LLM-as-a-Judge

Evaluating large language models (LLMs) is important for understanding their capabilities, comparing competing systems, and supporting the deployment of reliable models in practice. For open-ended tasks, pairwise evaluation has become a popular paradigm, in which two responses to the same prompt are compared and the resulting judgments are aggregated into an overall ranking. A central challenge of this paradigm is intransitivity: the induced comparison outcomes may fail to support any coherent global ranking. For example, one may observe cyclic preferences such as $A \succ B \succ C \succ A$, or inconsistencies involving ties such as $A \equiv B\equiv C\neq A$. Such contradictions make the resulting leaderboard unstable and challenging to interpret. In this paper, we propose a prompt perturbation framework for improving the consistency of pairwise LLM evaluation. Our approach generates perturbed variants of each prompt, uses the resulting comparison graphs to identify and filter out structurally inconsistent comparison patterns, and then applies standard ranking methods to the filtered comparisons. A key feature of the proposed framework is that graph-level structural consistency is incorporated explicitly into the evaluation pipeline before ranking aggregation. This provides a simple and principled way to reduce cyclic inconsistencies and improve the reliability of LLM rankings.

Generative AI and large language models (LLMs) are increasingly applied to question generation and automated assessment. However, deploying LLMs in preparation for high-stakes exams requires more than prompt engineering; it demands software pipelines that systematically ground model outputs in authorised curriculum artefacts and marking guidelines issued by education authorities. This paper presents a curriculum-grounded, configurable LLM-as-Judge pipeline for question-level marking, co-developed with an industrial partner, to support exam preparation for university admission. The pipeline identifies the relevant topics, subtopics, and cognitive demand of a question, and assembles verifiable and authorised context to support LLM judgement. Curriculum intent is operationalised through concrete syllabus artefacts, including prescribed verbs and outcomes, performance band descriptors, glossary definitions, and marking-guideline principles. A staged LLM workflow is employed to first generate question-specific rubrics, capturing structured expectations of performance, and then derive and evaluate marking criteria used to allocate marks to student responses. This design improves consistency, transparency, and alignment with official marking practices. Preliminary evaluation shows that the proposed LLM-as-Judge pipeline delivers marking outcomes comparable to human tutors, while yielding justifications that are more traceable to authorised curriculum artefacts and marking standards. The pipeline has also been integrated into an online study platform, where early deployment data provide initial insights into operational usage and manual overrides.

Large language models (LLMs) are increasingly considered for use in clinical consultation tasks, yet most medical evaluations remain static, single-turn, or narrowly outcome-based, limiting their ability to reflect the sequential, uncertain, and interactive nature of real-world care. Here, we propose AIPatient Arena, an EHRs-grounded evaluation framework for assessing the clinical utility of LLMs across eight dimensions of clinical competence. The framework integrates EHR data into patient-specific knowledge graphs, enabling multi-turn physician-patient interactions. We applied AIPatient Arena on a primary cohort of 437 patients and two out-of-distribution validation cohorts of 119 and 67 patients. We observe that LLMs performed well in medical interview questioning skills (QS; mean scores, 4.43-4.99/5), ethical and professional conduct (ET; 4.38-4.93/5), and clarity and transparency of clinical explanations (EX; 3.80-4.72/5). Performance was moderate in information integration (II; 3.19-4.21/5) and medication safety and justification (MS; 3.13-3.78/5), but persistent weaknesses were observed in handling of ambiguous patient responses (HR; 2.57-3.32/5), information coverage (IC; 2.08-3.02/5), and diagnostic accuracy and reasoning (Dx; 2.63-3.55/5). Process-based evaluation revealed recurrent interaction failures, including repetitive questioning, omission of past medical history, and inadequate handling of uncertainty. Richer conversational context improved diagnostic reasoning but yielded limited gains in treatment planning. These findings indicate that final-answer accuracy alone is insufficient for evaluating clinical readiness and highlight the importance of assessing how models gather, interpret, and communicate information throughout a consultation. AIPatient Arena provides an EHR-grounded framework for workflow-oriented pre-deployment evaluation of medical LLMs.

This study proposes a domain-specific LLM-based Visual Explanation Evaluation Framework for assessing Grad-CAM explanations in facial skin disease diagnosis models. While previous studies have primarily focused on improving classification performance through data augmentation techniques, relatively few studies have systematically examined whether model explanations are grounded in clinically relevant lesion regions. In this study, geometric augmentation, color-based augmentation, and mixed augmentation strategies were applied to facial skin disease classification models based on EfficientNet-B0, MobileNetV3, and ResNet18. Grad-CAM was employed to generate visual explanations representing the models' decision-making processes. Furthermore, an LLM-as-a-Judge evaluation framework was designed using GPT-5.5, Gemini 3.5 Flash, and Claude Sonnet 4.6 to assess Grad-CAM explanations from the perspectives of lesion localization and explanation trustworthiness. To improve evaluation consistency and clinical grounding, a progressive prompt engineering strategy was introduced, incorporating evaluation rubrics, clinical knowledge, penalty rules, and structured output formats.

Recent work motivates moving large language model (LLM) evaluation from mean-based to tail-aware metrics, including conditional value-at-risk and tail-index estimates of reward-model error. We ask whether the canonical extreme-value-theory tail-index parameter, which isolates how heavy a tail is from how large the tail mass is, adds discriminative information beyond the mean and a standard tail-magnitude statistic in LLM evaluation. We pre-register a protocol covering admissibility, goodness-of-fit, threshold-stability, and effect-size requirements for any positive tail-shape claim. The protocol is the contribution of this paper; the empirical study below is a demonstration of what its gates catch. Applied to a standard LLM toxicity-evaluation setup under two structurally different scorer families, the protocol catches three distinct modes of false positives that a naive analysis would have published, and rejects the headline tail-shape claim on both scorers. We conclude that tail-shape estimation in the LLM toxicity-evaluation setups we examined is more fragile than the recent literature suggests, and recommend the protocol as a starting point for tail-index claims in similar setups.

Current evaluation paradigms for Large Language Model (LLM) personalization rely heavily on brittle surface-matching metrics or computationally expensive LLM-as-a-judge protocols, both of which lack interpretability. To address these limitations, we introduce Natural Language Inference Constraint Verification (NLICV), a scalable, semantically invariant framework that maps sentence meanings to truth-condition sets to verify personalization constraints via a Natural Language Inference (NLI) model. Moving beyond binary scoring, NLICV categorizes LLM behaviors into four distinct modes: personalization, generalization, sycophancy, and failure. Extensive experiments demonstrate that NLICV aligns closely with human annotations while drastically reducing the latency and token costs associated with LLM judges (up to 2100 inference speedup). Finally, through an ablation-based procedure, NLICV pinpoints the exact sentences driving the constraint verification, yielding faithful, understandable evidence for its evaluations.

Large language models (LLMs) are increasingly deployed as UX judges that inspect interfaces, diagnose usability problems, and propose repairs. Yet no controlled benchmark measures whether the resulting critiques are reliable and actionable across heterogeneous product surfaces. We introduce UXBench, a benchmark for evaluating LLMs as interaction-grounded UX judges. UXBench comprises local-first runnable web fixtures spanning ten product-surface families, paired with coverage-gated browser exploration that forces models to collect interaction evidence before reporting. Each judge model produces a structured UX report over seven rubric dimensions; report quality is measured by whether a fixed downstream repair agent can improve the interface based on the critique. We evaluate eight frontier models under both an automated repair-lift protocol and a blind human validation study. Results show that UX judging is neither saturated nor one dimensional: models differ meaningfully in report actionability, exhibit distinct rubric-level repair signatures, vary in fixture-level reliability, and trade leadership across surface categories

High-quality time series forecasting is pivotal for real-world decision-making. However, traditional point-wise metrics often fail to reveal complex temporal patterns and align poorly with human intuitive preferences. While the ''LLM-as-a-Judge'' paradigm has revolutionized text evaluation by providing flexible, human-aligned judgment, its application to time series remains largely unexplored. In this paper, we leverage Vision-Language Models (VLMs) as judges for time series forecasting, harnessing their ability to comprehend time series plots grounded in textual information. Specifically, we propose a novel framework integrating micro- and macro-level judgments informed by contextual information to evaluate time series forecasting. To this end, we introduce TimeVista, a comprehensive VLM-as-a-Judge benchmark comprising 5563 time series samples paired with detailed evaluation rubrics. Extensive meta-evaluations demonstrate that VLMs are highly reliable judges, achieving significantly higher consistency with human preferences than conventional metrics. Building upon our benchmark, we comprehensively assess recent Time Series Foundation Models (TSFMs) under the VLM-as-a-Judge paradigm. Our results demonstrate that VLMs serve as robust and interpretable judges, providing a comprehensive, human-aligned standard for evaluating time series models.

