LLM-as-a-Judge

Autoraters, also referred to as LLM-as-judges, are increasingly used for evaluation and automated content moderation. However, there is limited statistical analysis of how modifications in a rubric presented to both humans and autoraters affect their score agreement. Rubrics that ask for an overall or holistic judgment - for example, rating the quality'' of an essay - may be inconsistently interpreted due to the complexity or subjectivity of the criteria. Conversely, rubrics can ask for *analytic* judgments, which decompose assessment criteria - for example, quality'' into fluency'' and organization''. While these rubrics can be edited to improve the individual accuracy of both human and automated scoring, this approach may result in disagreement between the two scores, or with the associated holistic judgment. Designing and deploying reliable autoraters requires understanding not just the relationship between human and autorater annotations but how that relationship changes as holistic or analytic judgments are elicited. The results indicate that rubric edits providing representative examples and additional context, and reducing positional bias in the rubric increased human-autorater agreement, while higher rubric complexity and conservative aggregation methods tended to decrease it. The findings from the automatic essay scoring and instruction-following evaluation domains suggest that practitioners should carefully analyze domain- and rubric-specific performance to move towards higher human-autorater agreement.

LLM-as-a-Judge pipelines have become the de facto evaluator for agent safety, yet existing benchmarks treat their verdicts as ground-truth proxies without checking whether the verdicts depend on the agent's behavior or merely on how the evaluation policy happens to be worded. We argue that any trustworthy safety judge must satisfy a basic property we call policy invariance, and we operationalize it as three testable principles: rubric-semantics invariance under certified-equivalent rewrites, rubric-threshold invariance under intentional strict-to-lenient shifts, and ambiguity-aware calibration so that verdict instability concentrates on genuinely ambiguous cases. Instantiating these principles as a stress-test protocol with four agent-class judges on trajectories drawn from ASSEBench and R-Judge, we surface a previously unmeasured failure mode: today's judges respond to meaningful normative shifts and to meaningless structural rewrites with comparable strength, and cannot tell the two apart. Content-preserving policy rewrites flip up to 9.1% of verdicts above baseline jitter, and 18-43% of all observed flips occur on unambiguous cases under such rewrites, so existing safety scores conflate what the agent did with how the evaluator was prompted. Beyond the diagnosis, we contribute the Policy Invariance Score and the Judge Card reporting protocol, which expose an order-of-magnitude spread in judge reliability that is invisible to accuracy-only leaderboards. We release the protocol and code so that future agent-safety benchmarks can audit their own evaluators rather than trust them by default.

Large-language-model (LLM) graders promise to relieve the grading burden of upper-division STEM courses, but most deployments to date send student work to third-party APIs, violating FERPA and exposing institutions to data risk while requiring substantial assignment modification. We present $\textbf{LaTA}\ (\textit{LaTeX Teaching Assistant})$, a drop-in, open-source autograder that runs entirely on commodity on-premises hardware and assumes a LaTeX-native workflow already adopted by many engineering and physics courses. LaTA implements a four-stage pipeline (ingest, segment, grade, report) using a locally hosted open-weight chain-of-thought LLM grader (gpt-oss:120b) that compares student work to an instructor-authored reference solution and applies a YAML rubric with binary per-item scoring. We deployed LaTA in Winter~2026 in ME 373 (Mechanical Engineering Methods) at Oregon State University, grading every weekly assignment for approximately 200 students on a single Mac Studio at $0 marginal cost per assignment and 1--3 minutes of wall-clock time per submission, enabling regrading of corrected assignments and greatly expanded TA office hour offerings. The instructor-confirmed grading-error rate held at roughly $0.02$--$0.04\%$ per rubric line item across the term. Relative to the same instructor's previous traditionally-graded cohort, the LaTA-graded cohort outperformed by approximately $11\%$ on the midterm exam and $8\%$ on the final exam, and reported large gains in self-assessed confidence on every stated learning objective ($N = 159$ survey responses, $Δ\geq +1.49$ Likert points, $p < 10^{-27}$ on every comparison). We release the code under AGPLv3.

