Llm-As-A-Judge

Large language models (LLMs) are increasingly considered for safety-critical engineering, yet their reliability in regulated functional-safety workflows remains underexplored. We introduce SAFARI (Safety-Aware Functional Automotive Risk Inference), the first industrial benchmark for LLM-assisted automotive Hazard Analysis and Risk Assessment (HARA) under ISO 26262. It contains 3,000 de-identified industrial HARA cases and evaluates two coupled tasks: open-ended hazard analysis and standards-grounded risk assessment. To evaluate open-ended HARA artifacts, we propose the first reference-anchored LLM-as-a-judge protocol with high expert correlation. Experiments with nine frontier LLMs show that models often produce plausible hazard narratives but remain weak at ISO 26262 risk classification, with the best ASIL macro-F1 reaching only 0.261. Chain-of-Thought prompting provides limited benefit and often degrades categorical risk assessment. Error analysis further localizes major failures to scenario-critical context omissions during hazard generation and to controllability misjudgments during risk assessment, indicating where expert oversight should be concentrated. The dataset can be obtained from https://github.com/xixi47520-hash/HARA.

Who the judge is can affect an LLM-as-judge result, but measuring that effect without confusing it with candidate quality is difficult. We study four open-weight families (Llama 3.1, Qwen 2.5, Gemma 2, and Yi 1.5) in a fully crossed pairwise design with 9,312 judgments. A common per-family statistic is strongly confounded with candidate quality and correlates with Bradley-Terry ability at r = 0.95. We derive a corrected estimator that holds the candidate family fixed and compares judges. All four families then show a positive same-family lift (3.4-8.4 percentage points), with global FPS 0.067 (95% CI [0.053, 0.084], permutation p = 0.0002). The effect remains under panel-based quality controls, an independent human-consensus anchor, and a float16 judging replication. Judge-side likelihood is closely related to the effect: adding likelihood advantage reduces the controlled coefficient by 61%, which we treat as descriptive attenuation rather than causal mediation. Position is a separate failure mode. Across the panel, 55.4% of AB/BA pairs reverse, and reversal above 50% is incompatible with a simple independent content-noise model. Relative to a family-balanced reference, panel composition changes 18.5% of pairwise outcomes. A complete reproducibility archive has been prepared for public release.

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.

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.

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.

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.

In high-stakes domains such as legal practice, a language-model answer is only useful to the extent that a reader can verify each claim against the source the system cites. Current grounded-generation pipelines score the answer as a whole, so a correct conclusion can rest on fabricated or loosely matched citations and still score well. Closing this gap requires both a system built for per-claim verification and an evaluation that measures it. We introduce GANDR (Grounded ANswer DRafter), a two-agent system in which a Drafter writes an answer in a structured legal-reasoning format and a separate Critic, with the same view as a human verifier, audits each claim against its cited source and emits a per-claim audit trace on every round. We pair it with a strict correctness criterion requiring every citation to resolve to a passage the retriever returned. On a 185-item legal benchmark where all six systems share one backbone, one retrieval surface, and one citation instruction, GANDR ranks first on every primary metric, reaching 70.8% strict accuracy and leading the strongest baseline by 11.3 points (p<0.01). Reverting the protocol-anchored commit rule lowers strict accuracy by 22.7 points, and the strict lead stays positive on three further backbones, at +3.2 to +6.5 points. This lead traces to the Drafter configuration and the protocol-anchored commit, not to rewriting. Against two law-trained annotators the audit flags under-supported claims at F1 0.84 as a binary detector, while its four-way verdict labels agree only weakly and are advisory. Code is available upon request.

LLMs have been used to simulate human decision-making in professional settings, yet their behaviors in common-law jury trials remain unexplored. We study when and how a defendant's courtroom statement affects LLM-simulated jurors, focusing on persuasion, ideological bias, and background-based affinity. To support the analysis, we introduce JuryBench, a benchmark containing controversial criminal cases in U.S. criminal law. In each case, a defendant can claim various plausible justifications to support acquittal or reduced liability. We fix the base case and design defendants of different backgrounds, who give courtroom statements with varying emotional appeal or rebuttal. Jurors with diverse ideological profiles across the spectrum are simulated. We examine 20 frontier LLMs, resulting in a total of 432K decisions and rationales, and quantify changes in verdict severity. Our findings show that LLM-jury simulation echoes many human-jury findings. First, emotional persuasion can be detrimental, since jurors may perceive it as evidence of guilt or inconsistency. Next, we show that background fit between jurors and defendants is a stronger and significant factor than other isolated factors, and that jurors are in general harsher toward opposite-background defendants and lenient toward same-background ones. Finally, we find that juror ideology also strongly shapes severity judgments. These findings highlight both the promise and risks of using LLMs to model jury reasoning and call for careful evaluation. The data and code are available at https://github.com/choyingw/JuryBench

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.

Oncology care operates at constant pressure of absorbing rapidly evolving evidence base in biomedicine. The American Society of Clinical Oncology (ASCO) addresses this through living guidelines, but the format introduces a new burden: any recommendation can change at any point, across multiple versioned documents. We present SENTRYLINE, a living guideline-aware clinical question answering system. SENTRYLINE retrieves guideline passages through a vectorless hierarchical RAG pipeline and returns a role-specific answer with inline citations, factual and temporal verification reports, and drift detection notes that surface when a guideline has been updated. We construct ASCOBENCH, a benchmark of 405 three-turn conversations across four question categories with gold answers from expert annotators(clinicians), and use test set to evaluate SENTRYLINE against five baselines under an LLM-as-judge framework. Experiments across three generation backbones show consistent improvements over four retrieval baselines and ASCO's guideline assistant, with particularly strong gains on Reasoning and Role-Specific questions where multi-hop synthesis and register adaptation are required

Large language models are increasingly used as sources of advice and information, including in high-stakes settings, yet little is known about how they respond to user disagreement. We study how a model manages its epistemic authority, referring here to its claim to knowledge, competence, or the right to advise, once a user challenges its answer. Building on Conversation Analysis, we introduce a taxonomy of six challenge types and a four-layer framework for analysing each response: whether the original claim is maintained or changed, where authority is located, how the disagreement is socially managed, and what kind of evidential support is offered. We construct a new dataset of 2,310 controlled challenge scenarios and 32,340 corresponding responses from 14 models, and analyse them using our framework with an LLM-as-judge pipeline, providing a vocabulary which future evaluation and benchmark design can build on. We find that models show conflicting behaviour: they validate users in 85% of responses but maintain their original claim in 65%. They explicitly apologise in 33% of responses, yet 59% of those apologies accompany maintenance of the original claim. They transfer authority most often in advice tasks, doing so in 28% of responses and reaching 57% in health advice and 49% in legal advice, compared with 6% in fact and 3% in explanation tasks. Abandonment of the original claim ranges from 0.8% for GPT-5.2 to 40% for DeepSeek 7B, while complete replacement of the original claim is rare overall at 1.5%.

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.

LLM-empowered paper-code discrepancy detection has received growing concern since the scaling of research submissions exceeds the manual review capability. However, the limited context capacity and one-sided discrepancy detection of existing single-agent LLM paradigms lead to an inferior recall performance in detecting discrepancies. In this paper, we propose Dude, the first Dual-Detection Multi-Agent System for paper-code discrepancy detection. We discover that the granularity asymmetry of the paper-language and code-language introduces over-interpretation and over-reporting challenges in a multi-agent system design for discrepancy detection, resulting in increasing false positives. To address this, we propose a granularity-aligned negotiation and a two-stage salience-filtering mechanism in Dude, which effectively prevents agents from falsely reporting discrepancies. Experimental results in real-world paper-code discrepancy datasets showcase Dude's significant recall and precision improvement by up to 22.8%, increasing F1 score by up to 18.7% compared to baseline methods.