Open-source LLM agent ecosystems are growing rapidly, yet the security of community-contributed skills - modular tool definitions that extend agent capabilities - remains largely unvetted. The gap we fill: existing scanners operate at the code layer and are structurally blind to instruction-layer and multi-agent risk - natural-language directives that hijack an agent, exfiltrate data through encoded side channels, or chain harm across pipelines - so what is needed is a semantic, multi-dimensional vetting system rather than another signature matcher. We present SKILLVETBENCH, a live public leaderboard on Hugging Face that uses an LLM-as-Judge to vet agent skills. What is new: SARS (Skill Agentic Risk Score), a five-dimensional agentic-risk metric with a principled weighted formula for instruction-following systems. What is integrated: full CVSS v4.0 vector decomposition and a ClawHub dual-view that places our LLM-generated review beside the official marketplace verdict. What is demonstrated: drawing on our companion benchmark paper [ 1], the LLM-as-Judge stage achieves zero false negatives across 78 confirmed-malicious skills and zero false positives across 22 benign controls, while the best static baseline (SKILLSIEVE) still misses 15%; for instruction-layer categories such as Prompt Injection and Memory Poisoning, conventional tools miss between 89% and 100% of threats (e.g., CODEBERT detects none of nine memory-poisoning skills). Detection rates vary from 35% to 95% across four LLM evaluators, motivating ensemble scoring in production deployments.

LLM-as-a-judge systems are now routinely used for open-ended model evaluation, where human preference annotation is costly, slow, and difficult to reproduce. Yet these judges are often reported as scalar accuracy, win-rate, or agreement devices. We argue that a judge should instead be reported as a measurement instrument. We introduce a Judge Datasheet protocol that measures dark current under true-vacuum inputs, stable cross-sensitivity to same-quality surface variation, positional false preference, target sensitivity on a controlled quality ladder, and the criterion or operating point induced by tie instructions. The direction-stability decomposition reveals that apparent Delta0 preference can be stable surface response or disguised position bias. In a three-judge open-weight case study, Llama-3.1-8B shows high dark current and presentation-conflicted Delta0 behavior, Qwen2.5-14B is vacuum-clean and target-sensitive but mixes stable and positional over-discrimination, and Qwen2.5-32B is vacuum-clean with low stable cross-sensitivity and low positional false preference. A strict tie criterion eliminates Qwen32B Delta0 false preference but absorbs marginal Delta1 target signals into ties while preserving Delta5 sensitivity. The results show that prompting moves the criterion, not the resolution. We do not claim that the downstream mechanism hypothesis that motivated this work is confirmed; the contribution is a metrological protocol for measuring the measuring device before downstream claims are made.

Multimodal Large Language Models (MLLMs) are increasingly used as automated judges, e.g., for image quality and safety assessment. However, their adversarial robustness remains largely unexplored, threatening the fairness and reliability of automated judging. To bridge this gap, we introduce RobustMLLMJudge, the first general framework for evaluating the adversarial robustness of general-purpose MLLMs when functioning as judges. It covers diverse attacks against popular judge approaches across quality and safety evaluation scenarios. Using RobustMLLMJudge, we reveal that i) different MLLM judges are highly vulnerable to score-inflating adversarial attacks; and ii) although effective, these attack methods face a critical challenge due to unique constraints in the evaluation protocols of MLLM judges. We further propose MGSIA, namely Manifold-Guided Semantic Induction Attack, a novel method that bypasses these constraints to enable more effective and transferable attacks on MLLM judges. The core idea of MGSIA is to combine affirmative semantic induction with high-score manifold alignment: it maximizes the probability that judges yield affirmative responses (e.g., "Yes") to binary semantic queries, while regularizing adversarial representations toward high-score centers estimated from proxy protocols. Together, these objectives yield transferable score-inflating perturbations. Extensive experiments demonstrate the superiority and generalizability of MGSIA in deceiving advanced MLLM judges under different evaluation scenarios, highlighting the need for robust MLLM judges. Code and data will be made available at https://github.com/mala-lab/RobustMLLMJudge.

Continuous evaluation of LLM products relies on a strong LLM judge treated as ground truth: a cheap monitor scores every interaction and a team is paged when the score drifts down. But the judge is itself a model behind an API, and a silent version bump or scoring-prompt update changes how it scores -- so every drift alarm is ambiguous between a worse product and a changed judge. We resolve the ambiguity with a fixed, human-labeled anchor set that the current judge re-scores at a steady interleave, a second betting e-process on the judge-versus-human gap, and a guard-window rule returning a verdict in {none, system, judge}. We prove anytime-validity, one-way identification (only the judge can move the anchors), an attribution race whose design law is that the anchors must out-run the main process they guard, and process orthogonality. On two real judge changes, a silent version bump is detected as judge drift in 60/60 runs with zero judge-to-system misattribution, and a contaminating strict-prompt change is correctly attributed on 110 of 120 runs at guard width 300 -- while the industry-default rolling z-test false-alarms on 75% of drift-free streams. Every experiment replicates on a second domain (TL;DR summarization) with nothing re-tuned, and where the domains differ the differences are the ones the race predicts: the strict-prompt change shifts scores harder there, so the anchors fire faster and attribution becomes perfect (240/240). The monitor runs at approximately 0.64 of the cost of strong-judging every item, or 0.21 in a cheaper-but-deafer regime.

Objective: To enhance the accuracy, interpretability, and robustness of large language models (LLMs) in medical question answering (MedQA). Method: We designed a multi-agent peer-reviewed reasoning method in which multiple LLM agents independently generate chain-of-thought reasoning with candidate answers, then act as peer reviewers to evaluate each other's reasoning for factual correctness and logical soundness. The highest-rated reasoning chain is selected to produce the final answer. Experiments were conducted with five state-of-the-art LLMs (Llama-3.1-8B, Qwen2.5-7B, Phi-4, DeepSeek-LLM-7B, GPT-oss-20B) on three benchmark datasets: HeadQA, MedQA-USMLE, and PubMedQA. Performance was compared against single-model chain-of-thought reasoning and chain-of-thought-based majority voting. Results: Peer-reviewed reasoning consistently outperformed both baselines. The best model combination achieved an average accuracy of 0.820 across datasets, exceeding the strongest single model (0.777) and majority voting ensembles (up to 0.789). The method also scaled effectively with more participating models, while peer assessments reliably distinguished high- from low-quality reasoning chains. Conclusion: The proposed multi-agent peer-reviewed reasoning method enables LLMs to act as both solvers and evaluators, yielding superior performance in MedQA. By emphasizing reasoning quality rather than answer agreement alone, this approach improves accuracy, interpretability, and robustness, offering a promising direction for trustworthy biomedical AI systems.

Large language models (LLMs) are increasingly used for technical document generation, yet single-model outputs often suffer from over-engineering, security blind spots, and incomplete coverage. We propose TriAdReview, a triangular adversarial review architecture that employs two independent reviewer models (engineering and boundary perspectives) and a triangular judging mechanism to iteratively improve a generator model's output. We evaluate TriAdReview across five benchmark tasks - architecture design, code generation, proposal review, security audit, and requirements analysis - using three configurations: single model (baseline), dual model (single review), and triple model (full system). Results across 75 experiments (n=5 per cell) show that the triple model configuration achieves a 10.1% overall improvement over the single model baseline (26.2 vs. 23.8 out of 50; p<0.05, paired t-test), with particularly strong gains on security audit (+27.6%), code generation (+20.8%), and architecture design (+15.6%). A second scorer (mimo-v2.5-pro) confirms the direction with a smaller effect (+2.7%), suggesting moderate inter-rater agreement. However, the system shows a -7.5% degradation on requirements analysis, revealing that adversarial review architectures have a structural bias toward simplification that is counterproductive for completeness-oriented tasks. We analyze this boundary condition through a task-type framework and demonstrate that reviewer prompt adaptation partially mitigates the issue. Our findings provide the first empirical characterization of when multi-model adversarial review helps versus harms, with implications for the design of collaborative AI systems.