This intervention study investigates the use of counterarguments in writing for critical thinking by students in the context of Generative AI (GenAI). This is especially as risks of cheating and cognitive offloading exist with the use of GenAI. We presented 36 students in a particular university course with 4 carefully selected thesis statements (from a set of popular debates) to write about anyone of them. We used six established rubrics (focus, logic, content, style, correctness and reference) to conduct three human assessments (two student peer-reviews and one experienced teacher) per writeup on a 5-point Likert scale for all the qualified samples (n) of 35 submissions (after disqualifying one for irregularity). Using the same rubrics and guidelines, we also assessed the submissions using six frontier LLMs as judges. Our mixed-method design included qualitative open-ended feedback per assessment and quantitative methods. The results reveal that (1) the students' self-written counterarguments to AI-generated content contains logic, among other things, which is a key component of critical thinking, and (2) GenAI can be successfully used at scale to assess students' written work, based on clear rubrics, and these assessments generally align with human assessments as shown with Gwets AC2 inter-rater reliability values of 0.33 for all the models except one.

Multi-constraint instruction following requires verifying whether a response satisfies multiple individual requirements, yet LLM judges are often assessed only through overall-response judgments. We introduce MCJudgeBench, a benchmark for constraint-level judge evaluation in multi-constraint instruction following. Each instance includes an instruction, a candidate response, an explicit constraint list, per-constraint gold labels in {yes, partial, no}, and controlled response-side perturbations. The evaluation protocol further includes evaluation prompt variants to test judge stability. We evaluate proprietary and open-source LLM judges using both correctness and inconsistency metrics, distinguishing intrinsic inconsistency under stochastic decoding from procedural inconsistency under prompt and response perturbations. Our results show that judge reliability has multiple dimensions: strong overall performance does not guarantee equally reliable detection across label categories, especially for rarer partial and no cases. Judges with higher correctness do not always have lower inconsistency. Evaluation with reasoning improves correctness but does not uniformly improve stability. These findings motivate evaluating LLM judges at the constraint level to study these failure modes.

Large language models offer a tempting solution to address the peer review crisis. This position paper argues that today's AI systems should not be used to produce paper reviews. We ground this position in an empirical comparison of human- versus AI-generated ICLR 2026 reviews and an evaluation of the effect of automated paper rewriting on different AI reviewers. We identify two critical issues: 1) AI reviewers exhibit a hivemind effect of excessive agreement within and across papers that reduces perspective diversity. 2) AI review scores are trivially gameable through paper laundering: prompting an LLM to rewrite a paper could significantly increase the scores from AI reviewers, demonstrating that LLM reviewers are easy to game through stylistic changes rather than scientific results. However, non-gameability and review diversity are necessary but not sufficient conditions for automation. We argue that addressing the peer review crisis requires a science of peer review automation -- not general-purpose LLMs deployed without rigorous evaluation.

Rankings produced by holistic LLM-as-a-judge scoring are sensitive to the bias of the chosen judge model. We show that switching to binary rubric scoring with multi-judge filtering removes this sensitivity: decomposing the judgement matters more than the judge model itself. To support this claim, we introduce Prosa, the first real user multi-turn Brazilian Portuguese chat benchmark: 1,000 WildChat conversations scored by three judges from three model families on 16 models. Under filtered rubric scoring the three judges agree on every one of the 16 ranks, whereas under holistic scoring they agree on only 7 of 16. Additionally, the rubric filtering pipeline increases the average score gap between neighbouring models by 47%, thereby improving Prosa's discriminative power. Evaluating a new model on Prosa costs approximately $2.1 when using Gemini 3 Flash as the judge. We release the benchmark and the filtering code to ensure that future models can be assessed under identical conditions. These artifacts also make our rubric-based scoring method reusable beyond Prosa, supporting other open-ended evaluation settings.

Vision-language models (VLMs) are increasingly used as automated judges for multimodal systems, yet their scores provide no indication of reliability. We study this problem through conformal prediction, a distribution-free framework that converts a judge's point score into a calibrated prediction interval using only score-token log-probabilities, with no retraining. We present the first systematic analysis of conformal prediction for VLM-as-a-Judge across 3 judges and 14 visual task categories. Our results show that evaluation uncertainty is strongly task-dependent: intervals cover ~40% of the score range for aesthetics and natural images but expand to ~70% for chart and mathematical reasoning, yielding a quantitative reliability map for multimodal evaluation. We further identify a failure mode not captured by standard evaluation metrics, ranking-scoring decoupling, where judges achieve high ranking correlation while producing wide, uninformative intervals, correctly ordering responses but failing to assign reliable absolute scores. Finally, we show that interval width is driven primarily by task difficulty and annotation quality, i.e., the same judge and method yield 4.5x narrower intervals on a clean, multi-annotator captioning benchmark. Code: https://github.com/divake/VLM-Judge-Uncertainty

