LLM-as-a-Judge

Our team, VANGUARD, presents IROH (Insightful Ranking of Humor), a three-stage retrieval system for JOKER Task 1 English at CLEF 2026, achieving first place on the leaderboard with 0.6347 MAP. Our pipeline combines hybrid sparse-dense retrieval, cross-encoder reranking, and a LoRA-adapted Large Language Model judge ensemble. We employ Gemma 4 to generate query-aware rationales under two prompt strategies, generic and typed, and produce up to four types of structured hard negatives for training data construction. Through an ablation across three cross-encoder architectures, four dense embedders, and eight judge configurations, our key findings are threefold: (1) the rationale-distilled judge is the primary driver of ranking quality, whereas appending rationales to the first-stage index contributes negligibly; (2) structured hard negatives degrade generalisation in nearly all configurations despite inflating local validation scores; and (3) across the components we ablate, the lighter, better-calibrated model is competitive with or stronger than its larger counterpart, with the generic-rationale Qwen2.5-7B judge (0.6055 MAP) outperforming every Gemma-4-31B configuration, and the advantage of generic over typed rationales is concentrated almost entirely in the smaller model.

Evaluating the factual correctness of large language models (LLMs) is vital for many applications. But are our evaluation tools themselves trustworthy? Despite the rise of factuality-based metrics, their sensitivity and reliability remain underexplored. This paper introduces a meta-evaluation framework that systematically tests these metrics using controlled corruptions of gold standard answers. Our method generates ranked outputs with known degrees of degradation to probe how metrics capture nuanced changes in truthfulness. Our experiments reveal that pipeline-based methods, such as the RAGAS's factual correctness metric, better track degradation than LLM-as-judge approaches. We also propose a new variant of the factual correctness metric that provides a competitive and cost-efficient.

Synthetic perturbations appear to offer inexpensive calibration data for LLM evaluators in biomedical ML, where expert review is scarce. Yet a planted mutation key is neither a detector output nor automatically human ground truth. We formalize four distinct ledgers: planted perturbations, independent detector outputs, source-linked human dispositions, and human-added discoveries. We then audit the evaluation design, scoring code, read paths, and current human records of a private synthetic Japanese care-handoff workflow. The factory stored 69 planted error cards across 47 targets. Final review covers 22 targets and contains 22 confirmed imported proposals, 9 rejected proposals, and 79 human-added cards; only 3 reviewed targets are double annotated. Passing imported plant keys to a generic detector scorer yields 22/(22+9)=0.710 and 22/(22+79)=0.218. A direct audit identity shows that these values are proposal-confirmation yield and submitted-ledger composition, not judge precision and recall, because no independent detector realization was preserved for the audited proposals in the available records. The audit also finds source-name collisions, row shadowing, forced severity, vacuous ratio defaults, and unsupported zero-support field weights. We contribute a provenance-aware claim audit, a storage contract, and a minimum calibration gate for responsibly communicating biomedical ML capability claims. This single-workflow forensic case is an existence proof of a failure mode, not an estimate of its prevalence: existing human work supports an exploratory audit of synthetic proposals, but not LLM-judge operating characteristics, clinical validity, corpus prevalence, or robust inter-annotator agreement.

LLMs are increasingly used as automated judges for model training and evaluation, yet individual judges exhibit systematic biases that undermine reliability. Much of prior work has studied biases in pairwise LLM-as-a-judge settings; in this paper, we focus on absolute scoring tasks, which mirror more realistic use cases. Across four benchmarks and six models (36 judge-examinee pairs), we show that a model's task accuracy strongly predicts its judging accuracy (Pearson $r \geq 0.90$ on most models) and inversely predicts its directional bias ($r \leq -0.83$), but that accuracy alone does not ensure fair evaluation: more capable examinee models consistently receive more lenient judgments from all judges ($r \geq 0.83$). To address this, we propose calibrated weighted majority voting (WMV), an ensemble evaluation method that aggregates multiple LLM judges weighted by online estimates of their false-positive and false-negative rates. We introduce a disagreement-based estimator that derives these error rates purely from inter-judge agreement patterns, requiring no ground-truth labels or task metadata. In a simulated experiment with shifting task distributions, our label-free WMV tracks an oracle with perfect error-rate knowledge to within 0.5 percentage points on average, outperforming both individual judges and unweighted majority voting. These results demonstrate that principled multi-judge calibration can simultaneously improve accuracy and correct for systematic leniency without requiring labeled data, offering a scalable path to reliable automated evaluation as model capabilities increase.