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

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.

Commercial short-drama production follows a multi-stage chain: script, storyboard, keyframe imagery, shot-level video, and the finished short drama. Most existing benchmarks evaluate solely the video-generation stage using pre-authored inputs instead of real upstream pipeline outputs. This leaves two critical questions unanswerable: whether each stage adheres to the original script intent (rather than only its immediate input prompt), and whether disparate shots remain coherent after assembly into multi-episode releases. We present DramaChain Bench, the first short-drama benchmark that evaluates every stage of the complete production chain. It is built upon three in-house systems sharing one dimension system, DramaChain Dimensions: five evaluation axes instantiated at every stage, resolving into 63 leaf dimensions. DramaChain Agent is calibrated against commercial short-drama platforms in both workflow and finished short-drama quality, enabling stage-wise fair comparison across models. DramaChain Labeling System has each of the 5,785 items scored independently by three professional annotators, with all defects spatio-temporally localised and selected from a predefined defect list. This process produces 17,488 valid scores and 255,925 traceable attribution records. The human annotations confirm that upstream defects cascade across the pipeline, demonstrating that final episode quality is not governed by video generation alone. DramaChain Agentic Judge then scores every leaf dimension automatically, gathering evidence over multiple agentic rounds before judging against a per-item checklist; it reproduces the model ranking at a mean PLCC of 0.918, enough to admit new models at no annotation cost.

Evaluating conversational voice agents at scale re- quires reliable assessment methods that capture both observ- able interaction quality and the contextual judgment typically provided by human evaluators. We investigate LLM-as-a-Judge evaluation by comparing human judgments with GPT-4.1 and GPT-5 on telecom and retail voice-agent conversations, across conversational quality and safety dimensions. The same interac- tions are scored under three evaluation configurations, p0, p1, and p2, to test whether automated judgments are sensitive to the evaluation setup and whether observed patterns generalize across configurations and judge models. Beyond aggregate agreement, we examine metric-level correlations, evaluator consistency, and systematic human-LLM disagreement to identify which conver- sational attributes can be judged reliably by automation and which remain sensitive to interpretation and context. Effective voice-agent evaluation is also shaped by pipeline-level factors such as speech generation, streaming, and error propagation across ASR, reasoning, and tool-calling stages, motivating our focus on comparing how human and LLM judges score the same interactions end to end. Our results show that LLM- based evaluation can serve as an effective component of large- scale voice-agent assessment, but that its reliability is metric- and configuration-dependent rather than uniform. This pro- vides an empirical framework for identifying which metrics suit automated evaluation and supports hybrid pipelines in which LLM judges handle scalable assessment while human evaluators remain engaged for metrics that demand contextual interpretation and higher-confidence judgment.

Benchmarks for scientific artificial intelligence are mostly written to be scored: multiple-choice questions, curated agent tasks with reference solutions, or simulators with a known generative structure. Real scientific requests arrive differently. They are underspecified, they carry attachments, and they lack ground truth. We report K-Bench 01, an evaluation built from first-turn requests sampled from live user traffic on K-Dense Web and run end to end by nine frontier models in identical sandboxes, yielding 1,602 completed agent runs. Three blinded language-model judges scored every run against an eight-dimension rubric. On a rubric whose 8-anchor is defined as work a domain scientist would accept with minor edits, no model clears the line under all three judges. gpt-5.6-sol has the highest pooled mean, 8.04, but its 95% interval [7.80, 8.23] spans the threshold, and two of the three judges rank claude-opus-5 first instead. We therefore report the ordering of systems as the reproducible quantity, the absolute level as an attribute of the instrument, and the top of the table as unresolved. Across all 39,934 scored judgments -- the eight dimension scores plus a holistic overall for each assessment, excluding not-applicable cells -- 47.6% fall below the 8-point threshold. Difficulty is not uniform across the rubric: scientific accuracy averages 6.22 against 7.33 for communication, on identical denominators and in the same direction within every one of the nine models. The single leading failure tag is overclaiming, on 31.4% of assessments. We argue that the informative quantity for scientific agents is not a leaderboard position but the joint distribution of what was delivered, what was claimed, and what artifacts were produced.

Agentic large language model (LLM) systems are reshaping scientific workflows in chemistry and drug discovery, but evaluating their open-ended, tool-augmented outputs remains a fundamental bottleneck. The LLM-as-a-Judge paradigm has emerged as a scalable alternative, but existing drug discovery benchmarks deploy LLM judges without validating their alignment with human experts. In this work, we present an LLM-as-a-Judge evaluation framework for ChatInvent, an agentic drug discovery assistant deployed at AstraZeneca, with five contributions. First, we define four output-quality evaluation dimensions---Completeness, Relevancy, Structural Clarity, and Scope Adherence---alongside deterministic Tool Call Correctness checks. Second, we validate the judge through a human alignment study with five expert annotators, comparing Gemini 3.1 Pro, Claude Opus 4.7, GPT-5, and Llama 3.1 70B as candidate judges. Third, we optimize the best-performing judge using few-shot demonstrations of human-annotated examples, improving alignment with the human majority vote from 0.80 to 0.86. Fourth, applying the optimized judge to 70 held-out questions, we surface concrete limitations and find no strong evidence that informal phrasing degrades output quality; it may, however, still be helpful to have the LLM rewrite the original question before querying the agent. Finally, we extend the framework to 38 adversarial questions that are ambiguous, invalid, out-of-scope or ethically sensitive, and show that the agent's refusal behavior is guided by the stated intent of a request. Our framework provides a reusable template for human-aligned evaluation of agentic systems in scientific domains.

LLM-as-a-Judge, which leverages a large language model to evaluate natural language generated by another AI application or model, has become a standard, scalable approach for accelerating and extending costly human evaluation. Yet most work treats a judge as a static artifact, evaluating it once at construction or against a fixed benchmark. We argue instead that an LLM judge operating in a deployed system is better understood as having a lifecycle. It must be built, trained, deployed, and continuously maintained as the surrounding data evolves, and each phase poses distinct technical and operational challenges. We present such a lifecycle for the LLM judges that evaluate recommendation explanations at Netflix. Everything we report comes out of a series of controlled online member-facing experiments, in which our pipeline generated and the judges assessed hundreds of thousands of distinct show-level explanations per week across a changing catalog. Our framework has four phases. (I) Birth defines the evaluation criteria and builds curated benchmark datasets with human labels and rationales. (II) Training refines the judges' rubrics via Reasoning-Aligned Rubric Tuning (RART), which uses a meta-judge over reasoning output as the learning signal. (III) Deployment puts one judge in two online roles, quality gating and reflective generation. (IV) Monitoring runs a continuous Human-in-the-Loop (HITL) alignment process that detects drift and triggers re-tuning behind a human review gate. We report results from a five-week online A/B test over tens of millions of members on the Netflix mobile app, in which judge-aligned explanations shifted member viewing toward novel content (previously unwatched) and increased successful browse-to-play sessions relative to a no-explanation control, with no quality-related escalations.