LLM-based financial agents increasingly produce investment rationales before the outcomes needed to evaluate them are observable. This creates a delayed-ground-truth evaluation problem: realized returns remain the eventual arbiter of investment quality, but they arrive too late and are too noisy to guide many model-development and governance decisions. LLM judges offer a tempting shortcut for pre-deployment evaluation of AI-finance systems, but unvalidated judges may reward verbosity, confidence, or rubric mimicry rather than financial judgment. This paper introduces ValueBlindBench, a preregistered agreement-gated stress-test protocol for deciding when LLM-judged investment-rationale claims are publishable, qualified, or invalid. In a controlled market-state capital-allocation prototype with 1,000 honest decision cycles and 100 preregistered adversarial controls (1,100 trajectories, 5,500 judge calls), ValueBlindBench clears the aggregate agreement gate at $\barκ_w = 0.7168$ but prevents several overclaims. Lower-rank systems collapse into a tie-class, one rubric dimension fails the per-dimension gate (constraint\_awareness, $\barκ_w = 0.2022$), single-judge rankings are family-dependent, and terse-correct rationales receive a $Δ= -2.81$ rubric-point penalty relative to honest rationales. A targeted anchor-specificity probe further shows that financial constructs such as constraint awareness are operationally load-bearing. The scientific object is therefore not a leaderboard and not a claim to measure true investment skill. ValueBlindBench is a pre-calibration metrology layer for AI-finance evaluation: it governs whether a proposed LLM-judge-based investment-rationale claim is stable enough, agreed enough, and uncontaminated enough to be reported at all.

Objective. Clinical AI documentation systems require evaluation methodologies that are clinically valid, economically viable, and sensitive to iterative changes. Methods requiring expert review per scoring instance are too slow and expensive for safe, iterative deployment. We present a case-specific, clinician-authored rubric methodology for clinical AI evaluation and examine whether LLM-generated rubrics can approximate clinician agreement. Materials and Methods. Twenty clinicians authored 1,646 rubrics for 823 clinical cases (736 real-world, 87 synthetic) across primary care, psychiatry, oncology, and behavioral health. Each rubric was validated by confirming that an LLM-based scoring agent consistently scored clinician-preferred outputs higher than rejected ones. Seven versions of an EHR-embedded AI agent for clinicians were evaluated across all cases. Results. Clinician-authored rubrics discriminated effectively between high- and low-quality outputs (median score gap: 82.9%) with high scoring stability (median range: 0.00%). Median scores improved from 84% to 95%. In later experiments, clinician-LLM ranking agreement (tau: 0.42-0.46) matched or exceeded clinician-clinician agreement (tau: 0.38-0.43), attributable to both ceiling compression and LLM rubric improvement. Discussion. This convergence supports incorporating LLM rubrics alongside clinician-authored ones. At roughly 1,000 times lower cost, LLM rubrics enable substantially greater evaluation coverage, while continued clinical authorship grounds evaluation in expert judgment. Ceiling compression poses a methodological challenge for future inter-rater agreement studies. Conclusion. Case-specific rubrics offer a path for clinical AI evaluation that preserves expert judgment while enabling automation at three orders lower cost. Clinician-authored rubrics establish the baseline against which LLM rubrics are validated.

Automated code review (ACR) bots are increasingly used in industrial software development to assist developers during pull request (PR) review. As adoption grows, a key challenge is how to evaluate the usefulness of bot-generated comments reliably and at scale. In practice, such evaluation often relies on developer actions and annotations that are shaped by contextual and organizational factors, complicating their use as objective ground truth. We examine the feasibility and limitations of automating the evaluation of LLM-powered ACR bots in an industrial setting. We analyze an industrial dataset from Beko comprising 2,604 bot-generated PR comments, each labeled by software engineers as fixed/wontFix. Two automated evaluation approaches, G-Eval and an LLM-as-a-Judge pipeline, are applied using both binary decisions and a 0-4 Likert-scale formulation, enabling a controlled comparison against developer-provided labels. Across Gemini-2.5-pro, GPT-4.1-mini, and GPT-5.2, both evaluation strategies achieve only moderate alignment with human labels. Agreement ratios range from approximately 0.44 to 0.62, with noticeable variation across models and between binary and Likert-scale formulations, indicating sensitivity to both model choice and evaluation design. Our findings highlight practical limitations in fully automating the evaluation of ACR bot comments in industrial contexts. Developer actions such as resolving or ignoring comments reflect not only comment quality, but also contextual constraints, prioritization decisions, and workflow dynamics that are difficult to capture through static artifacts. Insights from a follow-up interview with a software engineering director further corroborate that developer labeling behavior is strongly influenced by workflow pressures and organizational constraints, reinforcing the challenges of treating such signals as objective ground truth.