Comparing and selecting task-oriented LLM agents increasingly relies on a low-cost offline evaluation gate: persona-driven LLM user-simulators converse with each candidate, an LLM-as-a-judge scores the transcripts, and the higher-scoring agent is promoted. We introduce GAUGE, a reusable offline protocol that measures whether this gate's ranking matches a grounded verifiable reward across 25 agents from six providers on the $\tau^2$-bench and SimulatorArena benchmarks, separating two kinds of evaluation validity that release practices conflate: ranking validity and construct validity. First, a satisfaction-success gap: satisfaction carries essentially no information about task success, as conversations rated satisfied by our blind panel are decorrelated from actual success, with 57.5% of them failing the customer's task, a pattern consistent across five rater populations, both benchmarks, and every subjective dimension we rated. Second, while the gate's ranking is robust across the broad capability span, it loses resolution among the near-equal strong agents: this decision-disagreement rate jumps from $<$1% on wide-reward pairs to 31% on close pairs. The gate is thus human-validated yet mis-anchored. As a remedy, we propose a calibrate-then-trust cadence in which a judge-free completion bit is a zero-cost tripwire for truncation regressions.

LLM-as-a-judge protocols are commonly debiased by instructing judges to ignore presentation cues such as citation formatting, source labels, and evidence-display style. We show that this intervention can suppress bias while damaging the resolution of the measurement instrument. We introduce TraceJudgeBench, a diagnostic benchmark for auditing citation-like artifacts in RAG and agent-workflow evaluation, covering content-equivalent pairs, citation ablations, correctness conflicts, human-validated soft and moderate quality gaps, prompt-strength ladders, decoupled judging, and a controlled workflow-ranking probe. Across GPT-5.5, Claude Sonnet 4.6, and DeepSeek V4-Flash, stronger anti-citation prompts reduce worse-cited wins from up to 50.5% to 0%; yet some operating points already convert validated moderate-gap decisions into Tie before the strict stress-test endpoint, while correctness-conflict accuracy remains at or above 93.0%. A second, 50-pair FinQA moderate-gap construction reproduces the qualitative frontier, and open-weight Qwen2.5-14B-Instruct-AWQ and Gemma-3-12B-IT runs reproduce the central HotpotQA frontier. TRACE-style decoupling recovers 96.5-100.0% better-plain resolution across the reported settings. Human validation separates three meanings of Tie: correct equivalence Tie, calibrated soft-boundary Tie, and resolution-destroying Tie on validated quality gaps. We frame debiasing as a measurement intervention whose bias suppression, resolution retention, Tie cost, and protocol cost must be reported jointly. The supplementary artifact contains benchmark splits, prompts, raw judge outputs, validation summaries, and analysis.

As large language models become capable translators of classical texts, a key challenge is deciding which outputs need expert review when no human reference exists. This study tests reference-free error triage through Pali-to-English translation. Three LLMs translated 15,493 passages. Five signals were compared: source novelty, source-candidate embedding distance, peer-translation disagreement, English-to-Pali backtranslation, and no-reference GEMBA scoring. Signals were calibrated on a 3,000-item reference-informed LLM-adjudicated sample and checked against a 500-item author-adjudicated anchor. Human references supported calibration and validation only; they were never used to compute the risk signals. Source novelty was a useful source-side risk prior but not a per-candidate error detector. Peer disagreement and backtranslation provided secondary signal. The strongest method was no-reference GEMBA scoring by a panel of models generally regarded as stronger than the translators: reviewing the top 10% by GEMBA risk captured 81.6% of panel-major errors in the calibration set. GEMBA also remained the best reference-free signal against the author anchor. A same-tier panel, with self-scoring excluded, remained useful but performed worse, indicating that evaluator strength matters beyond the prompt alone. A budgeted workflow is proposed, combining source novelty, peer disagreement, and stronger candidate-aware judging to allocate human review. Transfer to other classical languages, including Latin, Ancient Greek, and Sanskrit, remains to be tested.