Predictive-distribution entropy makes a strong selection rule in retrieval-augmented question answering: across five QA benchmarks, keeping the candidate answer that a frozen respondent LLM produces with the lowest answer-token entropy lifts mean answer $F_1$ from 0.4769 to 0.5148 over the retriever's top-ranked passage, with no gold answers. Yet this lowest-entropy rule, which prior entropy-based selectors adopt, fails in a specific and consequential way: a misleading passage makes the respondent confidently wrong, driving its entropy down precisely where the signal looks most trustworthy. We show that the failure comes from the passage the respondent reads -- and the context that passage is read in is an input we can intervene on. We introduce LODESTAR, to our knowledge the first method to score a text intervention by the uncertainty it induces in a third-party frozen respondent, compared across one question's candidates. LODESTAR uses reinforcement learning to train, once and offline, a polarizer -- a short fixed natural-language string inserted into the respondent's prompt and never into its weights; its training labels are built offline from gold answers and two LLM judges, and inference reads neither. Evaluating every competing selector under the same frozen respondent and the same candidate pools on 5,008 questions, LODESTAR attains the highest mean $F_1$ of any inference-ready selector (0.5148 to 0.5339), the highest exact match (0.4136), and the highest GPT-4o judge score of the frozen-respondent configurations judged (0.6435); its three-seed mean wins all 70 method-by-dataset $F_1$ cells against fourteen published configurations while remaining paired-significant against every one. The gain holds both in-domain and out-of-domain, and ablating the polarizer shows it is what makes the respondent read a misleading passage less often (26.0% against 30.3%).

Mobile agent benchmarks increasingly rely on LLM-based judges to evaluate task completion, yet the reliability of these judges on mobile agent trajectories remains largely unexamined. We introduce MobileJudgeBench, a benchmark for systematically evaluating LLM-as-judge methods on mobile agent trajectories. Our benchmark comprises 931 human-annotated trajectories spanning 6 mobile agent benchmarks, 4 agent models, and 68 apps. We evaluate 6 judge methods (five adapted from SPA-Bench, A3 with two modes, AndroidArena, and AgentRewardBench, plus a simple baseline we design) across multiple LLM backends. Our experiments reveal three key findings. First, a simple baseline judge with sampled screenshots is competitive with, and often exceeds, purpose-built methods, indicating that more elaborate judge pipelines do not consistently improve judge quality; among competitive methods, the LLM backbone is the primary driver. Second, benchmark quality metrics reliably predict real-world judge utility: they correlate with both agent ranking fidelity for evaluation and downstream performance when judges serve as reward signals for on-policy reinforcement learning. Third, failure analysis across two LLM backends uncovers qualitatively opposite failure profiles, one conservative and the other permissive, linked to the backbone's precision-recall characteristics.

AI video generation has advanced rapidly and entered widespread commercial use. As a result, quality differences among videos produced by state-of-the-art AI video generation models~(AIVGMs) have become increasingly difficult to discern using conventional evaluation criteria, such as visual fidelity and semantic instruction following. Meanwhile, human evaluation now requires more expertise and sustained attention, substantially increasing annotation costs. This calls for automated evaluation that can reliably distinguish fine-grained differences among advanced AIVGMs with minimal human intervention. To address this challenge, we present RAVEN-Eval, a rubric-guided automated evaluation framework for AIVGMs, built primarily on the LMM-as-a-judge paradigm. Through an automatic task curation and quality-filtering pipeline, RAVEN-Eval curates 150 text-to-video~(T2V) tasks and 100 image-to-video~(I2V) tasks, and systematically collects more than 4,500 AIGVs. At its core, RAVEN-Eval adopts rubric-guided automated LMM preference judgement, in which LMM judges conduct pairwise comparisons according to task-specific rubrics. It further introduces an anchor-based model insertion approach to reduce the evaluation cost of incorporating new models. Finally, we evaluate 20 high-performance AIVGMs, as well as the judging capabilities of 13 LMM judges, and establish the RAVEN-Eval Leaderboards. Overall, RAVEN-Eval paves a scalable path for automatic and trustworthy evaluation of rapidly evolving AIVGMs.

Large language models are increasingly deployed as autonomous agents that interact with the web through browsers. While recent progress has been driven by benchmarks that evaluate end-to-end task success, these evaluations largely overlook two fundamental sources of difficulty in real web browsing: complex actions over rich user interfaces and visual perception of dynamically rendered content, especially in workflows that span multiple websites. We introduce CAP, a scalable benchmark for evaluating browser agents on cross-site, human-like web tasks that require non-trivial UI interactions and visual understanding. Specifically, we adopt a decomposition-and-recomposition pipeline that first abstracts each website into a structured site card capturing user-facing functions, complex execution operations, and perceptual requirements, and then recomposes these components into realistic cross-site workflows. Each task is therefore grounded in multiple specific operations on each website, enabling fine-grained diagnosis. Built on this framework, we construct 420 tasks across 108 real-world websites and 24 domains under careful quality control. Experiments on state-of-the-art browser agents using our verifiable agent-as-a-judge evaluation framework show low success rates and reveal that perception-heavy interactions remain a major bottleneck, exposing substantial gaps between current agents and real-world web browsing demands.

An LLM judge deployed inside a reasoning pipeline does not merely measure quality, it decides which answer ships. We show that the cost of that decision depends less on judge accuracy than on the decision rule the judge is embedded in. On frozen candidate pools from four GRPO policies, an unconstrained scalar DeepSeek-R1-7B judge buys almost nothing over answer-level majority vote (+1.0 pp on 500 GSM8K questions, +0.34 EM on 300 HotpotQA questions), and on a frozen-rule 30-question confirmation split it is 10 points worse than majority, a judge that destroys accuracy while scoring candidates confidently. We then subordinate the same judge to Evidence-Locked Derive-Gate-Repair (EL-DGR), a task-adaptive non-compensatory rule under which a judge preference may override evidence-supported consensus only with an extractive evidence certificate, and a repair only when neither alternative is certified and the repair is. With no change to the judge, the candidates, or the budget, EL-DGR reaches 58.2% on GSM8K (vs. 56.8% judge, 55.8% majority, 55.4% first candidate) and 17.33 EM / 25.46 F1 on HotpotQA (vs. 15.67/23.49, 15.33/23.19, 15.33/22.97), improving on first-candidate GRPO by +2.8 pp (exact McNemar p=0.0026) and +2.00 EM (p=0.070, borderline). A decision audit shows why: EL-DGR overturns consensus on only 8 of 30 pilot questions and never converts a correct consensus into an incorrect answer. We also report what did not work: the same seven-channel decomposition used as a step-level gated training reward is null, and corrected channel-drop ablations show no channel is individually necessary (p=1.0 throughout). The practitioner-facing finding is negative about judges and positive about admissibility, bound the judge's blast radius rather than trying to make it accurate.

Audio-language models (ALMs) are increasingly used as judges for speech-to-speech systems, but a judge that receives audio may not actually use paralinguistic evidence. We introduce counterfactual audits for paralinguistic response evaluation. Each audit item holds the transcript fixed while varying affect, prosody, or the timing of an affective shift, forcing a valid judge to track the audio cue rather than lexical content or response style. We evaluate ALM judges using a native one-context judgment protocol and a contrastive recoverability control, then further decompose each item into its constituent perception and response-mapping skills. This yields useful diagnostic states that identify different sources of judge failures. Across Gemini, GPT, and open audio models, we find that contrastive success often overstates native judge reliability, and that similar aggregate accuracies can hide different failure modes. These results suggest that ALM judges should not be evaluated by accuracy alone, instead requiring thorough behavioral audits before deployment.

Large Language Models (LLMs) are often used as evaluators of text quality, known as LLM-as-a-Judge, which can outperform conventional automatic evaluation metrics that rely on reference texts. However, LLM evaluators tend to generate particular scores regardless of the context of the evaluated text, which is known as scoring bias. This study proposes a novel method to mitigate this scoring bias. An LLM is instructed to randomly generate number tokens, and the latent numerical bias of the LLM is identified by measuring the deviation of the observed distribution of numbers from the uniform distribution. A definition of a downstream task, for which an LLM evaluator is used, is added to the prompts for random number generation to measure task-specific latent number bias. In the evaluation by an LLM, the token generation probabilities for a given input are rectified considering the LLM's latent number bias. Results of the experiment on four different tasks, evaluation of LLM alignment, evaluation of summarization, Semantic Textual Similarity, and Semantic Textual Relatedness, demonstrate that our proposed method outperforms the baselines, including an LLM without debiasing and previous calibration methods. In addition, it is confirmed that scoring bias varies across LLMs, tasks, and score ranges, indicating the importance of measuring latent number bias as the case may be.