Evaluating nuanced conversational travel recommendations is challenging when human annotations are costly and standard metrics ignore stakeholder-centric goals. We study LLMs-as-Judges for sustainable city-trip lists across four dimensions -- relevance, diversity, sustainability, and popularity balance, and propose a three-phase calibration framework: (1) baseline judging with multiple LLMs, (2) expert evaluation to identify systematic misalignment, and (3) dimension-specific calibration via rules and few-shot examples. Across two recommendation settings, we observe model-specific biases and high dimension-level variance, even when judges agree on overall rankings. Calibration clarifies reasoning per dimension but exposes divergent interpretations of sustainability, highlighting the need for transparent, bias-aware LLM evaluation. Prompts and code are released for reproducibility: https://github.com/ashmibanerjee/trs-llm-calibration.

Safety benchmarks such as HarmBench rely on LLM judges to classify model responses as harmful or safe, yet the judge configuration, namely the combination of judge model and judge prompt, is typically treated as a fixed implementation detail. We show this assumption is problematic. Using a 2 x 2 x 3 factorial design, we construct 12 judge prompt variants along two axes, evaluation structure and instruction framing, and apply them using a single judge model, Claude Sonnet 4-6, producing 28,812 judgments over six target models and 400 HarmBench behaviors. We find that prompt wording alone, holding the judge model fixed, shifts measured harmful-response rates by up to 24.2 percentage points, with even within-condition surface rewording causing swings of up to 20.1 percentage points. Model safety rankings are moderately unstable, with mean Kendall tau = 0.89, and category-level sensitivity ranges from 39.6 percentage points for copyright to 0 percentage points for harassment. A supplementary multi-judge experiment using three judge models shows that judge-model choice adds further variance. Our results demonstrate that judge prompt wording is a substantial, previously under-examined source of measurement variance in safety benchmarking.

Large language models are widely adopted as automated evaluation judges, yet the stability of their verdicts under semantically equivalent prompt rephrasings remains largely unexamined. We conduct a systematic empirical study of prompt-induced decision instability across multiple evaluation tasks and judge architectures. To facilitate this analysis, we release JudgeSense, a benchmark comprising hand-validated prompt-paraphrase pairs spanning factuality, coherence, relevance, and preference, drawn from established NLP benchmarks and accompanied by comprehensive decision logs. The benchmark enables the measurement of judge stability across equivalent prompts, allowing researchers to assess whether stability correlates with model scale or instruction-tuning, and to identify which tasks are most sensitive to prompt wording. Our evaluation reveals that coherence remains the primary task for distinguishing judge behavior, while factuality judgments demonstrate high stability under standard conditions. Pairwise evaluation tasks consistently exhibit position bias. Crucially, we find that model scale is not a reliable proxy for consistency; notably, as an interesting result in our analysis, the largest and newest models are not the most consistent.

LLM-as-a-Judge has become the dominant paradigm for evaluating language model outputs, yet LLM judges exhibit systematic biases that compromise evaluation reliability. We present a comprehensive empirical study comparing nine debiasing strategies across five judge models from four provider families (Google, Anthropic, OpenAI, Meta), three benchmarks (MT-Bench n=400, LLMBar n=200, custom n=375), and four bias types. Our headline practical finding is that a mid-tier model with the right debiasing can outperform frontier judges at a fraction of the cost: Gemini 2.5 Flash with the Combined Budget strategy reaches the highest agreement of any configuration we tested (71.0%, kappa=0.549) at ~$0.001 per evaluation, about 15x cheaper than the best frontier setup (Claude Sonnet 4, 69.5$0.015). Other key findings: (1) Style bias is the dominant bias (0.10-0.76 across models, favoring markdown over plain prose), far exceeding position bias (<=0.04), yet is rarely studied. (2) Verbosity bias is heterogeneous when measured length-aware: Pro, Flash, and Llama prefer longer answers (+0.24 to +0.44), Claude prefers concise (-0.12), and GPT-4o is neutral (-0.04); on truncation controls all models correctly prefer the complete response (0.88-1.00 accuracy). (3) Debiasing helps multiple models: Claude S8 (+11.5 pp), Flash S8 (+7.5 pp), and Claude S5 (+7.3 pp) survive Holm-Bonferroni correction, with Flash S1 (+4.7 pp) and Llama S8 (+4.5 pp) also significant. We release our evaluation framework, the 375-pair controlled dataset, and per-instance cached results for all nine strategies.