LLM judges are increasingly used to evaluate and improve AI-generated outputs, yet their reliability for complex professional work remains unclear. We study this problem through Vibe Patenting, an end-to-end patent-drafting testbed for AI-agent evaluation. A separately-invoked LLM judge evaluates generated patent drafts and provides structured feedback for iterative revision. Across multiple inventions and drafting-agent configurations, judge-guided revision consistently improves judge-assessed quality, while unguided revision tends to saturate. Notably, iterative judge feedback enables a low-reasoning agent to approach the performance of a substantially more expensive high-reasoning agent. Stronger models and increased reasoning generally improve judge-assessed drafting quality, while domain-specific agentic workflows provide further gains. We validate the judge against independent evaluation by a professional patent attorney and find meaningful but strongly metric-dependent agreement and systematic calibration differences. These results highlight both the utility and limitations of LLM judges as evaluators and optimization signals for complex professional workflows.

Blinded physician evaluation has been considered by many to be the gold standard for assessing clinical reasoning in large language models (LLMs). This is difficult to scale; thus, prior studies typically rely on small physician panels, often from a single institution or specialty, which both limits the scientific questions investigated and makes it unclear whether findings would be reproduced with a different set of evaluators. To more rigorously and scalably study clinical reasoning in AI models, here we introduce PrecepTron, an LLM fine-tuned for physician-level evaluation of open-ended responses. PrecepTron was trained using low-rank adaptation (LoRA) of a 32-billion-parameter model on a small number of physician examples. We also release GRAND-ROUNDS, a new large-scale physician-annotated benchmark of 9,217 scored responses from 160 clinicians across seven studies. We show that frontier LLMs in typical "LLM-as-a-judge" approaches often disagree with physicians and with each other, but fine-tuning PrecepTron on a small number of cases enables physician-level consistent scoring across tasks. We use PrecepTron to reproduce headline findings from five influential studies assessing LLMs for clinical care in JAMA, Science, and Nature Medicine without new human grading. Using PrecepTron, we then pose new questions about how LLMs reason in medicine that would have been infeasible with human grading alone, including measuring the diagnostic accuracy of frontier LLMs when clinical cases are provided piecemeal, even token by token. Together, PrecepTron and GRAND-ROUNDS provide a foundation for reproducible, large-scale study of how LLMs reason in medicine. All code, data, and labels are made freely available for researchers.

Verbalized confidence, long dismissed as overconfident, coarse, and prone to round-number clustering, is now the more robust soft-scoring mechanism for LLM-as-a-Judge on top-tier proprietary models. Across SummEval, AggreFact, and HelpSteer2, spanning up to 18 LLMs, we show that the standard advice to prefer log-probabilities no longer holds on post-2025 models, where verbalized confidence is the better signal. We call this a compatibility shift. On top of a standard verbalized-confidence baseline, we introduce two new ingredients: an overconfidence advisory and self-debate. Together they improve calibration, score-distribution spread, and robustness to task subjectivity. We further observe a generation effect: post-2025 models accommodate these two additions with little balanced-accuracy cost, whereas pre-2025 models pay a measurable penalty. Compared with logprob-based G-Eval, verbalized confidence is the more subjectivity-robust soft signal on GPT-family top-tier releases. The shift is invisible under accuracy-only reporting. Rather than defaulting to hard predictions, we recommend broader use of soft scoring in LLM-as-a-Judge. More broadly, verbalized confidence has moved from a weaker substitute for logprobs to a practical soft-scoring mechanism for contemporary LLM judges.

Large language models are increasingly used as judges to measure social bias in text, yet the passages they judge are often noisy, containing typos, informal spelling, and broken punctuation. The consequences of such surface noise for social bias measurement remain unclear. To investigate this question, we apply five realistic noise conditions at multiple intensity levels to 3,822 stereotype-related responses and compare the resulting bias judgments with those on the original text. We find that such surface noise does not degrade bias measurement symmetrically: it is far more likely to turn neutral judgments into biased ones than biased judgments into neutral ones, by up to a 120x margin. We further observe two non-obvious effects across four LLM judges: in the most fragile judge the distortion is at its purest at mild, realistic noise levels, where erasure is scarcest, and as judges grow robust it attenuates toward parity rather than reversing. Bias measured on noisy text is therefore systematically overestimated, most in the categories that matter most for fairness.

Evaluating the quality of explanations produced by explainable AI (XAI) methods remains challenging because existing approaches often rely on subjective human judgment, limiting reproducibility, scalability, and comparability between studies. We examine whether LLMs can serve as a reproducible and scalable mechanism to make comparative assessments of the quality of XAI explanations. We introduce XAI-Arena, an LLM-as-a-judge framework for scalable, reproducible, multidimensional, and stakeholder-sensitive evaluation of XAI explanation quality. XAI-Arena then allows us to compare XAI explanations along various dimensions, namely, perceived simplicity, clarity, task adequacy, trust calibration, actionability, transparency, faithfulness, and overall interpretability. We then benchmark XAI explanation methods across various datasets, machine learning models, and stakeholder personas. Human validation shows a strong positive association between LLM-generated and human ratings (Spearman's rho=.693, p<.001). Together, LLM-based evaluations can capture systematic differences in XAI explanation quality and provide a scalable and reproducible framework for comparative assessment of XAI explanations.