LLM judges are used to reduce the need for costly human labor in evaluating open-ended text generation. However, the reliability of these judges depends critically on their alignment with human raters -- a property that itself depends on costly human annotations. In this work, we develop a method (Metric Match) for estimating correlation-based reliability metrics of LLM judges from limited annotations. Metric Match selects a subset of samples for human annotation such that the subset matches the population reliability metric with respect to acquired synthetic labels. We empirically show that Metric Match achieves a win-rate of 0.838 against random subset selection across four different correlation metrics and 15 datasets, with an 18.7% decrease in average estimation error and reduces annotation needs by 32.5%. We provide a cost model and highlight a medical case study where our method saves $1,041.67 compared to random selection for expert annotation. Further, we shift our task from reliability estimation to reliability classification of whether a given judge is above a deployment threshold, outperforming random selection with Metric Match. All project code is publicly available, and we additionally provide an installable package for ease of use.

LLMs have substantially improved software engineering yet real-world development requires architectural understanding. Such understanding is prohibitively expensive to label manually and impossible to verify through tests alone. We propose an agentic judging pipeline using a strong LLM as a scalable proxy for expert architectural evaluation, comprising two judges: the Architecture Complexity Judge (ACJ), which estimates codebase-specific architectural understanding a task demands, and the Architecture Quality Judge (AQJ), which evaluates patch conformance to repository-specific architectural conventions via source-grounded rubrics. Fine-tuning Qwen3-8B/14B/32B on 3,360 curated instances achieves resolved rates of up to 27.2% on SWE-bench Verified - up to 540% over the base model and 256% over unfiltered fine-tuning. Meanwhile, the trained models achieve strong cross-language generalization and consistent improvements in architectural patch quality.

MLLM-as-a-Judge is conventionally validated by agreement with human annotations, but this metric is undefined when the human pool is culturally heterogeneous. We introduce VOIR DIRE, a multimodal benchmark of 626 culturally paired image--prompt artifacts spanning U.S. and mainland Chinese contexts across food, fashion, and architecture, with annotator pools that are within-pool reliable (a = 0.86/0.74) but cross-pool divergent on evaluation (Q1 r = -0.12). Across six MLLMs, the bias decomposes into two failures: a positivity-floor calibration failure (compressed scale use) and an orientation failure (default to one cultural norm). On this corpus, where contested items are sampled to split the two pools, the floor mechanically validates the more-permissive Chinese reading; persona prompting partially recovers calibration, but the orientation residual survives, evidence the tilt is not reducible to scale compression. Reference-pool in-context demonstrations deepen the orientation residual and inflate the high end rather than restoring use of the low end. Model origin adds a small additive tilt (~0.10 MAE) that is approximately invariant under demonstration. We recommend reporting alignment against each reference pool separately and treating cross-pool divergence as a judge property.

This work presents the design, implementation, and evaluation of a system for generating personalized reading content using Large Language Models (LLMs) combined with Retrieval-Augmented Generation (RAG). The proposed architecture consists of four modules: Input, RAG, Generation, and Judging and enables users to specify both a question and a target reading content complexity. RAG is employed to retrieve relevant information from the Internet, enriching and grounding the content produced by three modern LLMs: Meta LLaMA 4 Scout, LLaMA 3.1 8B Instant, and Google Gemma2 9B. Reading materials are generated using three prompting strategies (Chain-of-Thought, zero-shot, and few-shot), and the LLM-as-a-Judge module automatically evaluates answer quality and alignment with the desired readability level. Experimental results show that RAG consistently improves system performance across all models and prompting techniques, increasing relevance and particularly groundedness by up to 26-35 percentage points. Overall, the findings demonstrate that the RAG-augmented architecture effectively produces reading content tailored to user queries and desired textual complexity.

Large language models (LLMs) are now widely used as automatic judges for open-ended instruction-following evaluation. This practice is convenient, scalable, and often more semantically aware than reference-based metrics, but it also introduces a new reliability question: does a judge evaluate the quality of an answer, or does it also react to the language in which the comparison is presented? We propose Judge-LS, a lightweight meta-evaluation protocol that transforms LLMBar response-pair items into English, Chinese, and Chinese-English language-switched variants. A reliable judge should preserve its preference under label-preserving language transformations and should not prefer a language when two answers are translation-equivalent. We evaluate four API-accessible judges on the full 419-item LLMBar benchmark, producing 13,408 successful pairwise judgments. Across models, Chinese and language-switched presentations induce 10.7--14.4% preference flips relative to English, and all judges achieve their highest accuracy in English. However, translation-equivalent tie probes do not reveal a systematic English preference: most probes are judged as ties, and non-tie decisions more often favor Chinese. We add confidence intervals, paired significance tests, and an automatic transformation audit with a sensitivity analysis that excludes mechanically flagged high-risk variants. The experiment requires no model training, uses only API calls, and is feasible on modest local hardware.

Large Language Models (LLMs) have made notable progress in automated theorem proving, yet existing formal benchmarks remain limited in both mathematical coverage and difficulty. Most are concentrated in areas that are easier to formalize, such as algebra and elementary number theory, and provide limited coverage of subfields that require deeper reasoning, including mathematical analysis. To address this gap, we introduce MA-ProofBench, to the best of our knowledge, the first formal theorem-proving benchmark dedicated to Mathematical Analysis. The benchmark contains 200 formalized theorems covering 6 core topics and 27 subcategories, including measure and integration theory, complex analysis, and functional analysis. The problems are divided into two difficulty levels, an undergraduate level (Level I, 100 problems) and a Ph.D. qualifying level (Level II, 100 problems), to evaluate how well LLMs perform formal reasoning at different mathematical depths. Each problem is constructed through a human-led, LLM-assisted formalization pipeline followed by independent expert review, ensuring that the formal statements remain faithful to the original mathematics. We evaluate a range of recent general-purpose reasoning models and formal theorem provers on MA-ProofBench. However, most models perform poorly: even the best-performing model, GPT-5.5, achieves only 16% Pass@8 on Level I and 5% on Level II, while most models stay close to 0% on Level II. Further analysis identifies Mathlib hallucinations and incomplete proofs as the two dominant failure modes, while an evaluation on the natural-language version of the benchmark exposes a clear gap between informal and formal reasoning. MA-ProofBench is intended to serve as a reliable reference for tracking progress in formal mathematical reasoning in advanced domains.

Supramolecular chemistry, which includes the study of non-covalent host-guest assemblies, has advanced various applications. However, designing host-guest systems remains time-consuming, requiring days of dry-lab verification per candidate pair. Although LLMs have emerged as a fast alternative with strong performance on molecular binding tasks, no benchmark currently systematically evaluates LLMs for host-guest reasoning across fundamental supramolecular chemistry tasks, e.g., binding affinity prediction. To this end, we collaborate with domain experts to release the first Supramolecular Benchmark, called SupraBench, to evaluate LLMs in chemistry reasoning. Specifically, we design four fundamental tasks, i.e., binding affinity prediction, top-binder selection, solvent identification, and host-guest description, plus an auxiliary vision-based task for molecular identification. We also release SupraPMC, a curated 16M-token corpus of Supramolecular chemistry articles distilled from Europe PMC, to support the adaptation to the supramolecular domain. We benchmark a broad range of open and proprietary LLMs and find that LLMs leave substantial headroom across all tasks. Domain adaptation pretraining over SupraPMC transfers cleanly to in-distribution regression but trades off against strict letter-format output. Moreover, the difficulty profile differs sharply across task families, revealing distinct failure modes that indicate specific gaps in current supramolecular chemistry reasoning. Our source codes and benchmark datasets are available at https://github.com/Tianyi-Billy-Ma/SupraBench.