LLM agents increasingly execute long-horizon tasks through tool use and environment interaction, shifting evaluation from final-response scoring to verification of complete executions. For skill-augmented agents, verification additionally requires the procedural knowledge encoded in task-time skills, because this knowledge indicates what evidence to inspect and which failures are task-critical. However, existing judge benchmarks often expose final responses or static trajectories, and rarely combine task-time skills with directly inspectable artifacts and environments. We therefore introduce SkillTV-Bench, a 681-case benchmark of real agent trajectories from 50 tasks across eleven domains, designed to evaluate skill-aware trajectory verification for both LLM-as-a-Judge and Agent-as-a-Judge methods. Additionally, we propose SkillTV-Evolve, which externalizes verification knowledge as a reusable JudgeSkill that guides an agent judge to plan targeted inspections and issue evidence-grounded verdicts. On a disjoint development pool, an automated evolution loop further refines the JudgeSkill using misjudged cases. On SkillTV-Bench, the refined skill increases the same agent judge's accuracy by 14.8 percentage points. In offline rollout-pool selection, it increases selected-trajectory success from 22.9% with one rollout to 45.5% with ten rollouts. The code and data are available at https://github.com/HanZhi306/SkillTV-Bench

In-the-wild Bengali scene text recognition is largely unmeasured: existing resources target handwritten documents or constrained sign-board parsing, report only aggregate edit-distance metrics, and evaluate either conventional OCR or VLMs, never both on the same in-the-wild data. To address this gap, we introduce BANGLAWILD, a benchmark of 2,535 Bengali scene text images, each paired with a verbatim gold transcription, two categorical axes, four diagnostic attributes, and an orthographically standard form where the in-image text deviates from canonical spelling. We evaluate fifteen VLMs and three conventional OCR systems under three prompting strategies, fine-tune 6 open-source models with LoRA, and complement edit-distance metrics with an LLM-as-a-Judge evaluation. Our results reveal a persistent gap in which larger models within the same family do not outperform smaller ones. Our fifteen-class error taxonomy shows that visual mis-recognition accounts for ~60% of errors in the strongest systems, while conjunct-related errors contribute under 2%, challenging a long-standing assumption in Bengali OCR research; the same visual dominant profile also holds across architectures, including the one conventional baseline that reads Bengali reliably. Prompt language mainly affects cross-script drift and LoRA reduces catastrophic failures in weak models without lifting the ceiling on already competent ones. Code and data will be publicly released.

Rubric-based LLM-as-judge pipelines often assume that evaluation criteria provide independent signals. In practice, however, criteria can be behaviorally coupled: improving one criterion may systematically change scores on another, distorting aggregate scores used in model-release or product-update decisions. We introduce RADAR, a lightweight preflight diagnostic framework for estimating such coupling before large-scale evaluation. Given a rubric, RADAR generates targeted synthetic probes, scores each probe on all criteria, and produces a directional coupling matrix that shows which criteria co-score and how. We validate RADAR on three industry-relevant evaluation settings: NVIDIA HelpSteer2, SumPubMed, and the Yale-Salesforce SummEval benchmark. Using only a small number of probes per criterion, RADAR recovers human inter-criterion correlation structure (Pearson r > 0.84) and provides practitioners with concrete audit signals about redundancy, hierarchy, and aggregation sensitivity before committing to large-scale judging.

However, whether these judges truly evaluate the scientific substance of ideas or are influenced by superficial stylistic presentation remains an open question. To address this question, we propose SciStyleBench, a unified three-component benchmark for diagnosing and mitigating stylistic bias in LLM-based idea evaluation: (i) First, SciStyleStage, a three-stage evaluation environment that applies controlled stylistic perturbations to fixed scientific content across three settings no context, fixed-domain context, and open-domain retrieval context, covering 600 scientific ideas and 15 style variants, with 9,000 evaluation instances per setting; (ii) Second, SciStyleMetrics, a set of quantitative measures, including Style Bias Index (SBI), Substance Recognition Rate (SRR), and Adversarial Win Rate (AWR), to characterize how stylistic variation affects scoring stability, substance discrimination, and ranking robustness; (iii) Third, SciStyleExtractor, a plug-and-play evaluation module that separates presentation style from scientific content by predicting style type and deviation before style-conditioned evaluation, enabling us to assess whether style awareness reduces stylistic bias. Experiments on SciStyleBench show that direct LLM judges remain sensitive to writing style and struggle to distinguish scientific substance. In contrast, SciStyleExtractor reduces SBI from 0.566 to 0.501 while increasing SRR and AWR from 0.504 and 0.554 to 0.759 and 0.899, respectively. These results suggest that robust idea evaluation requires invariance to stylistic variation without sacrificing sensitivity to scientific substance. Overall, SciStyleBench provides a systematic framework for identifying, quantifying, and mitigating stylistic bias in scientific idea evaluation.

Uncommon and off-guideline cases are difficult for clinical decision support, because physicians must make a series of management decisions under diagnostic uncertainty and rarely see the full case at once. Most large language model (LLM) benchmarks for medicine score only the final diagnosis, yet much of clinical care turns on the next appropriate action: the next test to order, the imaging study to obtain, the specialist to involve, or the differential to pursue. We introduce MedUPSQA, a dataset of 21,874 mid-stream clinical decision points built from 5,535 real case reports, and MedUPS, an alignment framework that supervises models on these intermediate decisions as they unfold along a patient's trajectory. We segment free-text case presentations into chronologically ordered, accumulating clinical chunks and align models to predict the next step with reinforcement learning (GRPO), using an external LLM-as-a-Judge reward. This objective mirrors how clinicians actually meet patients, reasoning forward from accumulating evidence toward the next decision, rather than committing to a final label. Across three backbones, mid-stream alignment raises next-step accuracy from 55.2 to 66.7 for Qwen3.6-27B, from 47.2 to 57.8 for Qwen3.5-9B, and from 37.8 to 44.4 for HuatuoGPT-3-8B, with 95% CI. In several model scales we test the objective improves accuracy more than scale, with smaller models surpassing larger, frontier models we evaluate. We further train supervised fine-tuning (SFT) baselines on the mid-stream task, SFT improves all backbones above base, indicating the target framwork carries signal independently of the optimizer. We release the dataset, code, and aligned checkpoints.

Qualitative researchers increasingly encounter interaction corpora whose scale exceeds what manual coding alone can address, and large language models (LLMs) are frequently proposed as analytic assistants. The open questions are not whether LLMs can participate in qualitative analysis but to what extent, in what phases, and under what safeguards. This article provides a detailed procedural account of a multi-phase human-LLM collaborative pipeline that adapted open, axial, and selective coding to develop a hierarchical codebook from 45,000 messages exchanged between K-12 educators and a generative AI platform. Across three phases, LLMs generated candidate labels and structured annotations at scale, while human researchers retained conceptual authority over category definitions, merging decisions, and interpretive frameworks. The resulting instrument was then tested through systematic human coding, in which three trained coders with educational domain expertise applied the codebook to an independent sample of 2,560 messages, established reliability through iterative calibration using set-valued agreement measures appropriate for multi-label annotation, and extended the instrument with five codes that the LLM-assisted phases had not surfaced. The final codebook comprises 72 items within 19 categories and six domains. We reflect on the methodological decisions the pipeline required, including the choice of a conversational unit of analysis, the treatment of the LLM as a labeling instrument rather than an interpretive agent, the measurement of intercoder agreement under multi-label coding, and the conditions under which human domain expertise remained decisive. The account is offered as an auditable template for qualitative researchers considering LLM assistance in codebook development while preserving human interpretive authority.