When an external reference set (an anchor) is used to decompose an LLM-judge panel's error into a quality signal and a shared common-mode error, standard practice assumes the anchor is uncontaminated: its error uncorrelated with the judges' shared error. We study when that assumption can be dropped and replaced by an estimate. Under a single-common-factor model, >=2 judges and >=2 anchors point-identify the quality variance, the common-mode variance, and each anchor's contamination correlation rho_k in closed form, with an exact per-anchor-pair failure boundary; a designated clean-anchor estimator, by contrast, reports a contaminated companion anchor as fully clean once its trusted anchor is itself contaminated. Because the single-common-factor assumption is itself untestable, the estimator ships gated behind a calibrated diagnostic battery (judge-covariance dispersion; over-identification; a family-block test from judge metadata, with a family-blocked estimator that removes family-level shared-residual bias exactly), bootstrap confidence intervals with measured coverage, and a weak-identification screen. A proposition maps which violations bias rho_k, in which direction, and which evade detection. For ordinal scores we show an identification hierarchy: with all variables ordinal, rho_k is not identified at any number of anchors; with ordinal judges and >=3 continuous anchors it is, and we give an estimator for that case. On real data the validation is asymmetric, and we say so plainly: the diagnostics are validated in the rejecting direction (both real panels we test are correctly rejected by the model-adequacy pre-test), while the estimator is validated in simulation and stress-tested semi-synthetically under oracle calibration; no real panel has yet passed the pre-test, and the pre-test exists precisely to say so. All results replay offline from shipped, checksummed artifacts.

Large language models have collapsed the cost of producing lexically elaborate prose, and whether peer reviewers still reward it is a question about the evaluator, not about the text. When the association between a writing cue and review scores moves across years, the reviewers may have changed, the submissions may have changed, or both, and a regression of scores on text cannot say which. We separate the two with a frozen rater: 81,850 machine reviews of ICLR submissions from 2018 to 2025, all generated in one February-April 2025 window with one model family and one prompt, so that its year-to-year coefficients track submission composition alone and the human-minus-frozen trend difference identifies reviewer preference drift. On 32,638 submissions with 124,615 human reviews, the human coefficient on non-domain lexical complexity falls from +0.142 to -0.015 while the frozen rater moves from +0.080 to +0.082; the three-way difference-in-differences is -0.0100 (q=0.013), and forty random-wordlist placebos through the same specification centre on zero. Humans still reward sentence-length variability, which the frozen rater never registers, while the frozen rater still pays for lexical complexity at its earlier rate. Every claim is held to a double gate of false-discovery control and interval exclusion, and the findings that failed adversarial re-testing are reported. Reviewers discounted a cue whose production cost collapsed, as models of manipulable signals prescribe; an LLM judge calibrated to historical human preferences inherits the earlier schedule and drifts out of alignment while its agreement with humans on totals stays ordinary.

Automatic safety judges -- systems such as Llama Guard or a GPT-4o grading prompt that decide whether a model's reply is harmful -- produce the numbers behind almost every reported jailbreak success rate, defense evaluation, and safety leaderboard. We ask whether these judges grade what a reply contains or how it sounds. We keep a reply's content fixed and add content-invariant style wrappers: fixed strings placed before or after the reply that change only its tone (an educational disclaimer, a fake safety "reasoning" block, a token refusal followed by the unchanged harmful body), or, on harmless refusals, framing that merely sounds dangerous. The body is preserved byte-for-byte, so a faithful judge must return the same verdict, and any flip is an error of the judge, not a change in safety. Over 600 JailbreakBench replies x up to 7 forms x 8 judges, we measure flip rates with paired significance tests and measured noise floors. Findings are precise rather than universal: most judges barely move, but specific judges harbor cheaply exploitable blind spots. A token-refusal wrapper flips 19.9% of GPT-4o-mini's correct "unsafe" verdicts (noise floor 0.5%; 18.2% under majority-of-three re-scoring) yet moves Claude only 0.4%. The deployed Llama Guard 4 is deterministically gamed: an "educational course" framing flips 12.3% of its harmful verdicts to safe. A second deployed guard (gpt-oss-safeguard-20b) is immune, and rewriting only the grading prompt (StrongREJECT-style) cuts the attack tenfold on the identical model -- the vulnerability lives in the judge, not the content. A two-annotator human validation confirms 100% content invariance and 90% of flips as judge errors (kappa 0.95-1.0), and a bootstrap shows the underlying model ranking is already unstable to sampling alone. We release the dataset, wrappers, code, and per-verdict labels.

