduces three testable principles—rubric‑semantics invariance, rubric‑threshold invariance, and ambiguity‑aware calibration—and applies them in a stress‑test protocol across four agent‑class judges using trajectories from ASSEBench and R‑Judge. The study finds that current judges often change verdicts in response to both meaningful normative shifts and meaningless structural rewrites, with up to 9.1% of verdicts flipping due to content‑preserving rewrites and 18‑43% of flips occurring on unambiguous cases, indicating that safety scores conflate agent actions with evaluator wording.
whyItMatters":"The findings reveal a hidden failure mode in safety evaluation that could mislead researchers about agent performance, underscoring the need for policy‑invariant judges and transparent reporting protocols."
By Shihao Weng, Yang Feng, Xiaofei Xie
arXiv:2610.07177v1 Announce Type: cross
Abstract: An open contrastive decision model is near chance as a judge on the hard public benchmarks: Contrastive-LM/CLM-v0.1-8B scores between 0.351 (best- of...
By Gowthamkumar Nandakishore
The paper introduces a portable validity screen for evaluating large language model (LLM) confidence signals, mirroring clinical personality assessment practices. It computes metrics such as the rate of high‑confidence errors (L), low‑confidence correct answers (Fp), and their contrast (RBS), without adding new statistics, to determine whether a signal truly tracks correctness. Applied to 20 frontier models on 452 two‑alternative items and 33 models on 1,500 MMLU items, the screen classifies models as Valid, Indeterminate, or Invalid, revealing that many models fail to meet validity criteria depending on confidence thresholds.
By Jon-Paul Cacioli
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.
By Jianlin Chen, Wenhui Chen, Ziyao Lin, Chi Man Vong
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.
By Idil Gozel
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."
By Delip Rao, Chris Callison-Burch
The paper investigates how different scoring rules affect the ranking of large language models (LLMs) on the IndiaFinBench benchmark, which contains 406 expert‑annotated items covering regulatory interpretation, numerical reasoning, contradiction detection, and temporal reasoning. Using twelve LLMs evaluated under strict string‑matching, a cross‑model judge, and a composite regime, the authors find that model rankings vary dramatically across regimes, with no significant rank correlation (Spearman ρ = -0.273). Item‑level analysis reveals that a substantial portion of items are answered identically by all models, while a higher‑discrimination band isolates more challenging cases.
By Rajveer Singh Pall
The paper evaluates 21 LLM-as-a-Judge models from nine providers across MT-Bench, JudgeBench, and RewardBench, using agreement, consistency, and bias audit protocols over 118 runs and about 541,000 judgments. It finds that exact-match agreement overstates performance, with Cohen’s kappa lower by 33–41 percentage points, judge rankings shift up to 14 positions, and high test–retest reliability coexists with significant position bias in some production judges. Verbosity bias is minimal across the cohort. The authors propose a Minimum Viable Validation Protocol based on these findings.
By Justin D. Norman, Michael U. Rivera, D. Alex Hughes
The paper presents a closed‑form estimator for the contamination correlation between anchors and judges under a single‑common‑factor model, requiring at least two judges and two anchors. It introduces a diagnostic battery—including judge‑covariance dispersion, over‑identification tests, a family‑block test, bootstrap confidence intervals, and a weak‑identification screen—to validate the estimator and detect violations. The authors also discuss identification limits for ordinal data and report that real panels have not yet passed the model‑adequacy pre‑test, while simulation studies confirm the estimator’s performance.
By Veerendra Kumar Sunkavalli