Clinical diagnosis at hospital admission must be made rapidly from limited, incomplete evidence. Existing diagnosis-prediction benchmarks are poorly suited to this setting: they restrict prediction to closed code sets, exclude free-text notes, and supervise with discharge diagnoses that incorporate the full inpatient course. We introduce EarlyDx, a large-scale benchmark for open-ended early diagnosis, built from 154,834 emergency department encounters in MIMIC-IV. Each encounter is restricted to records available at admission time $t_0$ and supervised by the diagnoses recorded during the ED encounter rather than at discharge. An LLM auditor further verifies every free-text label as supported, partially supported, or unsupported by that evidence; the primary evaluation scores only fully supported labels. Under a semantic LLM-as-judge protocol, no evaluated system --- frontier general, medical-specialized, or in-domain post-trained --- synthesizes admission-time evidence reliably. Zero-shot models score largely by extraction, recovering only 3-31% of diagnoses that must be inferred rather than read from the record; post-training raises inference-dependent recall to 56%, but a sizeable margin remains, and on time-critical conditions no system attains a clinician's balance of sensitivity and precision. We release the full construction and evaluation pipeline at here.

LLM tutoring poses a measurement problem: can a general-purpose helpfulness rubric distinguish direct answer-giving from pedagogical guidance? We audit this signal in a pre-registered study. Within each of three tutor bases, we compare conversational and pedagogical policies instantiated with the same underlying model and paired with one fixed weak simulated student. Deterministic detectors measure answer leakage and next-turn independent work. Claude Opus 4.8 is the frozen, condition-blind primary judge. After the Opus scores were fixed, GPT-5.6 Sol was prospectively specified for a post hoc robustness audit of the same 1,179 confirmatory answer-phase tutor turns under the frozen helpfulness and pedagogy rubrics. On the primary base under Opus, the policies do not differ significantly in helpfulness but are perfectly rank-separated under the pedagogy rubric (Cliff's $|δ|{=}0.10$ vs. $1.0$). Across the two judges, pedagogy contrasts retain their direction where detected, whereas the helpfulness ordering is judge-contingent, reversing between judges on two of three bases. In an Opus-only ablation, seven primary-base policies span $2.3$ points in mean judged pedagogy within a $0.25$-point band of mean judged helpfulness. Separately, answer-revealing turns are followed by less independent student work on every base, a result that is judge-invariant by construction. In this controlled setting, general-purpose helpfulness is not a reliable pedagogy signal. Tutor evaluation should pair pedagogy-targeted rubrics with deterministic process measures.

Search-augmented language agents should retrieve external information only when necessary and ground their answers in retrieved evidence. Existing external rewards provide either sparse outcome supervision or richer feedback from process annotations and LLM judges. Outcome rewards scale readily but cannot distinguish grounded retrieval from redundant search, whereas richer signals require costly annotation or inference during training. Internal rewards based on policy-side signals such as entropy, likelihood, or information gain are graded and inexpensive to evaluate, yet mainly reflect model confidence rather than evidence grounding. We propose Search-G1, a representation-based intrinsic reward framework that measures the operational grounding of an agent's answers through two intervention-calibrated readouts. A prompt-state readout predicts closed-book sufficiency, whose complement defines policy-relative retrieval necessity; an answer-commit readout estimates evidence reliance from answer-stage sensitivity to evidence deletion. Together, they provide additional credit to correct searched trajectories when retrieval is estimated necessary and the answer is evidence-sensitive, favor correct direct answers when closed-book knowledge suffices, and penalize repeated search. After calibration, reward scoring requires neither process annotations nor LLM-as-judge inference during policy optimization. Because reinforcement learning changes policy representations, Search-G1 periodically refits both readouts on trajectories from the latest checkpoint, allowing the reward to co-evolve with the policy. Experiments across multiple search-based question-answering benchmarks and two model scales show that Search-G1 improves the grounding--search-cost trade-off, producing shorter response-side trajectories at competitive task accuracy. Code is available at https://github.com/Rosy0912/Search-G1.

Long-term memory is becoming a core component of LLM agents, but most memory benchmarks evaluate conversations or compact summaries, while research agents need to restore evidence from full scientific papers. We introduce two full-text scientific-memory benchmarks, Public AI Memory (PAIM; 81 papers, 66 questions) and Public Transformers (PTr; 252 papers, 98 questions). We evaluate eight memory/retrieval systems, including our own proposed Theoria, plus a no-retrieval baseline. Our results show that memory leaderboards are not interpretable without the full protocol: ingestion granularity, raw-text preservation, retrieval budget, retrieval modality, rubric audit, and judge choice all affect the outcome. For example, on PAIM Graphiti wins convincingly but uses 2.6M characters of retrieved context per query, and after controlling for retrieval budget the lead disappears. On PTr, for the systems where BM25 retrieval can be added cleanly, the sparse-dense hybrid is the single most significant intervention: hybrid variants of Simple RAG, Mem0, and Theoria tie for the lead within 0.03 points. Multi-judge and human side-by-side calibration show that LLM-as-a-judge rankings are consistent across frontier judges and agree with human evaluation, with an effective resolution of roughly one point on a ten-point scale. We argue that scientific memory should be evaluated as budgeted, modality-aware context restoration rather than as an unconstrained architecture leaderboard, and we release the datasets, harness, raw outputs, judgments, and scripts to reproduce our results and serve as tools for such evaluation. Our code is available at http://gitlab.com/quantellence/research/scientific-recall-bench , and the datasets are available at http://huggingface.co/datasets/quantellence/srb-data .

Probably not, at least for meta-analyses in economics. In a pre-registered, identity-masked, within-paper experiment, the authors of 44 meta-analyses ranked three AI reports on their own paper by usefulness for improving it: a single pass by a frontier model against two multi-agent debate tools we built and expected to win. All reports were held to a common length and template. The authors preferred the single pass, by 0.66 rank points over mad-research (95% CI 0.32 to 1.00) and 0.57 over paper-workshop (0.16 to 0.95), though paper-workshop spent roughly thirty times the tokens. Authors who recalled their journal referee report usually placed it first and never last; in a separate exercise, three AI judges almost always placed the real journal referee report last. Among the three AI reports, Gemini (the judge whose model family wrote none of the reports) would have ranked paper-workshop first in the authors' place, reversing the single-pass preference. The reversal warns against substituting an AI judge for the author. We measure perceived usefulness for finished papers; whether AI should referee papers is a separate question.