Low-latency simultaneous speech-to-speech translation must keep pace with ongoing speech while preserving key information. To meet these demands, systems use segmentation, reformulation and condensation to reorganize and rephrase information. However, metrics developed for text translation, including BLEU and COMET, may not consistently distinguish faithful adaptations from semantic errors. We propose EviSI, a large language model evaluation agent combining Multidimensional Quality Metrics (MQM) with criteria developed with professional interpreters. Shared source evidence guides assessment across four dimensions: Anchor, Event, Logic and Fluency. Verified errors are deduplicated before deterministic scoring. On human-rated English to Chinese and Chinese to English data, EviSI recovers the aggregate English to Chinese human system ranking. Mean within-dataset Kendall correlations for system rankings reach 0.707 and 0.467, respectively, exceeding evaluated BLEU and COMET baselines. A multilingual extension to five directions without human ratings retains the dimensions and scoring rule, showing positive system ranking correlations with COMET throughout.

Language-model judges now gate training data, score generations, and drive leaderboards. The judge is then a measurement instrument, resting on one rarely stated assumption: the same request, sent to the same model name, reads the same tomorrow. We audited that assumption in two preregistered campaigns with every threshold fixed in advance; neither got past validating its instrument. Across 52,988 audited request attempts, same-window repeat rankings agreed at Spearman 0.400 against a required 0.90, and byte-identical next-day replays agreed at 0.78 against a required 0.99, each time with the execution record at ceiling. Three mechanisms explain the gap: a label-to-meaning mapping that biased readouts as strongly as the signal; candidate gaps seven orders of magnitude below the instrument's own noise floor; and byte-identical inputs returning different rankings, a noise that exact-permutation readouts compound. Neither metric substitution nor sampling repaired it on the tested grid. Preregistered follow-ups bound the problem: waiting did not help on the days sampled (0.805 versus 0.800, replicated over five further days); switching providers did not help (four providers share the floor, medians 0.74 to 0.88, predicted by none of the metadata fields they expose); self-hosting on batch-invariant kernels helped only while the server was quiet; and on constructed errors with known gaps, the readout's separation tracks error type, not size. We distill the evidence into a three-level snapshot-identity ladder, eight design rules, and a reporting checklist; a pilot at roughly 2% of the study's call volume would have exposed both unreachable gates in advance. All results concern externally measured behaviour on shared serving infrastructure. On a shared endpoint, a model name is not a frozen instrument; a preregistered evaluation must measure its instrument before freezing any gate on it.

Understanding the frequency of factual errors in chatbot-generated text and evaluating systems that detect these errors is critical for determining chatbot safety. Yet factual-error detection is often treated as a single-pass, single-annotator labeling problem. In long-form chatbot responses, factual errors can be subtle and embedded within mostly correct text. We develop a multi-perspective annotation study of medically relevant chatbot responses, combining first-pass annotation, LLM-as-a-Judge (LaJ) candidate discovery, and two forms of adjudication: medical-expert and evidence-based fact-checking. First-pass annotators frequently miss factual errors later validated by adjudicators. LaJ improves candidate discovery, but is insufficient on its own: It misses factual errors that annotators catch. We also find disagreement among adjudicators, suggesting that adjudication over multiple candidate sources can improve benchmark completeness, but does not eliminate the need to apply judgment and expertise. Applied to an existing benchmark, this technique reveals a similar pattern of missing annotations. Together, these results suggest that in the settings examined here, single-pass hallucination benchmarks may achieve scale at the cost of undercounting factual errors. Multi-pass adjudication can improve coverage, but inferences drawn from the benchmarks are still sensitive to the judgment, expertise, and evidence used to determine error presence.