This study examines training dynamics in a small Llama-style language model trained under a fixed, compute-constrained token budget. Rather than evaluating efficiency solely through endpoint performance, the study uses a quantitative experimental repeated measures design to analyze how validation loss, validation perplexity, rolling volatility, backslide behavior, spike behavior, and between-seed variability change across token-based training intervals. Six independent training runs were conducted on a 4.26-million-parameter model using the TinyStories corpus, CPU-based full-precision training, and a target budget of approximately 20 million cumulative training tokens. Metrics were collected across 21 intervals, producing 126 seed-by-interval observations. Repeated measures ANOVA showed statistically significant interval effects for validation loss, validation perplexity, and rolling volatility. Descriptive trajectories revealed rapid early improvement followed by non-monotonic degradation during later training intervals. Mean validation loss decreased from 8.3552 at initialization to 2.7996 near 4 million tokens, but increased to 3.9010 by the final checkpoint. Validation perplexity followed the same pattern, falling sharply early in training before rising later. Derived telemetry further showed recurrent validation-loss backslides and no interval-summary evidence of a stable phase under the predefined criteria. These findings suggest that compute-aware language model evaluation should examine training trajectories rather than endpoint metrics alone. In constrained compute settings, additional token exposure may increase computational cost without producing proportional generalization gains, and interval-level telemetry can reveal instability, regression, and diminishing returns that final metrics may obscure.

Large Language Models (LLM) have shown impressive capabilities in Register Transfer Level (RTL) code generation, particularly for Verilog. However, evaluating their performance with other Hardware Description Languages (HDL), especially VHDL, remains limited although its distinct language characteristics, such as stricter semantic rules, introduce evaluation considerations that differ from Verilog. This lack of coverage restricts fully understanding of how well current models generalize across hardware design languages with differing structures and semantics. To address this gap, we introduce VHDLSuite, a benchmark-centered infrastructure for scalable VHDL generation evaluation, integrating automated benchmark synthesis, executable validation, and multi-model diagnostic analysis. First, we propose a data pipeline that automatically converts Verilog designs and their accompanying testbenches into executable VHDL benchmark instances, followed by VUnit/GHDL-based validation to ensure each released task is compilable, runnable, and consistently checkable in the VHDL environment. Second, we introduce VHDLBench, a benchmark with over 200 VHDL problems with complete and validated testbenches across a wide range of complexity levels. Third, we extensively evaluate cutting-edge LLMs and uncover key challenges specific on LLM-aided VHDL generation. Our findings provide important insights and support future work in multi-language hardware design automation.Our data pipeline, benchmark, and evaluation framework will be open-sourced.

Evaluating new large language models typically requires costly human annotation campaigns at scale. LLM-as-a-judge offers a cheaper alternative, but judge scores carry systematic errors - such as position bias, self-preference, or intransitivity - that can strongly miscalibrate the resulting rankings. We quantify the resulting judge-human disagreement at two complementary levels. At the local level, we estimate per-battle uncertainty from the judge's own score differences by propagating calibrated win probabilities rather than hard labels into the Bradley-Terry procedure. This alone provides a drastic improvement to Elo estimation accuracy, bringing LLM-derived ratings within 17.9 Elo MAE of human-derived ones when averaged over 55 held-out models on LMArena. At the global level, we apply split conformal prediction to the residual gap between LLM-derived and human-derived Elo ratings across held-out models, producing prediction intervals with distribution-free marginal coverage guarantees that account for irreducible LLM-human disagreement. Together, these two layers yield a low-cost evaluation tool that provides developers with calibrated Elo estimates and honest uncertainty bounds, without access to large-scale human annotations. To facilitate reproducibility, we release our code at https://github.com/kargibora/SoftElo .

We present MÖVE (Modelle für die Öffentliche Verwaltung Evaluieren), a holistic benchmark for evaluating large language models (LLMs) in the context of the German public sector. While LLMs are increasingly adopted in public administration, model selection remains largely ad hoc, and existing benchmarks offer limited guidance: they are predominantly English-centric, US-centric in content, and focus exclusively on task performance. MÖVE addresses these gaps by evaluating 39 models across two complementary dimensions. Performance criteria cover summarization, question answering, and topic extraction. Governance criteria assess hallucination tendencies, energy consumption, provider transparency, and alignment with German constitutional values and knowledge about positions by German political parties. In total, we utilize ten German-language datasets, including gold- and silverstandard datasets that we constructed to reflect public-administration domains. We employ a multi-metric evaluation strategy combining classical NLP metrics, embedding-based methods, and LLM-as-a-judge approaches. Our results show that no single model dominates across all criteria: top performers differ between tasks, and model size alone is a poor predictor of quality. We further evaluate the benchmark itself, analyzing its statistical precision, LLM judge reliability, the impact of our private datasets on model rankings, the sensitivity of our results to prompt formulation, and the validity of our energy consumption estimates. MÖVE is designed as a living benchmark under active development; results are publicly available at https://moeve.bundesdruckerei.de/.

Image-based AI assistants are now deployed at production scale on e-commerce platforms, where a single uploaded image can trigger fundamentally different user intents: product search, style recommendation, visual encyclopedia, or utility tool calls, each demanding its own response format, tool invocation, and domain knowledge. Without per-intent behavioral constraints, LLM-based systems conflate these heterogeneous modes and fall short of domain quality standards, while the breadth and dynamism of the intent space render manual engineering infeasible. To address this, we present SkillChain, which closes the production feedback loop on Skill evolution, automating the lifecycle of Skills through three stages: Skill Creator for bootstrapping from task specs and trajectories, Route Optimizer for routing alignment, and Body Refiner for iterative Skill Body refinement via dual-path LLM-Judge evaluation. Deployed on a production-scale e-commerce image assistant, SkillChain substantially improves aggregate response quality, with the strongest gains on structural compliance and content quality; a one-week online A/B experiment further confirms significant gains in user engagement, content consumption, and long-term retention.

Automated testbench generation has become a critical bottleneck in large language model (LLM)-driven Register Transfer Level (RTL) workflows, where large numbers of candidate designs must be verified rapidly and reliably. Existing prompt-based approaches treat testbench generation as unconstrained code synthesis, yielding stochastic outputs with high token cost, low reproducibility, and insufficient coverage. To address this gap, we present STG, a Structured Testbench Generation framework that exploits the inherent structure of hardware designs to generate deterministic testbenches. As a direct verification tool, STG runs 720x faster than an iterative LLM-based testbench generation flow and higher rate of successful compilation, achieves higher coverage, and reduces false-pass verdicts on incorrect DUTs. STG also helps identify errors in RTL generation benchmarks by exposing faulty benchmark testbenches. As a data curation engine, it is 11x faster than LLM-based filtering on a single CPU core with 127x less energy, and the resulting distilled models provide state-of-the-art performance in our multi-benchmark evaluation. As a test-time scaling oracle, it reduces node count by 14-47%. Our models are available at https://huggingface.co/collections/AS-SiliconMind/siliconmind-v12.

Despite strong performance in competitive programming, the role of Large Language Models (LLMs) in supporting human learning in the same setting remains largely unexplored. In this work, we introduce UOJ-Bench, a benchmark designed to evaluate not only the problem-solving ability of LLMs, but also their ability to identify errors in human-written code -- a crucial educational activity traditionally supported by running test cases over online judge systems. UOJ-Bench consists of three distinct tasks: code generation, code hacking, and code repair, all constructed from real-world code submissions on the Universal Online Judge (UOJ) and evaluated through UOJ's native judging infrastructure. Our results show that under one-shot evaluation, even the strongest models fail to identify errors in more than 50% of a set of submissions that have been found to be incorrect by UOJ users. While test-time scaling improves success rates to above 90%, the substantial computational costs incurred from model inference limit its practicality for large-scale deployment. Despite these limitations, we find that the best-performing models under test-time scaling can uncover errors in over 5% of full-score submissions across roughly 30 problems, suggesting that frontier LLMs can already provide complementary signals beyond standard judging systems.

Irrelevant numbers in a prompt can shift language model judgments, producing anchoring effects in numerical reasoning. We study where this anchor-sensitive signal is carried inside language models using a controlled multiple-choice setup with shared answer options. We define a logit-difference metric comparing the correct answer option with the answer option corresponding to the anchor, and validate that it tracks behavioral anchoring. Using attribution-based circuit localization on 7B--8B Qwen and Llama base and instruction-tuned models, we find that edge-level methods recover this signal more faithfully than node-level methods. Low- and high-anchor circuits transfer strongly within a model, suggesting shared pathway structure across anchor direction. However, sparse transfer across base and instruction-tuned variants is less reliable, indicating that post-training changes which pathways matter most. Overall, our results provide a mechanistic account of how anchoring-related decision signals are carried inside language models.