Recent advancements in large language models have led to significant improvements across various tasks, including mathematical reasoning, which is used to assess models' intelligence in logical reasoning and problem-solving. Models are evaluated on mathematical reasoning benchmarks by verifying the correctness of the final answer against a ground truth answer. A common approach for this verification is based on symbolic mathematics comparison, which fails to generalize across diverse mathematical representations and solution formats. In this work, we offer a robust and flexible alternative to rule-based symbolic mathematics comparison. We propose an LLM-based evaluation framework for evaluating model-generated answers, enabling accurate evaluation across diverse mathematical representations and answer formats. We present failure cases of symbolic evaluation in two popular frameworks, Lighteval and SimpleRL, and compare them to our approach, demonstrating clear improvements over commonly used methods. Our framework enables more reliable evaluation and benchmarking, leading to more accurate performance monitoring, which is important for advancing mathematical problem-solving and intelligent systems.

Evaluating LLM-generated business ideas is often harder to scale than generating them. Unlike standard NLP benchmarks, business idea evaluation relies on multi-dimensional criteria such as feasibility, novelty, differentiation, user need, and market size, and expert judgments often disagree. This paper studies a methodological question raised by such disagreement: should an automatic judge approximate an aggregate consensus, or model evaluators individually? We introduce PBIG-DATA, a dataset of approximately 3,000 individual scores across 300 patent-grounded product ideas, provided by domain experts on six business-oriented dimensions: specificity, technical validity, innovativeness, competitive advantage, need validity, and market size. Analyses show substantial expert disagreement on fine-grained ordinal scores, while agreement is higher under coarse selection, suggesting structured heterogeneity rather than random noise. We then compare three judge configurations: a rubric-only zero-shot judge, an aggregate judge conditioned on mixed evaluator histories, and a personalized judge conditioned on the target evaluator's scoring history. Across dimensions and model sizes, personalized judges align more closely with the corresponding evaluator than aggregate judges, and evaluator agreement correlates with similarity of judge-generated reasoning only under personalized conditioning. These results indicate that pooled labels can be a fragile target in pluralistic evaluation settings and motivate evaluator-conditioned judge designs for business idea assessment.

LLM-as-a-Judge has become a dominant approach in automated evaluation systems, playing critical roles in model alignment, leaderboard construction, quality control, and so on. However, the scalability and trustworthiness of this approach can be substantially distorted by Self-Preference Bias (SPB), which is a directional evaluative deviation in which LLMs systematically favor or disfavor their own generated outputs during evaluation. Existing measurements rely on costly human annotations and conflate generative capability with evaluative stance, and thus are impractical for large-scale deployment in real-world systems. To address this issue, we introduce a fully automated framework to quantifying and mitigating SPB, which constructs equal-quality pairs of responses with negligible quality differences, enabling statistical disentanglement of discriminability from bias propensity without human gold standards. Empirical analysis across 20 mainstream LLMs reveals that advanced capabilities are often uncorrelated, or even negatively correlated, with low SPB. To mitigate this bias, we propose a structured multi-dimensional evaluation strategy grounded in cognitive load decomposition, which reduces SPB by 31.5% on average.