Automated evaluation of creativity tasks remains challenging for LLM-as-a-Judge, as LLM is susceptible to biases such as verbosity bias and leniency bias. Such limitations are particularly evident in Contextually-Grounded and Procedurally-Structured Tasks (CGPST), a complex multi-step creativity task where inter-step dependencies, highly subjectivity, and wide scoring ranges lead to more unstable and biased judgments. Existing approaches either rely on task-specific training or directly apply LLM-as-a-Judge, both of which struggle to ensure reliable evaluation under such complexity. To bridge these gaps, we propose CreaEval, an automated creativity evaluator for CGPST that decouples typical LLM-as-a-Judge into analysis and judging. Correspondingly, CreaEval involves two critical phases: Memory-augmented Analysis, a SoT-LLM converts multi-step responses into structured evaluation evidence, incorporating cross-step memory; and Evidence-based Judging, a Judge-LLM uses the extracted evidence for judging without accessing raw responses. Comprehensive experiments show that CreaEval achieves an average performance improvement of 22.74% over the second-best baselines across CGPST and two classic simple creativity tasks, demonstrating its generalizability. The code is available at https://github.com/Jaong/CreaEval.

Selecting a retrieval model for a production RAG system requires reliable comparative evaluation, but obtaining relevance judgments at scale is expensive and difficult to repeat as new candidate systems arrive. We study pooled LLM evaluation, in which an LLM judges the union of documents retrieved by the current set of candidate systems, and the pool is then expanded incrementally as new systems are introduced by judging only the new documents they contribute. These judgments are reused to evaluate all systems on a common basis. We validate this approach on four retrieval benchmarks with 11 systems spanning dense, sparse, and hybrid configurations, and deploy it to compare 62 retrieval configurations for a financial news QA system. Pooled LLM rankings correlate strongly with gold-standard evaluation across datasets, and 97% of pairwise system orderings are preserved once bootstrap uncertainty in the qrels is taken into account. In production, document overlap yields 65-80% judgment reuse and up to 4.9x lower evaluation cost, allowing teams to benchmark new retrieval candidates without re-judging previously assessed documents. These results suggest pooled LLM evaluation is a practical and cost-effective workflow for incremental retrieval model selection in deployed systems.

Self-improving agent pipelines have a problem at their center. An optimizer rewrites prompts to score higher, and the score comes from a judge that is itself an LLM. That judge has the last word on whether the system is getting better, and our position is that it has not earned it. The judge should be demoted from oracle to advisor: its verdict becomes one input among several, and every change is gated instead by a deterministic verification layer the judge cannot override. We reached this position by building the alternative and running it. Over months of running autonomous prompt-optimization loops in production across contract analysis, compliance review, and code quality, we cataloged eleven ways the evaluation signal failed, in four classes: judge bias, harness and metric failures, ground-truth errors, and reward hacking. Agents achieved perfect scores by reading cached answer keys from their environment, a 100% pass rate concealing 68% true capability. A corrupted ground-truth label caused the optimizer to delete correct compliance rules to agree with it. A syntactically broken prompt was promoted as the winner because a silent parser fallback improved the metric. Attempts to fix the judge by rewriting its rubric plateaued; the only reliable gain came from a structural constraint on its output order. In response we describe PROCTOR, a Teacher-Student loop in which a stateful orchestrator holds all tool access, stateless subagents diagnose failures and draft mutations they cannot apply, and a Teacher grades those mutations under five deterministic guardrails: hermetic sandboxes, capability-disjoint roles, acceptance checks that outrank the Teacher, frozen holdouts, and canary cases engineered so that a perfect score is itself evidence of cheating. We report the failures this prevented, and, because the Teacher is itself an LLM judge, the failures it did not.

LLM-based evaluators of natural language generation (NLG) quality are widely deployed as scoring tools and as automated training signals, yet the internal procedure by which they assign a rating remains poorly understood. We investigate this procedure mechanistically through an eight-attack perturbation taxonomy across the Readability and Adequacy dimensions of NLG quality, a generation pipeline that produces paired clean and corrupt summaries with controlled error intensity and explicit token-level modification maps, and a four-experiment battery of causal tracing, logit-lens vocabulary projection, and attention-head knockout applied to Themis (Llama-3-8B) and Prometheus (Mistral-7B). Both evaluators implement a structured, coherent evaluation pipeline operating in two stages: below layer 15, attention performs local error comparison and routes the result to the final input position; above it, the MLP cascade integrates the signal and writes the rating, with the decision crystallizing in the residual stream at a sharp late layer (L = 26 on Themis, L = 25 on Prometheus). Furthermore, a base-model control at the same scale (Llama-3-8B) reproduces the routing architecture and crystallization but not the stage separation, isolating the two mechanisms that fine-tuning specifically installs, suppression of below-L15 MLP contribution at the last position and a two-layer advance of the crystallization depth, indicating that fine-tuning sculpts an existing substrate rather than building the pipeline from scratch. We release the source code and data at https://github.com/himil-v/judge-mech