Large language models (LLMs) now reach expert-level scores on medical licensing exams, encouraging the assumption that high scores imply safe medical judgment while patients increasingly use them for health advice. We show this assumption is fragile: when misleading context is injected into questions that LLMs originally answer correctly, they abandon the correct answer. We call the ability to maintain correct judgment under adversarial context epistemic resilience, and introduce MedMisBench to measure it. MedMisBench contains 10,932 medical question items and 48,889 misleading context-option pairs spanning medical reasoning, agentic capability, and patient-journey evaluation. Across 11 model configurations, mean accuracy falls from 71.1% on original questions to 38.0% under focused misleading context, with 51.5% attack success. The most damaging injections are formal, rule-like fabrications: authority-framed falsehoods reach 69.5% attack success and exception-poisoning claims reach 64.1%. A 14-member clinical panel from 7 countries identified serious potential harm in 38.2% of reviewed cases. MedMisBench exposes a structural blind spot in LLM evaluation in medical settings: existing benchmarks measure what models know, but not whether they preserve correct medical judgment under misleading context.

Benchmark scores often misrepresent a large language model's (LLM's) knowledge, because they rely, e.g., on the model's ability to follow specific formatting requirements. This especially penalizes base models that may know the correct answers but lack the ability -- typically introduced in post-training -- to structure them as instructed. To overcome this, we propose soft-prompt tuning, an efficient, fair, and architecture-agnostic model evaluation. By optimizing only 10 soft-prompt vectors (roughly 0.0006% parameters for a 7B model) over a short tuning period, we adapt models to specific benchmark formats, closing gaps in format-following and ensuring that underlying knowledge is accurately reflected in benchmark scores. This allows one to fairly compare different base models -- trained with various pre-training recipes -- on benchmarks without the need for full post-training. We evaluated soft-prompt tuning across 7 models and 7 datasets. The results show that (a) soft-prompt tuning saturates format-following within 80 steps (~640 samples) making it highly efficient, (b) soft-prompt tuning significantly outperforms zero- and few-shot prompting, surfacing base model knowledge that standard prompting misses, that (c) even post-trained models can benefit from soft-prompts to maximize format compliance, and that (d) soft-prompted base model performance predicts post-trained model rankings more reliably than zero- and few-shot baselines, offering a low-cost proxy for downstream model quality. Our contributions include (1) metrics which disentangle format-following and knowledge accuracy, (2) a fairer benchmarking protocol of LLM knowledge, and (3) a cost- and memory-effective recipe to identify optimal pre-training strategies early in LLM development.

LLMs are increasingly used to generate and judge scientific ideas. This makes novelty evaluation a central problem. Full idea evaluation is difficult because it often requires judging a method, its feasibility, and its empirical promise. We therefore study a cleaner upstream object: the research question (RQ). RQ generation is a prerequisite for scientific ideation, and RQs can be compared against questions pursued in real papers. We introduce RQ-Bench, a benchmark built from recent arXiv papers. For each paper, we reconstruct author-anchored RQs from its cited background, gaps, and contributions. These RQs are not the only valid questions for the same background. They are author-anchored reference points for testing novelty judgments. We evaluate model-generated RQs with standalone LLM judging, comparative LLM judging, and human expert evaluation. LLM judges consistently rate model-generated RQs as highly novel, producing a novelty mirage; in comparative evaluations, this preference becomes even stronger. Domain experts, however, reach the opposite conclusion and prefer the author-anchored reference questions. We further find that many generated RQs are narrow or source-bound, a dimension that LLM judges often miss unless explicitly tested. Overall, the contradictory novelty evaluations between LLM judges and human experts raise a serious concern about the reliability of using LLMs to assess the scientific novelty of research questions.

Large language models (LLMs) have achieved remarkable progress in language understanding, reasoning, and generation, sparking growing interest in their creative potential. Realizing this potential requires systematic and scalable methods for evaluating creativity across diverse tasks. However, most existing creativity metrics are tightly coupled to specific tasks, embedding domain assumptions into the evaluation process, and limiting scalability and generality. To address this gap, we introduce an automated, domain-agnostic framework for quantifying LLM creativity across open-ended tasks. Our approach separates the measurement apparatus from the creative task itself, enabling scalable, task-agnostic assessment. Divergent creativity is measured using semantic entropy, a reference-free and robust metric for novelty and diversity, validated against human annotations, LLM-based novelty judgments and baseline diversity measures. Convergent creativity is assessed via a novel retrieval-based multi-agent judge framework that delivers context-sensitive evaluation of task fulfilment with over 60% improved efficiency. We validate our framework in three qualitatively distinct domains: problem-solving (MacGyver), research ideation (HypoGen), and creative writing (BookMIA), using a broad suite of LLMs. Empirical results show that our framework reliably captures key facets of creativity, including novelty, diversity, and task fulfilment, and reveal how model properties, such as size, temperature, recency, and reasoning, impact creative performance. Our work establishes a reproducible and generalizable standard for automated LLM creativity evaluation, paving the way for scalable benchmarking and accelerating progress in creative AI.

Adversarial robustness evaluations of large language models (LLMs) typically report attack success rate (ASR) under fixed query budgets, implicitly treating all attacks as equally costly. In practice, the computational expense of different attack strategies can vary by orders of magnitude. Consequently, ASR at a fixed budget can obscure the true effort required to jailbreak a model, thereby making it hard to determine whether an attack's cost justifies its payoff to the attacker. We propose a compute-aware evaluation framework based on computational pressure, measured in cumulative floating-point operations (FLOPs), as a proxy for adversarial effort. We introduce risk-compute curves, which map compute budgets to attack risk, and derive two metrics that summarize the average pressure required for a given attack to succeed. Across ten models spanning three families and four different stages in language model training and alignment, evaluated with three attack strategies (gradient-based, iterative refinement, and template-based) on two jailbreak robustness benchmarks, we find: (1) alignment training has non-monotonic effects on compute-space robustness; (2) scaling model size reduces gradient-based attack effectiveness but has limited impact on cheaper template-based attacks; (3) gradient-based attacks optimized on a surrogate model can transfer to a separate target model, providing a way to reduce attacker costs; (4) compute cost varies by up to ${\approx}5{\times}$ across harm categories within a single model; and (5) safety-aligned RL increases aggregate cost while leaving some categories disproportionately accessible. We release our framework to enable compute-aware risk assessment and evaluation.

Large Language Models (LLMs) are increasingly described as performing at the level of human experts on knowledge economy tasks. These claims are primarily based on how LLMs perform on benchmarking tasks that measure average performance across standardized datasets. Primary limitations of many benchmarking tasks are that they often measure performance based on content directly included in LLM training data, and they frequently do not assess the reliability of LLM performance or the magnitude of LLM errors. However, in high stakes contexts, these qualities are critically important. Through a novel LLM benchmarking task that requires writing computer code to complete a data analysis task, we compare the performance of a frontier LLM against submissions from human experts and explicitly measure the variance of responses and the magnitude of errors. Our study reveals that the human experts perform better on average on a range of metrics and demonstrate less variability in performance. Our results provide evidence that LLMs do not consistently perform at the level of human experts and demonstrate the importance of measuring variance and assessing error magnitude in LLM benchmark evaluations.

Automated Essay Scoring (AES) systems must judge interdependent discourse elements (e.g., lead, claim, evidence, conclusion), yet most approaches treat these in isolation, harming coherence and generalization. We investigate task-aware fine-tuning of LLaMA-3.1-8B for AES using parameter-efficient LoRA with 4-bit quantization and compare three training curricula: (i) Sequential (progressively fine-tuning on lead, then position, then claim, then evidence, then conclusion), (ii) Independent (task-specific models), and (iii) Randomized (shuffled multi-task). Experiments on the PERSUADE~2.0 corpus show that modeling task dependencies matters: Sequential fine-tuning yields the strongest overall results, including F1 scores of 65% (evidence) and 87% (conclusion) and corresponding accuracies of 63% and 85%, surpassing Independent training and outperforming a general-purpose LLaMA-70B baseline on conclusion despite its far larger capacity. Randomized training improves position scoring (57% F1) but is less consistent elsewhere. These findings indicate that (1) curriculum design aligned with discourse structure can materially improve AES, and (2) small, task-optimized models can be competitive with substantially larger Large Language Models (LLM), offering a practical path to scalable, cost-effective assessment. We release templates and implementation details to facilitate reproduction and future work on curriculum design for educational NLP.