Large language models routinely produce fluent answers to single-shot prompts, yet deploying them as reliable components of a domain decision system is substantially harder. Closing this gap is the work of harness engineering: the deliberate composition of deterministic scaffolding (symbolic filters, retrieval, schema-typed I/O, LLM-as-judge loops, HITL gates, persistent state, audit trails) around an LLM core. We present a case study in academic supervision, a domain combining high-stakes recommendation, longitudinal accountability, and structured operational workflows. We compare a baseline (ASA), a GPT-5 chatbot with no scaffolding, against a multi-module system (ASuS) that wraps the much smaller GPT-4o-mini in a LangGraph harness with symbolic-semantic retrieval, schema-validated outputs, LLM-as-judge with bounded retry, HITL gates, deterministic weighted risk scoring with LLM narration, and a per-node SQLite audit trail. The evaluation rubric is retargeted at six harness-mechanism dimensions (grounding, explainability, consistency, process integrity, cognitive load, constraint adherence). A blind ten-rater hybrid evaluation, supplemented by a 2 x 2 model-harness ablation, finds that ASuS, despite using a much smaller base model, outscores ASA on every dimension. Across ten raters the pooled mean for ASuS is 4.08 versus 1.23 for ASA, and 8 of 10 raters reject the null at alpha = 0.05 on a paired Wilcoxon test; full numbers are in Sections 6.4 and 6.7. The ablation confirms that the structural contributions of the harness are largely model-invariant. We extract seven recurring harness-engineering patterns and argue that where reliability, traceability, and institutional consistency matter more than open-ended fluency, harness engineering challenges the prevailing 'bigger model is better' intuition.

Multi-reference-to-audio-video (MR2AV) generation aims to generate coherent audio-video content conditioned on multiple references and textual instructions. Existing benchmarks mainly focus on text-driven generation, single-reference subject preservation, or isolated audio-video alignment, leaving the emerging MR2AV setting largely unexplored. Compared with these settings, MR2AV requires models to jointly reason over multiple references while generating synchronized visual and audio content. Models must not only preserve each reference faithfully but also correctly bind and compose multiple referenced entities into coherent audio-visual events. To address this gap, we introduce MultiRef-Compass, a unified benchmark for MR2AV generation. It comprises $350$ carefully curated samples constructed through a scalable and controllable asset-composition pipeline, covering multi-view subject preservation, multi-entity binding, and human-object-scene composition. To provide interpretable assessment, MultiRef-Compass defines an evaluation protocol with four dimensions: Basic Quality, Reference Consistency, Audio-Visual Consistency, and Instruction Following, using 14 sub-metrics. MultiRef-Compass integrates automatic metrics with a rejudging-enhanced MLLM-as-a-Judge framework, enabling scalable and auditable evaluation of both perceptual fidelity and reference-conditioned composition. Extensive experiments on eight representative MR2AV systems reveal substantial room for improvement across multiple evaluation dimensions, underscoring the need for a comprehensive benchmark and positioning MultiRef-Compass as a foundation for future MR2AV research.

Studies of bias in LLM-as-judge systems typically build synthetic corpora by prompting an LLM to generate a hallucinated answer to pair with a factual one, then presenting both to a judge. We report a case in which this generation step silently failed, and use it to argue that the failure mode is structural rather than incidental. In a multilingual (Turkish/English) faithfulness-judgment corpus, a decoding-budget parameter shared between judging and generation calls truncated one producer's hallucinated answers to a few words. The resulting items produced a large, statistically robust effect: a 32-point cross-lingual collapse in one judge's selection accuracy, replicated from N=50 to N=500, explained by a three-layer mechanistic account, and confirmed by a controlled producer-swap experiment, none of which was real. The effect vanished to ceiling once the shared parameter was corrected, and only manual reading of the raw generations, not any aggregate statistical check, exposed the fault. A second measured bias (Markdown-formatting preference) was not fabricated but distorted by the same fault, its magnitude and in one case its sign shifting with stimulus length, a mode aggregate metrics cannot distinguish from the first. We frame the underlying vulnerability using the test oracle problem: corpora whose negative examples are LLM-generated carry no mechanical way to verify item integrity, while corpora built by deterministic perturbation of a gold answer carry an item-level oracle for free. A positive control supports this claim directly: an analogous fault injected into a minimal perturbation-based corpus is caught with 100% accuracy by a zero-cost, zero-human gold-to-negative string comparison. We close with a validation protocol, derived from our own case, for analysts working in the oracle-less regime that we argue describes most contemporary multilingual LLM-as-judge corpora.

LLM-as-a-judge is widely used to provide feedback and selection signals in closedloop regeneration, but this use remains insufficiently validated. We study it in table recognition, where deterministic TEDS evaluation provides a controlled testbed, using FinTabNet and OmniDocBench. Three findings emerge. First, judge signals were weak on both datasets: scores frequently tied, rankings were not reproducible, and the only selection policy that beat random on both datasets depended on an earliest-iteration tie rule, so its advantage cannot be attributed to the judge scores alone. Iteration produced better candidates, but the judge failed to recover them. Second, severe losses occurred even without specific judge feedback. A structurepreserving instruction significantly reduced the severe-loss rate on FinTabNet and was directionally consistent on OmniDocBench. The contrasts support target-preservation failure under unconstrained regeneration as a proximate mechanism of the observed severe losses. Third, the structure-preservation constraint reduced the severe-loss tail but produced no improvement. In an exploratory 2x2 analysis, the same protection was not stably observed when judge feedback was retained. These results do not dispute the value of LLMs as evaluators. Instead, they show that evaluation ability does not imply optimization utility. Iterative refinement requires, at minimum, a verification signal that deterministically detects structural change, rather than judge scores alone.

LLMs now sit on every side of evaluation: as examinees scored on benchmarks, judges of other models' outputs, and raters of human-generated content. Each paradigm can be viewed as a measurement problem, where a latent property of an object is probed with items from an instrument (e.g., benchmark) by judges or raters. Standard evaluation practices often neglect the contributions of each core component to the end result, limiting our understanding of what is being measured. Rasch measurement theory (RMT) is well-suited to this problem. RMT decomposes ordinal ratings into separable facets on a common scale. It further provides a battery of diagnostics that can identify miscalibrated measurements and rater biases. We present a case study of RMT applied to the LLM-as-rater paradigm using the Measuring Hate Speech corpus, whose construct was itself built under RMT. We fit a series of many-facet Rasch models to annotations from nine LLMs spanning families and capability levels. Our analyses show that LLMs systematically differ from human raters in severity, item-level calibration, question-order robustness, target-identity sensitivity, and rating scale use, all of which standard evaluation practice would largely obscure. Overall, we argue that RMT belongs in the toolkit for evaluating LLM-as-examinee, -judge, and -rater paradigms.

In the rapidly evolving landscape of information retrieval systems, the ability to adapt and improve through user feedback is paramount. This study introduces a novel methodology for refining the performance of a primary Retrieval Augmented Generation (RAG) system by strategically integrating an auxiliary feedback RAG system. By systematically harnessing human-generated feedback, the approach aims to enhance the accuracy, relevance, and overall quality of responses, driving the system towards self-improvement. Central to this methodology is a human-in-the-loop implementation, where user feedback is continuously collected, classified, and integrated into the inference workflow, enabling the system to learn and evolve iteratively. To validate the effectiveness of this approach, the study employs rigorous testing against three diverse benchmark datasets focused on general and custom domain knowledge, utilizing a LLM-as-a-Judge evaluation strategy. This comprehensive framework not only underscores the transformative potential of feedback-driven enhancements in RAG systems but also sets a precedent for future research in adaptive information retrieval technologies, marking a significant step in the journey towards autonomous refinement and optimization through user engagement.

LLM-as-a-judge evaluation is widely used for retrieval-augmented generation (RAG), but reusing the same model family as both generator and judge makes self-leniency difficult to identify. We introduce Eval-Pair Matrix, a controlled meta evaluation protocol for source-grounded RAG. Starting from GaRAGe questions and grounding passages, we induce one hidden answer-causal contradiction per record, generate answers from perturbed passages with GPT, Grok, and Gemini models, and then use the same models as blind judges to evaluate each answer against the original passages. The experiment contains 300 core records, 897 labeled generator outputs, and 2,683 judge verdicts in a crossed 3 x 3 matrix; the primary analysis uses 275 fully validated records. Instead of comparing diagonal and off-diagonal cells across different answers, we estimate same-model effects by pairing judges on the exact same candidate answer. This changes the interpretation: diagonal and off diagonal F1 are similar, and the paired same-model recall effect is near zero (-0.5 pp; 95% cluster bootstrap CI [-2.7, +1.7]). The only robust paired gap is lower matching-judge flagging for answers that avoided the induced claim (-4.3 pp). A targeted human evaluation finds that reviewed apparent false positives are alternate source-error detections, mistakes in labeling whether the induced claim was adopted, or unclear cases; none were adjudicated as genuine false alarms. The lesson is methodological: RAG judge studies should report full matrices, answer-paired effects, behavior strata, and label-task alignment.