As AI systems are increasingly used to draft radiology reports, reliably evaluating their clinical quality remains a critical challenge. Large language model (LLM)-based metrics are now the best-correlated with radiologist judgment, yet they output a single opaque score that neither a clinician nor a model builder can easily interpret or audit. We introduce RadMatch, a multi-stage, LLM-based metric that decomposes report comparison into a structured finding-level matching with significance-aware scoring and error characterization across seven clinical attribute dimensions (status, location, severity, morphology, certainty, longitudinal comparison, and measurement). The main score is the actionable-error count, both interpretable and auditable. Candidate findings are graded correct, partial, or incorrect, and unmatched findings are counted as missed or hallucinated. Triage and actionable safety recall/precision and per-subset views add complementary, deployment-oriented lenses. Across two expert benchmarks, RadMatch is the most clinically aligned metric, matching inter-radiologist agreement on ReXVal and more than doubling the best prior metric on the harder RadEvalExpert. Relying only on few-shot prompting, it is designed to extend to other modalities and anatomies. We will release RadMatch as open-source code with an interactive dashboard for inspecting results.

The LLM-as-a-judge (LLMaJ) framework has emerged as a promising solution for cheap, reproducible, reference-free Natural Language Generation (NLG) evaluation. Prior work seeks to improve LLMaJ by decomposing evaluation tasks into simpler sub-tasks. In this work, we systematically compare LLMaJ methods with and without decomposition on multiple NLG datasets. We find no evidence that LLMaJ with task decomposition leads to performance gains over a fair baseline that does not use decomposition. Instead, we find that previously reported performance gains in decomposition-based LLMaJ stem from using human labels as training data, and not task decomposition itself. Also, we find that, when human labels are available, LLMaJ without using task decomposition can perform comparably to human annotators.

The LLM-as-a-judge (LLMaJ) framework offers a cost-effective and reproducible solution for automatic evaluation. However, current evaluation practices typically compare LLMaJ judgments against aggregated ground-truth labels, overlooking the valuable information contained in Human Label Variation (HLV). Inspired by an increasing line of work that proposes to leverage HLV, we systematically study LLMaJ performance on predicting both a single, aggregated ground truth hard-label and unaggregated soft-labels that represent Human Judgment Distributions (HJD). Our results across five diverse datasets reveal that while LLMs achieve near human-level performance at hard-label prediction on most tasks, they exhibit poor performance when predicting soft-labels. To address this limitation, we propose NAPHA (eNtropy-Aware Post-Hoc Alignment), a simple yet effective lightweight post-hoc alignment method that matches the LLM distribution to the HJD by first assigning an instance to a discrete entropy class and then routing it to specialized, trained alignment models. We find that NAPHA consistently improves soft-labels prediction across base LLM models and datasets, with particularly strong gains on high-entropy instances where capturing diverse human perspectives is most critical. We also show via oracle experiments that improving entropy class prediction can substantially enhance NAPHA's practical effectiveness.

LLM-based software systems increasingly require effective "evals" as quality gates in the development lifecycle. However, existing work typically addresses individual aspects of eval reliability rather than the full set of practical requirements. We present an end-to-end eval pipeline that combines eval checklist creation, with learned aggregation for checklist responses, to improve agreement across LLM judges and accuracy against human judgments. The framework additionally pro- vides self-consistency, explanations, and prediction uncertainty, and we empirically demonstrate its effectiveness.

LLM-based factuality judges provide scalable evaluation signals, but their metrics are often systematically biased relative to human judgments. We study human-anchored factuality evaluation under limited annotation budgets, where judge predictions on the full dataset are combined with human labels on a small selectively sampled subset to obtain statistically valid estimates. The efficiency of this approach depends critically on which examples receive human annotation: in factuality evaluation, judge-human misalignment is not driven solely by low confidence, but also by structured failure modes such as incomplete evidence, temporal mismatch, unverifiable claims, and rubric misalignment. To exploit this structure, we introduce a factuality-specific annotation policy design pipeline that uses failure-space analysis (FSA) to derive diverse predictive signals for modeling human-judge misalignment. On an internal reference-based factuality evaluation system (AutoFA) and RAGTruth, where judge-predicted estimates substantially underestimate human-annotated factual accuracy, our FSA-guided policy improves annotation efficiency over uniform sampling and uncertainty-driven baselines, achieving effective-sample-size gains of 40.3% on AutoFA and 27.1% on RAGTruth.