LLM-as-a-Judge is now widely used to rank model outputs, train reward models, and populate public leaderboards, but its run-to-run reliability remains under-characterized. We study repeated identical evaluations on 29 tasks spanning 10 categories using two OpenAI judge models (GPT-4o-mini and GPT-4.1-mini), with 50 pairwise trials and 50 pointwise trials per question, supplemented by temperature and prompt-sensitivity ablations. Across judges, pairwise preferences flip on average 13.6% of the time, with 28% of questions exceeding a 20% flip rate and one question reaching 56%. GPT-4o-mini also exhibits a significant first-position bias (72% A-majority, p = 0.024). At the same time, mean pointwise score gaps are small (0.19--0.36 on a 10-point scale) and not statistically significant in aggregate, producing a pairwise--pointwise gap: judges frequently choose a winner even when their own scalar scores provide little evidence of a meaningful quality difference. Beyond within-judge instability, cross-judge agreement is only 76% ($κ= 0.51$), semantically equivalent prompt templates change majority outcomes in 25% of tested cases, and deterministic decoding reduces but does not eliminate inconsistency. A reliability curve analysis shows that, in our dataset, 11 repeated trials are needed for a majority vote to recover the 50-trial reference verdict with 95% probability on average, rising to 15 for high-variance questions. These findings suggest that single-trial LLM judging is often too noisy for high-stakes evaluation, and that multi-trial aggregation, position randomization, and explicit uncertainty reporting should be standard practice. Because both judges are from a single provider, cross-provider replication remains an important next step.

This work explores the role of prompt design and judge selection in LLM-as-a-Judge evaluations of free text legal question answering. We examine whether automatic task prompt optimization improves over human-centered design, whether optimization effectiveness varies by judge feedback style, and whether optimized prompts transfer across judges. We systematically address these questions on the LEXam benchmark by optimizing task prompts using the ProTeGi method with feedback from two judges (Qwen3-32B, DeepSeek-V3) across four task models, and then testing cross-judge transfer. Automatic optimization consistently outperforms the baseline, with lenient judge feedback yielding higher and more consistent gains than strict judge feedback. Prompts optimized with lenient feedback transfer better to strict judges than the reverse direction. Analysis reveals that lenient judges provide permissive feedback, yielding prompts with broader applicability, whereas strict judges produce restrictive feedback, leading to judge-specific overfitting. Our findings demonstrate algorithmically optimizing prompts on training data can outperform human-centered prompt design and that judges' dispositions during optimization shape prompt generalizability.

Automatic systems are increasingly used to assess the originality of responses in creative tasks. They offer a potential solution to key limitations of human assessment (cost, fatigue, and subjectivity), but there is preliminary evidence of a self-preference bias. Accordingly, automatic systems tend to prefer outcomes that are more closely related to their style, rather than to the human one. In this paper, we investigated how Large Language Models (LLMs) align with human raters in assessing the originality of responses in a divergent thinking task. We analysed 4,813 responses to the Alternate Uses Task produced by higher and lower creative humans and ChatGPT-4o. Human raters were two university students who underwent intensive training. Machine raters were two specialised systems fine-tuned on AUT responses and corresponding human ratings (OCSAI and CLAUS) and ChatGPT-4o, which was prompted with the same instructions as human raters. Results confirmed the presence of a self-preference bias in LLMs. Automatic systems tended to privilege artificial responses. However, this self-preference bias disappeared when the analyses controlled for the idea elaboration. We discuss theoretical and methodological implications of these findings by highlighting future directions for research on creativity assessment.

Evaluating the writing capabilities of large language models (LLMs) remains a significant challenge due to the multidimensional nature of writing skills and the limitations of existing metrics. LLM's performance in thousand-words level and open-ended writing is inadequately assessed by traditional reference-based metrics or modern LLM-as-a-judge methods. We propose Tree-of-Writing (ToW), to resolve the implicit inconsistency often found when LLM-as-a-judge aggregates all sub-features in text evaluation. ToW incorporates a tree-structured workflow by explicitly modeling the aggregation weights of sub-features. We also present HowToBench, a large-scale Chinese writing benchmark encompassing 12 genres and 1302 instructions across three task categories: contextual completion, outline-guided writing, and open-ended generation. ToW successfully mitigates the biases, achieving a 0.93 Pearson correlation with human judgments. Furthermore, we detect that both overlap-based text generation metrics and popular LLM-as-a-judge practices are vulnerable to textual disturbances, while ToW is robust to them. We also uncover a negative correlation between input length and content-related scores in the Guide task, showcasing that it cannot be simply improved by input-side information piling.