LLM-as-judge is the default instrument for evaluating conversational agents, yet its reliability is almost always reported as agreement with human ratings, not recall of real defects. We study a deployed multi-turn food-and-beverage ordering agent and measure how many genuine quality problems its built-in LLM judge catches, using exhaustive human transcript review as ground truth. Across three batches the judge surfaces well under a quarter of human-confirmed systematic problems -- 2 of 9 patterns (22%) in one batch, and its operational gate flagged zero of 100 rounds in a batch where humans confirmed 23 distinct defects and 7 new cross-cutting patterns. Our blind-spot taxonomy shows the failure is structured, not random: the judge catches turn-local issues (a fabricated statistic, a wrong language) but misses cross-turn state issues (confirm-gate lockout, cart hallucination, escalation lockout, stale referents). The mechanism: the scoring rubric exposes only three coarse axes (intent, brand-voice, personalization) and has no category for the behavioural dimensions -- state-tracking, guardrails, recovery -- where most defects cluster. The failure is routing, not perception: 113 of 114 rounds whose raw judge note describes a confirm-gate or cart-state defect are scored "brand voice", and none reach an operational failure -- the gate is wired to hangs and hard assertions, not the rubric -- so the 0% is a routing-and-wiring failure, not blindness. The consequence for prevalence estimation is sharp: when the apparent defect rate is zero the Rogan-Gladen correction degenerates -- no signal can recover the true rate -- while where the gate reports a nonzero rate the same estimator implies a 3-6x undercount under our measured sensitivity. For production multi-turn agents, automated judging is a regression floor, not a substitute for human review.

Evaluating reasoning quality in multi-agent LLM systems is challenging, especially for open-ended tasks without reference answers. We investigate whether intrinsic confidence signals, token-level log-probabilities from decoding, can predict reasoning quality as assessed by LLM-as-judge evaluation. Using a debate-based essay scoring framework, we compare confidence proxies against rubric-based judge scores across two ASAP essay sets. We find that early-token confidence, particularly within the first few generated tokens, is consistently the strongest predictor of reasoning quality, outperforming full-sequence statistics. Analysis of log-probability trajectories shows that the opening phase of generation is the most heterogeneous and therefore most informative. We also observe a systematic asymmetry between agent roles, with stronger alignment between confidence and quality for supportive reasoning than for adversarial critique. These results suggest that early decoding dynamics provide a lightweight and effective signal for estimating reasoning reliability in multi-agent LLM systems.

Multi-agent debate systems are typically evaluated only on whether the final answer is correct, overlooking the quality of the intermediate reasoning that debate is designed to produce. This paper studies the relationship between three signals in multi-agent debate: token-level log-probability distributions over reasoning tokens, LLM-as-judge rubric scores assigned to those tokens, and final task accuracy. We examine whether internal confidence signals predict externally evaluated reasoning quality, and whether either signal aligns with task correctness, across three domains: rubric-based scoring, mathematical reasoning, and factual question answering. Our framework pairs a two-agent debate architecture -- a Constructor and an Auditor -- with an LLM-as-judge that scores each agent's reasoning along instruction following, justification quality, and evidence grounding, together with a critical-failure flag. Experiments in the rubric-scoring domain reveal a consistent four-phase confidence trajectory and a substantial role asymmetry: confidence aligns with judged reasoning quality roughly twice as strongly for the Constructor as for the Auditor, and confidence-based detection of critical reasoning failures is markedly more reliable for the Constructor (AUROC 0.804) than for the Auditor (0.634). These findings motivate the broader cross-domain investigation proposed in this paper.

While Large Language Models (LLMs) have achieved near-perfect performance in \emph{solving} high-school mathematics, their ability to \emph{evaluate} the diverse reasoning processes of real human students remains under-examined. To bridge this gap, we introduce \textbf{RealMath-Eval}, a rigorously annotated benchmark of 224 real-world exam responses from high schools. Our initial evaluation reveals that even state-of-the-art LLM judges struggle significantly on this task, exhibiting a high Mean Squared Error ($\sim$2.96) against expert human grading. To probe a plausible explanation, we contrast this performance with a control setting where the same judges evaluate synthetic LLM-generated solutions. We identify a stark Evaluation Gap'': judges are considerably more accurate and consistent on synthetic text (MSE $\sim$1.17) but struggle to generalize to authentic student reasoning. Through semantic embedding analysis, we find that synthetic errors suffer from a structural collapse'' into predictable, low-dimensional linear subspaces, whereas human errors form a more diverse error space. Furthermore, generative probability probes suggest that human reasoning involves significantly higher information-theoretic surprisal, indicating that student reasoning transitions are more out-of-distribution for current models. Finally, we find that surface-level style transfer fails to close this gap. Our findings suggest that current LLM evaluation pipelines relying heavily on synthetic data may not adequately capture the diversity of authentic student mathematical reasoning.

LLMs have recently achieved strong results on formal proving benchmarks. However, existing evaluations remain heavily concentrated on competition-style problems and often fail to capture how models behave on longer, more dependency-rich mathematical developments. We introduce TheoremBench, a Lean4 benchmark designed to evaluate theorem provers beyond contest settings. The benchmark is built from nearly one hundred classical theorems and is released in two complementary forms: a plain main version containing one target theorem per instance, and a premised version that expands each theorem into a structured family of related proving tasks consisting of the main theorem together with automatically extracted supporting subtheorems. This design enables evaluation of not only whether the final theorem was proved from scratch, but also of partial progress through the internal proof structure of a theorem. Our experiments show that explicit premises substantially improve performance for Lean4-capable prover models. To provide a comprehensive evaluation, we introduce theorem-level coverage and token-efficiency metrics that expose qualitative differences in proof behavior. The results show that current provers remain strongly biased toward easy subtheorems and often solve theorems through long and inefficient tactic traces rather than compact proof plans. TheoremBench therefore provides a more fine-grained view of formal reasoning ability and highlights the importance of structural benchmark design for evaluating Lean4 theorem provers.

Safety judges are increasingly deployed to evaluate model outputs against evolving criteria, yet recent meta-evaluation work shows they remain brittle under prompt and rubric variation, with false negative-rate swings of up to 0.24 reported for stylistic perturbations alone. We argue that safety judgment is fundamentally a rubric-following problem: a robust judge must apply the given evaluation criteria consistently across rubric formulations rather than memorize one specific template. We propose a training strategy that combines (i) instance-conditioned dynamic rubrics generated from prompt-response-label triples to expose the judge to the variability of evaluation criteria, and (ii) a reliable-to-expressive curriculum that begins with clean fixed-rubric supervision and progressively introduces noisier dynamic-rubric data. We evaluate on a single human-labeled set under three contrasting rubric prompts (HarmBench-style, ShieldGemma-style, and a domain-specific rubric). Our 12B curriculum judge achieves 94.12-94.88% accuracy across the three rubrics with a cross-rubric range of only 0.76, outperforming general-purpose LLMs, dedicated safety classifiers, and reasoning-oriented judges up to 30B in both peak accuracy and stability. An ablation shows that naively mixing dynamic rubrics into SFT increases cross rubric variance (1.44 -> 3.60); only the curriculum schedule recovers and improves on the fixed rubric baseline (variance 0.76).

LLM-based RTL generation and reasoning is a promising direction for hardware design automation. High-quality benchmarks are critical infrastructure for tracking progress in this direction. However, existing RTL benchmarks face inherent limitations in both scale and task scope. The designs they cover are typically small and simple, and the tasks focus almost entirely on specification-to-RTL generation. Frontier models' performance already saturates on the existing benchmarks. Scaling these benchmarks up is fundamentally difficult because aligned labels are required for benchmarking, such as specifications and testbenches. Such aligned high-quality data are rarely available for real-world designs. We introduce RTL-BenchLS, a large-scale benchmark addressing both limitations above. It contains over 10,000 formally verified Verilog designs, covering substantially larger and more complex designs than existing benchmarks. Beyond specification-to-RTL generation, we propose three novel tasks that jointly evaluate reasoning and generation: round-trip reasoning, masked-content reasoning, and repository-issue reasoning. The first two are self-supervised, which directly resolves the scaling bottleneck. All tasks are verified through formal equivalence checking without any manual testbenches. We evaluate eight LLMs on RTL-BenchLS. Even the best model reaches only 23% on natural-language round-trip reasoning, 28% on masked-content reasoning, and 12% on repository-issue fixing. RTL-BenchLS is substantially more challenging than existing benchmarks. It leaves ample room for future improvement and offers guidance for developing LLM-based methods for hardware design.