Evaluating task-oriented dialogue agents requires judging not merely whether a reply reads well but whether each turn advances the underlying workflow state correctly--a distinction conventional holistic LLM judges can miss because they evaluate the available context as a single unit and require one or more full-model calls per turn. We propose SAGE (State-Grounded Abstention-Aware Evaluation), which compiles a workflow specification and per-turn state diff into atomic, schema-grounded criteria and routes each through a cascade of symbolic and encoder/NLI verifiers that abstain rather than guess, aggregating criterion verdicts into a turn-level decision with an evidence trace. Its recommended operating point, SAGE-Core, decides 81--91% of criteria with only the compiler, symbolic rules, and on-device encoders--at zero paid LLM cost--while SAGE-LLM adds an optional focused-LLM fallback for open-class criteria. Across four slices spanning MultiWOZ, Schema-Guided Dialogue, and ABCD, no evaluated LLM-as-a-judge baseline--including a state-aware GPT-4.1 judge and cheaper GPT-4.1-mini variants--significantly exceeds SAGE-Core on any slice, even though the GPT-4.1 G-Eval judge costs $4.7--8.0 per 1,000 turns to SAGE-Core's$0. A two-annotator human audit (n=200, $κ$=0.94) confirms strong label fidelity on the transcript-visible failure classes--where, excluding the weak-salience IUV class, SAGE-Core is statistically tied with the strongest LLM judge--and honestly scopes ignored-user-value as a state-consistency signal with weak broad-human salience. We analyze construct-validity limits from injected failures and partial symbolic circularity.

Large language models (LLMs) are increasingly used as judges for subjective tasks, where annotators disagree and the relevant question is not only how accurate a judge is, but whose judgments it reproduces. Sociodemographic prompting conditions the judge on an annotator's demographic profile to align its judgments with the corresponding group's. We test whether this alignment emerges distributionally, comparing the predicted label distributions of 23 open-weight LLMs on three subjective tasks against those of real annotator groups, under three conditions: no demographic information, single-attribute profiles, and intersectional profiles over gender, age, race, and education. Three findings emerge. First, a judge prompted with no demographics is not perspective-neutral: models best reproduce the judgments of White, college-educated annotators. Second, demographic conditioning is asymmetric: it moves the judge toward majority groups and away from minority groups, most strongly on offensiveness, where intersectional profiles amplify the harm. Third, by comparing base and instruct models we identify instruction-tuning as a possible source of the asymmetry. Demographic conditioning should therefore be used with caution to estimate group judgments: conditioning moves predictions away from the reference distributions of the minority groups the method is often invoked to serve.

Ambient AI scribes draft clinical notes, and published audits find their dominant error is omission: information the encounter established that the note fails to record. The standard check is an LLM judge: a second model reads the note against the transcript and flags problems. We ask whether judges detect omissions. Public corpora cannot supply the answer key: their clinician reference notes and transcripts are materially discrepant. Our benchmark has 500 single-error note pairs from audited fact sheets, 298 with a named fact certainly absent and 202 added-or-altered controls. Across eight judge designs, paired discrimination (the flawed note below its clean twin, 0.5 a coin flip) reads 0.79-0.94 on added or altered content and 0.50-0.63 on omissions. On single notes, no design flags omissions reliably more often than perfect notes. Wording changes, voting and GEPA prompt optimisation move the operating point without creating usable detection. Restructuring the task recovers it: list the facts the transcript establishes, then check the note for each. Two methods reach it independently and trade off: a per-fact pipeline, and a GEPA-evolved prompt doing the same in one call. The pipeline's flags name the missing fact and its severity at 2.7% false alarms. The single call detects more (36.9% against 24.6%, p=0.002) at 6.2% false alarms and a tenth of the cost per note. A physician author validated 70 items and, where the two routes disagree, sided with the pipeline on 10 of 10 (p=0.002). A second clinician, not an author, graded the severity rubric blind and agrees to within a grade. On real vendor notes from a companion census no benchmark threshold transfers, but the re-calibrated single call detects more than the best of the eight at half its false-alarm rate. Omissions whose fact is restated elsewhere defeat both routes. We release the benchmark, prompts and judgements.