We propose a scalable, multifactorial experimental framework that systematically probes LLM sensitivity to subtle semantic changes in pairwise document comparison. We analogize this as a needle-in-a-haystack problem: a single semantically altered sentence (the needle) is embedded within surrounding context (the hay), and we vary the perturbation type (negation, conjunction swap, named entity replacement), context type (original vs. topically unrelated), needle position, and document length across all combinations, testing seven LLMs on tens of thousands of document pairs. Our analysis reveals several striking findings. First, LLMs can exhibit a within-document positional bias distinct from previously studied candidate-order effects: some models penalize semantic differences more harshly when they occur earlier in a document. Second, when the altered sentence is surrounded by topically unrelated context, it raises similarity scores in almost all settings. This is consistent with a contradiction account, in which topically related context allows the alteration to conflict with the surrounding sentences. Third, all models exhibit the same hierarchy in how harshly they score different perturbation types, with negation penalized the most, but express this hierarchy on significantly different scales. Finally, varying the prompt's conception of semantic similarity across several alternatives changes the magnitude but not the direction of these effects. Together, these results demonstrate how LLM semantic similarity scores are sensitive to the aforementioned parameters in ways that go beyond the semantic change itself; our framework offers a practical toolkit for auditing and comparing scoring behavior across current and future models.

As reinforcement learning continues to scale the training of large language model-based agents, reliably verifying agent behaviors in complex environments has become increasingly challenging. Existing approaches rely on rule-based verifiers or LLM-as-a-Judge models, which struggle to generalize beyond narrow domains. Agent-as-a-Judge addresses this limitation by actively interacting with environments and tools to acquire verifiable evidence, yet its capabilities remain underexplored. We introduce a benchmark AJ-Bench to systematically evaluate Agent-as-a-Judge across three domains-search, data systems, and graphical user interfaces-comprising 155 tasks and 516 annotated trajectories. The benchmark comprehensively assesses judge agents' abilities in information acquisition, state verification, and process verification. Experiments demonstrate consistent performance gains over LLM-as-a-Judge baselines, while also revealing substantial open challenges in agent-based verification. Our data and code are available at https://aj-bench.github.io/.

Retrieval-augmented generation (RAG) enhances large language models (LLMs) with external knowledge to answer questions more accurately. However, research on evaluating RAG systems-particularly the retriever component-remains limited, as most existing work focuses on single-context retrieval rather than multi-hop queries, where individual contexts may appear irrelevant in isolation but are essential when combined. In this research, we use the HotPotQA, MuSiQue, and SQuAD datasets to simulate a RAG system and compare three LLM-as-judge evaluation strategies, including our proposed Context-Aware Retriever Evaluation (CARE). Our goal is to better understand how multi-hop reasoning can be most effectively evaluated in RAG systems. Experiments with LLMs from OpenAI, Meta, and Google demonstrate that CARE consistently outperforms existing methods for evaluating multi-hop reasoning in RAG systems. The performance gains are most pronounced in models with larger parameter counts and longer context windows, while single-hop queries show minimal sensitivity to context-aware evaluation. Overall, the results highlight the critical role of context-aware evaluation in improving the reliability and accuracy of retrieval-augmented generation systems, particularly in complex query scenarios. To ensure reproducibility, we provide the complete data of our experiments at https://github.com/lorenzbrehme/CARE.

Multimodal Large Language Models (MLLMs) have been increasingly used as automatic evaluators-a paradigm known as MLLM-as-a-Judge. However, their reliability and vulnerabilities to biases remain underexplored. We find that many MLLM judges fail to reliably integrate key visual or textual cues, yielding unreliable evaluations when evidence is missing or mismatched, and exhibiting instability under semantically irrelevant perturbations. To address this, we systematically define Compositional Bias in MLLM-as-a-Judge systems and introduce MM-JudgeBias, a benchmark for evaluating it. MM-JudgeBias introduces controlled perturbations across Query, Image, and Response, and evaluates model behavior via two complementary metrics: Bias-Deviation (BD) for sensitivity and Bias-Conformity (BC) for stability. Our dataset of over 1,800 curated and refined multimodal samples, drawn from 29 source benchmarks, enables a fine-grained diagnosis of nine bias types across diverse tasks and domains. Experiments on 26 state-of-the-art MLLMs reveal systematic modality neglect and asymmetric evaluation tendencies, underscoring the need for more reliable judges.