When an LLM judge calls a peer review analytical and a human committee calls another review high quality, are they tracking the same thing? We argue they are not, and that the difference matters philosophically. We operationalise Kahneman's dual-process theory into a structured rubric for peer review and release Kahneman4Review, a benchmark of 3,563 rated reviews scored along nine theoretically motivated textual dimensions, eight bias diagnostics, and a continuous reasoning-quality score. Three findings bear on trustworthiness: decision tier is not detectably aligned with the rubric's text-grounded epistemic-quality proxy; public-showcase agentic reviews receive higher raw scores than pooled human reviews, but length and venue explain most of the gap and the samples are not paper-paired; and ICLR review-text diagnostics shift at the 2022--2023 transition, temporally coincident with widespread LLM availability but without identifying its cause. A matched function-probe pilot further shows that the rubric distinguishes textual probes designed to contrast genuine fault-finding with surface fluency. We argue that a trustworthy reliability benchmark for LLM judges must separate analytical form from epistemic function, and propose concrete design choices toward that goal. An interactive demo is available at https://huggingface.co/spaces/nuojohnchen/Kahneman4Review.

Selecting the correct answer from a pool of candidate reasoning chains is the engine of test-time scaling, yet the standard selectors each carry a cost: self-consistency inherits the errors of the single model it resamples, and trained reward models need labeled data and transfer poorly off-distribution. We study a third signal, free at inference time: cross-model consensus, the degree to which independently trained models, each solving the problem once, agree on a final answer. We treat the panel as an LLM-jury, in which the structure of agreement, not any model's score of another, is the verification signal. Across seven benchmarks it selects correct answers better than self-consistency and far better than a model scoring its own candidates: on competition math it closes the entire gap to an oracle selector, while self-scoring closes almost none. The mechanism is error decorrelation: independently trained models err differently, so their wrong answers scatter while the correct one accumulates agreement. We make this precise with a parameter-free law, derived in closed form, that predicts consensus accuracy from three measured panel statistics to a mean absolute error of $0.03$ and exposes the method's ceiling: a shared-error floor where models share a misconception, near zero on math but non-trivial on science. Against four trained verifiers spanning discriminative, outcome, and generative reward models, the free LLM-jury matches the strongest inside their math training domain and is the top selector outside it. Cross-model consensus is thus a verifier we can characterize in advance: a law that says when to trust it, and a floor that marks where it cannot.

An LLM-as-judge score can move even when the candidate responses stay fixed, simply because the evaluator has changed. We treat this evaluator-replacement ambiguity as a measurement-validity problem. Across four judgment datasets, we compare two upgrade paths available in practice: scaling Qwen3 dense judges from 1.7B to 32B parameters and moving across MiniMax M2-M2.7 released APIs. The main pattern is that judge upgrades are not interchangeable: only Qwen3 1.7B to 4B gives a robust adjacent gain, while MiniMax adjacent releases do not. Stronger judges reduce but do not remove position and verbosity bias. Repeated-sample juries add little when errors are correlated. Structured debate can move decisions substantially, but without parser and fallback logs those shifts cannot be attributed to deliberation. We argue that LLM-as-judge reports should include dataset slices, bias probes, error-dependence estimates, and protocol audit trails.

Personality recognition has traditionally been constrained by theory-dependent formulations, where models are trained to fit predefined psychological taxonomies rather than uncovering shared underlying behavioral structure. This limits generalization, as personality itself is better understood as theory-invariant, while existing annotations reflect only partial and sometimes inconsistent views of the same latent traits. In this work, we introduce JAM ((J)udge for (A)daptive (M)etric-Alignment), a theory-agnostic framework that shifts learning from adapting to predefined personality theories toward discovering unified latent pseudo-facets that capture shared psychological structure. Rather than constraining the model to any personality taxonomy during training or inference, the framework learns generalizable psychological representations and can infer an individual's latent psychological profile directly from the textual samples, without requiring theory-specific labels. JAM achieves this through an Attention-Pooled Graph Prototypical Network that learns structured representations via clustering in embedding space, together with a Cross-Theory Harmonization (CTH) approach that integrates (i) Human-Guided Linkage and (ii) Machine-Induced Consensus to unify heterogeneous datasets without relying on predefined labels. To further improve robustness and data quality, we incorporate an LLM-as-a-Judge mechanism operating in two configurations, (i) LLM-before-the-loop and (ii) LLM-in-the-loop which identifies ambiguous samples to guide adaptive metric learning. Experiments show that JAM improves cross-framework generalization and performance, establishing a strong step toward theory-agnostic personality inference and supporting low-resource personality theories. The related code repository, model weights, and artifacts are available at https://research.jingjietan.com/JAM

Audio Descriptions (ADs) narrate visual content for Blind and Low Vision (BLV) audiences during gaps in audiovisual media. There is growing momentum around ADs in movies and TV shows, and with mandates from India's Central Board of Film Certification (CBFC), there is a need to expand ADs beyond English. Yet, there is no work that generates ADs for any Indian language. To address this gap, we present the first systematic study of ADs in Hindi, contributing to aspects such as data, generation, and evaluation. We introduce Andha-Dhun, the first dataset of human-authored Hindi ADs collected from 8 full-length movies. We explore two approaches for generating ADs in Hindi: (i) directly from English dense video descriptions, and (ii) translating English ADs into Hindi. We evaluate these approaches using perplexity and LLM-as-a-judge metrics to assess fluency and quality respectively. We also analyze movies that have both English and Hindi human-authored ADs and find that naive translation introduces artifacts and narrows diversity compared to original Hindi ADs. Direct machine translation fails to adapt cultural references, while human-translated ADs do better but still fall short. Our findings emphasize that the purpose of Hindi ADs is accessibility for Indian BLV audiences, and that this requires adapting content for the audience more than strict fidelity to the source.

Training a language model against its own reference-free judgments (the premise of self-rewarding, self-play, and LLM-as-a-judge pipelines) assumes a model's verdict on a shown answer tracks correctness. We show it fails structurally: conditioned on a candidate, a judge scores plausibility, not correctness, leaving false-positive basins a policy learns to exploit. We measure this with a hidden-anchor audit: a held-out, cross-source exact-match check the judge never sees. On GSM8K with Qwen3 policies, self-play drives the judge's pass rate from 0.72 to 0.94 while true accuracy stays at 0.20 (three seeds). This reward hacking is not white-box gaming: the errors transfer across judge families (Qwen, Llama, Gemma) and scales, a strict three-judge ensemble still accepts 55% of them, and no plausibility-scoring defense closes the basin. The decisive variable is whether the judge commits an answer of its own before using the candidate: committing first drops the false-positive rate from 0.719 to 0.012, blind solving lifts discrimination to 0.96, and used as the training reward the de-anchored channel keeps false positives at zero, preventing the basin rather than only detecting it. A falsifiable bound (the gap is at most 1 - accuracy) predicts which regimes are exposed. The full arc replicates without training under best-of-N selection in code and competition math, and with a Gemma policy.