The rapid rise in LLM capabilities has made AI agents increasingly viable across a broad range of tasks. Among the most promising applications is building production-ready customer-facing agents, a challenge that demands coordinated excellence in evaluation methodology, context engineering, training, and online measurement. Yet these critical pillars are typically developed in isolation, creating blind spots that only surface after deployment. In this paper, we present a unified framework that bridges offline development with online impact for customer support AI agents at Nubank, a company with 100M+ users. Our approach integrates several key components: (1) structured context engineering tailored to customer support agents, (2) systematic human-in-the-loop prompt iteration, (3) rigorous LLM judge evaluation with measured inter-rater agreement and GEPA optimization for consistency, and (4) ideation-to-production validation. A central insight is that evaluation-pipeline quality directly determines iteration velocity. We present results from five production deployments spanning distinct domains: card delivery, debt management, credit-limit support, card management, and product explanation. These deployments deliver consistent customer-satisfaction gains while substantially accelerating iteration. In our card-delivery deployment, large-scale A/B testing yields a 37 percentage-point improvement in AI transactional Net Promoter Score and a 29 percentage-point gain in self-service rate over prior agent variants, alongside a strong correlation between offline simulation metrics and online outcomes, demonstrating that eval-driven development reliably predicts production impact. On most use cases, AI satisfaction reaches within a few percentage points of expert human agents.

Code generation models are typically compared using compact execution benchmarks and aggregate pass rates, but such summaries obscure how performance varies across programming languages, problem families, and failure modes. We present a large-scale, execution-grounded evaluation of 9 openly accessible LLMs specialized for coding on 2,707 free LeetCode problems across 12 programming languages. Our corpus contains 325,343 problem-model-language jobs, each linked to prompt metadata, extracted code, LeetCode execution outcomes, and static-analysis signals. The results show that current open models remain far from the human acceptance reference: the best model, Yi-Coder-9B-Chat, reaches 23.64% mean correctness, compared with a 57.2% human acceptance baseline. Rankings are also slice-dependent: Qwen2.5-Coder-14B-Instruct is strongest on hard problems and distinct-problem coverage, while Gemma-2-27B-IT achieves the highest all-language lint pass rate. Failure analysis shows that compile errors account for 63.25% of non-accepted best submissions, indicating that many failures occur before semantic correctness can be tested. Static quality further diverges from functional correctness. Together, these findings show that multilingual, artifact-preserving evaluation reveals tradeoffs hidden by single-language or single-metric leaderboards.

As Large Language Models (LLMs) advance toward open-ended autonomous agents, the mechanisms used to evaluate and guide their behavior must evolve accordingly. This work introduces the rubric as a unifying framework capturing this evolution, characterizing rubrics as a dynamic response to successive LLM paradigm shifts that recurs across otherwise independent efforts in evaluation, reinforcement learning, and safety alignment. We define rubrics as explicit criteria sets that transform complex quality judgments into structured and actionable standards, and demonstrate that their recurrence across these research threads is not coincidental. We systematically organize existing rubric designs, examine their construction and optimization, and analyze their role across evaluation and training. Rubrics manifest at three progressively deeper levels: at the evaluative level, they decompose holistic judgments into verifiable dimensions; at the training level, they serve as dense feedback signals providing process-level guidance where scalar rewards fall short; at the intrinsic level, they emerge dynamically from model behaviors, driving self-improvement. We further assess rubric reliability across generation quality, execution fidelity, theoretical constraints, and security threats, before surveying rubric-based benchmarks across diverse domains. By rendering assessment transparent and decomposable, rubrics translate human value expectations into machine-learnable signals, serving as the enduring bridge between human intentions and machine behavior.

Large language models (LLMs) are increasingly evolving from simple text-based interaction systems into LLM agents that can maintain memory, use tools, access external environments, and execute tasks. As their capabilities and autonomy expand, the safety risks they face also become more diverse. Existing evaluations often rely on manually written scenarios, static prompts, or final-output judgments, making it difficult to capture the diverse risks that agents may face during task execution. We introduce ForesightSafety-SAGE, a fully automated scenario generation and safety evaluation framework for LLM agents. Based on five risk dimensions,we instantiae abstract and diverse safety risks in real-world task execution into 1,072 measurable evaluation scenarios. Using the automated evaluation pipeline, 12 LLM agents are evaluated under two authority contexts. The results show that current agents still face substantial behavioral safety risks during task execution, with an average ASR of 47.1% and several models exceeding 70%. These findings demonstrate the importance of executable, process-level evaluation for understanding and improving LLM agent safety.

Evaluating LLM-powered interactive social agents is challenging because socially relevant behaviors depend not only on isolated outputs, but also on prior interactions, social roles, and downstream actions. Existing methods typically allow a target agent to act freely in an environment and then score the resulting trajectory. However, this passive setup can miss capabilities that only become observable under specific social circumstances; for example, conflict handling may remain untested if no disagreement arises. We propose Online Agent-as-a-Judge, a situation-generating evaluation framework for interactive social agents. Online Agent-as-a-Judge deploys an in-world evaluator agent that interacts with the target agent through the environment's native dialogue and action protocol, actively eliciting situations relevant to the evaluation criteria. The resulting trajectories provide evidence for assessing both immediate responses and subsequent behavior. In a life-simulation environment with $32$ designer-authored social criteria, Online Agent-as-a-Judge improves criteria coverage and agreement with human labels, yielding more reliable evidence-grounded evaluations of behaviors that passive methods can leave unobserved.

Multilingual LLM-as-a-judge is widely used to evaluate model outputs across languages, but suffers from cross-lingual inconsistency (Fu and Liu, 2025). Existing methods typically treat this inconsistency as noise and mitigate it through voting or aggregation. In this work, we instead show that multilingual inconsistency can provide complementary evaluation signals. Our oracle analysis finds that sampling judgments across languages yields a higher performance upper bound than single-language judging, indicating that different languages potentially include complementary judgments. Motivated by this finding, we propose SEMJ, a self-evolving multilingual judge that leverages cross-lingual inconsistency for iterative refinement. SEMJ constructs multilingual variants of each input, collects independent judgments and rationales, and feeds inconsistent outputs back for self-reflection and re-evaluation. Experiments on multiple benchmarks show that SEMJ consistently outperforms voting and reflection baselines in both accuracy and cross-lingual consistency. Further analysis shows that inconsistency triggers useful re-evaluation, which improves judgment quality.

Reliable evaluation of large language models in surgery remains underdeveloped. Broad medical benchmarks test clinical knowledge, while surgery requires procedural reasoning, management trade-offs, negation handling, and selection among plausible operative decisions. We present SurgiQ, a text-only, source-grounded benchmark of 13,055 four-option multiple-choice questions spanning six surgical domains and four question formats: case-based, reasoning, best-option, and negative. SurgiQ is constructed from surgical textbooks, open-access papers, and examination material using a multi-stage generation, verification, and expert-audit pipeline. We evaluate 35 open-weight LLMs under a unified log-likelihood protocol. Our results show substantial remaining headroom: smaller models often remain near the 25% random baseline, while the best model reaches 68.1% accuracy. General-purpose models, especially Qwen2.5, outperform most biomedical models, suggesting that current medical specialization does not yet provide sufficiently broad surgical coverage. Calibration and error analysis further show that even strong models make confident mistakes on clinically plausible distractors, motivating more reliable and broader surgical LLM evaluation.