Decentralized LLM inference networks need lightweight, reference-free quality evaluation for Proof of Quality (PoQ). We present PoQ-Judge, a framework that trains dedicated judge models to score query-output pairs without ground-truth references. We study three architectures across the quality-cost tradeoff: a TextCNN judge, a MiniLM cross-encoder, and a DeBERTa judge. Using two-stage training on UltraFeedback plus GPT-labeled in-domain data, the best model reaches 0.747 Pearson correlation with the ground-truth proxy on a held-out test set, outperforming reference-based evaluators from prior work. As a reference-free component in composite scoring, it achieves 0.645 Pearson correlation, matching the best single reference-based evaluator while removing the need for reference answers. We also show that online calibration identifies semantic quality as the dominant dimension and that cascade evaluation reduces cost by 72.7 percent with only modest quality loss. Results are much stronger on QA than summarization, pointing to proxy quality as the main remaining limitation.

Instruction Following (IF) is a core capability of LLMs, requiring strict adherence to diverse constraints, ranging from verifiable ones (e.g., output length) to unverifiable ones (e.g., tone). Reinforcement learning with verifiable rewards has emerged as a paradigm for IF tasks, leveraging LLM-as-a-judge to assess unverifiable constraints. However, we empirically find that this approach remains a significant bottleneck, suffering from severe reward hacking and higher computational overhead. In this work, we first analyze the generalization capabilities of unverifiable constraints and discover that specific constraints exhibit distinct, high-generalization patterns. Motivated by this, we propose TinyJudge, a framework that employs an ensemble of specialized tiny language models ($\sim0.6B$) to provide rewards for soft constraints. By distilling expertise from frontier models into these tiny models, it achieves high-precision, lightweight evaluation. Extensive evaluations across five benchmarks demonstrate that TinyJudge outperforms the baselines by $\sim10\%$ in average performance and $12\%$ in reward precision. Crucially, it also achieves a $3\times$ speedup in total training time. Our work provides a scalable and robust path for aligning LLMs with unverifiable human instructions.

Large Language Models are increasingly used as judges to evaluate code artifacts when exhaustive human review or executable test coverage is unavailable. LLM-judge is increasingly relevant in agentic software engineering workflows, where it can help rank candidate solutions and guide patch selection. While attractive for scale, current practice lacks a principled account of reliability and bias: repeated evaluations of the same case can disagree; small prompt edits can swing outcomes; and seemingly semantics-preserving, human-equivalent perturbations may elicit divergent verdicts. This paper studies LLM-as-a-Judge for code through a measurement-first lens. We analyze two pointwise judging regimes across code generation, code repair task, and test generation, and we systematically probe prompt-induced biases. Our study considers difficulty levels for repeated runs and controlled prompt interventions that isolate one presentation cue at a time, and it evaluates judges using consistency and sensitivity to bias. We find that judge decisions are highly sensitive to prompt biases even when the underlying code snippet is unchanged. Across all three tasks, several biases systematically shift preferences toward the option favored by the prompt, improving accuracy when that option aligns with the gold answer but substantially reducing it otherwise. In some settings, these effects are large enough to change task-level conclusions and alter relative model rankings. These findings show that reported judge performance may reflect prompt artifacts rather than stable assessment ability, posing a direct threat to the validity and reproducibility of code evaluation. We therefore argue that LLM-as-a-Judge studies should report bias sensitivity alongside accuracy and incorporate explicit controls to support more trustworthy model comparison in software engineering.

Automated evaluation of tool-using large language model (LLM) agents is widely assumed to be reliable, but this assumption has rarely been validated against human annotation. We introduce AgentProp-Bench, a 2,000-task benchmark with 2,300 traces across four domains, nine production LLMs, and a 100-label human-validated subset. We quantify judge reliability, characterize error propagation, and evaluate a runtime mitigation. Substring-based judging agrees with human annotation at kappa=0.049 (chance-level); a three-LLM ensemble reaches kappa=0.432 (moderate) with a conservative bias. Under validated evaluation, a parameter-level injection propagates to a wrong final answer with human-calibrated probability approximately 0.62 (range 0.46-0.73 across models). Rejection (catching bad parameters) and recovery (correcting after acceptance) are independent model capabilities (Spearman rho=0.126, p=0.747). A tuned runtime interceptor reduces hallucination on GPT-4o-mini by 23.0 percentage points under a concurrent n=600 control, but shows no significant effect on Gemini-2.0-Flash, whose aggressive parameter rejection eliminates the target failure mode. All code, data, traces, and human labels are released at https://github.com/bhaskargurram-ai/agenthallu-bench.