LLM-as-a-Judge has become the dominant evaluation paradigm for many natural language generation tasks (albeit mostly in English) due to shortcomings of conventional metrics and high correlations with human judgment. There are now attempts to extend LLM-as-a-Judge to multilingual settings including low-resource languages. However, LLMs have limited proficiency in low-resource languages, and there is often no adequate human validation in these settings. To highlight the scope of the problem and current practices, we explore the use of LLM-as-a-Judge evaluators in ACL Anthology papers focusing on multilingual settings and low-resource languages across a diverse set of tasks. Out of 650 papers mentioning LLM-as-a-judge, only 33 of them focus on low-resource or multilingual settings. Our in-depth analysis of these papers indicates inconsistent evaluation outcomes, a tendency to overtrust LLM judgments in multilingual settings, and the widespread reliance on a single judge model per study. To help the NLP community further, we conclude with a checklist of recommendations about how to use LLM-as-a-Judge in multilingual and low-resource settings

Large Language Models are increasingly deployed in emotional-support contexts and crisis-related situations. Nevertheless, their cross-lingual abilities in these circumstances remain underexplored. Existing benchmarks emphasize multilingual performance but rarely examine crisis-related empathy and cultural grounding in low-to-mid-resource languages. We introduce SPLIT, a 500-prompt benchmark designed to evaluate LLM consistency in generating emotionally grounded responses across five categories: Stress, Panic, Loneliness, Internal Displacement, and Tension. We evaluate three technically diverse LLMs across three dimensions: Empathetic Accuracy, Linguistic Naturalness, and Contextual & Cultural Grounding. The framework aims to assess and compare the quality of LLM responses in both English and Ukrainian languages, as well as to explore the reliability of the LLM-as-a-jury paradigm. Our findings reveal that Gemini-2.5-Flash and LLaMA-3.3-70B-Instruct degrade when transitioning to Ukrainian, while DeepSeek-V3 remains comparatively stable within our benchmark. We additionally find that human and AI evaluators agree weakly on empathy and naturalness but diverge on cultural grounding. We further argue that producing Ukrainian text is not equivalent to producing Ukrainian emotional support. Our findings may assist in the future development of more culturally tailored benchmark designs, as well as encourage a stronger emphasis on human-centered evaluation.

Open web-scale pre-training corpora remain concentrated in English, limiting multilingual LLM development. We introduce MultiSynt/MT, an open synthetic parallel corpus with approximately 4.8 trillion target-language tokens across 36 languages, produced by translating 100 billion high-quality Nemotron-CC tokens with Tower+ and OPUS-MT/HPLT-MT systems. For many medium- and lower-resource European languages, this is the largest openly available pre-training resource. Across five high- and medium-resource languages, reference LLMs trained on MultiSynt/MT reach the final score of HPLT 2.0, a native-data baseline, using roughly 72% fewer pre-training tokens, and outperform it by approximately 15% relative at a matched 100B-token training budget. Our analyses also identify evaluation blind spots: standard multiple-choice benchmarks miss translation-quality differences that a fluency-sensitive LLM-as-judge protocol recovers on the trained LLMs without detecting a deficit relative to its native-data baseline, while Norwegian idiomatic and culturally grounded tasks remain better served by native data. We release the corpus, including row-aligned translations from multiple systems, to support controlled research on multilingual pre-training data and evaluation.

Current large-language-model (LLM) physics benchmarks are usually scored by answer accuracy, which cannot distinguish genuine reasoning from recall of familiar problem patterns and reveals little about where a model's reasoning breaks down. We introduce an auditable four-stage diagnostic that evaluates whether an LLM can reason inside an unfamiliar physics framework through induction, formulation, prediction, and review. The diagnostic combines locked pre-registrations, fresh sessions between stages, dual-LLM judging, and a human-audit pathway, and we apply it to three parallel physics worlds: a single-equation counterfactual world ($F=mv$), a historical framework (Aristotelian mechanics), and a four-domain counterfactual world (Decay World). Across Claude Opus 4.7, GPT-5.5, and Gemini 3.1 Pro, the three worlds yield composite PASS rates are 6/15, 6/15, and 0/15 respectively (content $\land$ structural for $F=mv$ and Aristotelian, content axis only for Decay World where the structural axis is out of scope). The most pointed empirical pattern is a qualitative-versus-quantitative asymmetry: in Decay World, models almost never predict the wrong direction of change, but frequently compute the wrong ratio by slipping back to standard-physics relations. The protocol also surfaces two methodology findings: LLM-judge reliability does not transfer across frameworks, and Stage 4 self-review is weak in every framework, with the model's own review wrongly reporting no earlier error in at least two-thirds of the trials that actually contained one. We release the full prompts, responses, verdicts, and audit records.

Large language models (LLM) trained on web-scale corpora generate output that may infringe copyright, yet existing technical safeguards focus narrowly on verbatim memorisation. EU copyright doctrine applies a broader standards: substantial similarity, which extends to stylistic choices, narrative structure, and creative elaboration. This mismatch between what current methods detect and what the law protects leaves a significant compliance gap. We introduce PSALM, an LLM-as-a-judge framework that operationalises EU copyright doctrine through ten evaluators assessing computational overlap, stylistic dimensions (writing style, narrative voice), content dimensions (character, plot, scene, world building), and statutory exceptions (parody, pastiche, quotation, scènes à faire). Applying PSALM to Llama~3.2 models fine-tuned on translated historical Dutch literary works, we find that: 1) instruction-tuned models exhibit non-trivial baseline stylistic similarity prior to corpus exposure; 2) fine-tuning induces systematic stylistic appropriation across all infringement-relevant dimensions, extending beyond verbatim memorisation to abstract narrative patterns; 3) Negative Preference Optimisation unlearning substantially reduces similarity but leaves detectable residual stylistic patterns. These findings indicate that safeguards targeting literal copying alone are insufficient to mitigate broader copyright risks. PSALM provides infrastructure for auditable, legally informed compliance evaluation, though the relationship between automated similarity scores and infringement determinations requires validation by legal experts. This work bridges qualitative legal standards and quantitative technical measurement, exposing fundamental tensions between generative AI and EU intellectual property law.

The LLM Jury, a Panel of LLM Evaluators (PoLL) reporting consensus scores, has become a practical alternative to single-judge LLM evaluation, yet its statistical behavior remains poorly understood. We formalize the LLM Jury under the Huber contamination model and show that PoLL incurs unbounded bias under any positive contamination, regardless of jury size, whenever a single judge fails in a biased, LLM-typical way (mode collapse, sycophancy, safety refusal). Framing jury consensus as classical robust mean estimation, we propose RoPoLL (Robust Panel of LLM-as-Judge), which preserves the PoLL panel but replaces the aggregation function with a robust mean estimator, instantiated with the geometric median (GM): tuning-free, with the optimal finite-sample breakdown point 1/2. A finite-sample error bound and a matching information-theoretic minimax lower bound agree on the parametric rate sigma*sqrt(d/N) and differ on the breakdown floor by a factor of sqrt(d), a statistical-computational gap that polynomial-time RoPoLL pays relative to the intractable Tukey halfspace median. Across 13 open-weight judges (4B-675B), three reward-model benchmarks, and four corruption regimes at rates up to 50%, RoPoLL dominates PoLL on every biased corruption type: by about 19% on cross-dimensional attacks at matched compute, and by orders of magnitude on heavy-tailed Byzantine adversaries. A 3-judge RoPoLL committee at 38B beats Mistral-Large-3 (675B) by 1.31x on HelpSteer-2 under 30% bimodal-random corruption, an 18x parameter advantage at better accuracy; a Noisy-GT control confirms the premium is paid against biased contamination, not benign imprecision.