Large Language Model (LLM) safety has often been evaluated at the behavior level, which provides limited evidence of internal robustness, as these evaluations target outputs rather than representation-level vulnerability under intervention. We formalize this discrepancy as the audit gap: the difference between behavioral safety and robustness under intervention. To study this gap, we construct dissociated models that preserve safe outward behavior while remaining vulnerable in the latent space. We introduce an intervention-based evaluation framework to test model robustness through soft interventions in parameter and latent spaces, including harmful fine-tuning and layer-wise latent perturbations. To formalize the evaluation, we propose the Latent Vulnerability Score (LVS) to measure how easily harmful behavior can be elicited by bounded latent perturbations. Using this evaluation framework, we show that behavioral safety metrics are insufficient measures of representation-level robustness across multiple safely and unsafely aligned state-of-the-art models. Notably, dissociated models show substantially elevated LVSs despite comparable refusal behavior under harmful intervention, with intermediate representations being the most sensitive to intervention. Our results suggest that behavioral safety evaluation alone provides an incomplete picture of model robustness, motivating representation-aware audits of latent vulnerability and observable behavior.

LLMs-as-judges are the only way to evaluate safety at scale. Despite their importance, LLM-judges themselves are rarely evaluated beyond human agreement in simple, static benchmarks. We therefore investigate two under-explored but crucial properties of LLMs-as-judges: their susceptibility to relying on in context-information, and their steerability to differing safety definitions, which may not align with their internal safety priors. We evaluate the safety judging abilities of many generalist LLMs and safety-specific judges, and investigate the impact of task demonstrations, novel in-context information, and changing safety definitions. We find that while LLM-judges can learn from new information, they are broadly unlikely to adjust their evaluations if the context or safety definition contradicts their prior.

LLM judges increasingly turn verdicts into system commitments. Under mixed evidence (claims with both supporting and refuting sources) this is unsafe: when the schema exposes CONFLICTING as the authorized non-directional verdict, returning SUPPORTS/REFUTES is an unauthorized directional commitment, a failure we name Cherry-pick Override (CCO). We define CCO under an explicit task contract and report it with a same-denominator diagnostic protocol paired with matched-coverage bootstrap and an apples-to-apples random-veto null. On AVeriTeC's Conflicting subset (N_C = 150), three-option judges return a directional verdict on more than 84% of mixed-evidence claims; under the typed schema, three-judge majority voting amplifies direction-on-conflict on AVeriTeC (0.887 vs. 0.840; 95% CI [+0.013, +0.080]) but does not replicate on VitaminC-Mixed. Walking an intervention ladder of common single-channel fixes (typed vocabulary, panel aggregation, confidence thresholding, validator-only filtering), each leaves a distinct residual failure: panel aggregation suppresses single-judge CONFLICTING dissent in 48% of CCO cases; the panel is well-calibrated for direction (ECE = 0.07 on pure-S/R) so confidence cannot operationally separate CCO from correct directional commits; validator-as-classifier nearly halves pure-evidence accuracy. A minimal two-channel reference probe reaches operating points neither single channel reaches; under the random-veto null its promotion to CONFLICTING is structurally targeted on AVeriTeC (empirical p < 1/2001) and weaker but in the same direction on VitaminC-Mixed, a selectivity result rather than a magnitude one. We argue for an external commitment-control layer that separates verdict generation from commitment authorization, using structural evidence and confidence as orthogonal channels and NO-COMMIT as a routed controller state.

Large language models (LLMs) are widely used as judges for evaluating model outputs, but their high cost, latency, and opacity limit scalability. We introduce SLMJury, a framework for evaluating small language models (SLMs) as judges across two paradigms: closed-ended binary correctness and open-ended quality scoring. We benchmark 16 SLM judges (0.6B-14B parameters) from four model families across ten benchmarks: eight closed-ended tasks spanning mathematical, scientific, and general reasoning (N=64,824 judgments per configuration), plus SummEval and MT-Bench for summarization and conversational scoring. We formalize judging as a budget-conditioned function and study five dimensions. Four findings emerge. (1) The overthinking effect is domain-dependent: for most judges quick 10-token verdicts match or beat extended reasoning on mathematical judging (by 2-7% where they help), while reasoning wins on general tasks by up to 23%. (2) Domain generalization separates model families, with math-to-general accuracy gaps ranging from under 10% to nearly 40%. (3) Closed-ended and open-ended judging draw on different capabilities: the best binary judge (Phi-4) drops to rank 9 on MT-Bench, while reasoning-trained models invert this ordering. (4) Under the Reflect-Critique-Refine (RCR) debate protocol, multi-agent debate degrades accuracy across all tested configurations, whereas the top judges resist six adversarial personas with <=0.55% variance. Reliable automated evaluation does not require large proprietary models, yet no single SLM dominates. The leaderboard is available at https://anishh15.github.io/SLMJury/, and our framework code and pip package are publicly available at https://github.com/anishh15/SLMJury and https://pypi.org/project/slmjury/.

Language is a vehicle for thought, intricately tied to sounds, symbols, and meaning. However, most large language model (LLM) research focuses on meaning (semantics) and symbols (spelling) while largely overlooking sounds. Existing benchmarks on LLMs' phonological abilities are either solvable through rote memorization or intertwined with other abilities, making them inadequate to measure LLMs' genuine ability in phonological understanding. Here, we present Phun-Bench, a purpose-built Chinese benchmark with diverse tasks and settings across three dimensions (Homophony, Rhyme, and Phonetic Similarity), designed to systematically evaluate LLMs' phonological understanding. Our results show that while LLMs excel at recalling correct pronunciations, they generally struggle to leverage phonological knowledge in the flexible and intuitive way that human speakers do. Moreover, through detailed analyses, we propose a hypothesis regarding the underlying mechanism of LLMs' phonological understanding and "perception", highlighting an underexplored frontier for future research.

Hallucination detection is essential for the reliable deployment of large language models (LLMs). However, existing evaluations face two core challenges: inconsistent inference configuration and evaluation, and limited coverage of downstream domains and tasks. Consequently, reported detector performance is often difficult to compare, reproduce, and generalize beyond specific experimental settings. We introduce OpenHalDet, a unified benchmark for hallucination detection across diverse generation scenarios. OpenHalDet standardizes the evaluation pipeline, from prompt construction and response generation to truthfulness annotation, detector scoring, and metric computation. It supports heterogeneous detector families under different access settings, including black-box methods that use only generated outputs, gray-box methods that rely on probability-based signals, and white-box methods that exploit internal model signals. By bringing diverse tasks, models, and detectors into a shared framework, OpenHalDet enables controlled comparison and provides a systematic view of how different detection paradigms behave in LLM applications. We release OpenHalDet as an open and extensible codebase to facilitate reproducible evaluation and future development of hallucination detection methods. The code and datasets are available at https://github.com/Nellie179/Hallucination-Detection.

We present MADRAG, a training-free framework for analytic essay scoring that combines multi-agent reasoning with retrieval-augmented grounding. Unlike standard LLM-as-judge approaches, which are prone to bias and unstable scoring, MADRAG decomposes evaluation into an interactive process: an Advocate identifies strengths, a Skeptic critiques weaknesses, and a Judge aggregates their arguments into a final score. Crucially, the Judge is augmented with rubric-aligned exemplar retrieval, enabling calibration through comparison with scored examples. Our results show that MADRAG significantly outperforms prompt-based baselines while approaching the performance of supervised systems without requiring task-specific training. Ablation studies demonstrate that retrieval drives calibration gains, while debate improves reasoning on higher-level traits. Our findings highlight the complementary roles of structured interaction and external memory in reliable LLM-based evaluation.

Qualitative product feedback can reveal nuanced user experiences, but its implicit sentiment is difficult to measure. This paper presents a scalable and interpretable framework that uses large language models (LLMs) to quantify product desirability from such data. Using two Product Desirability Toolkit (PDT) datasets from ZORQ and CARMA comprising 106 respondent term groupings with gold-standard human annotation, zero-shot continuous numerical sentiment scoring and categorical sentiment classification are evaluated without relying on explicit review scores. Across the datasets, LLMs generated numerical sentiment scores directly from qualitative responses and closely matched expert labels, achieving Pearson correlations up to 0.97 and classification accuracy up to 94%. LLMs maintained robustness even when handling data presented in multiple forms and consistently expressed high confidence. In contrast, lexicon-based and transformer baselines did not produce statistically significant results. Among the models tested, GPT-4o-mini achieved performance comparable to larger models at 94% lower cost, supporting scalable deployment. The framework also incorporates model confidence ratings and human-readable rationale explanations (xAI), improving interpretability, transparency, and trust while supporting practical use in product satisfaction assessment. In general, using the PDT tool as a survey method along with a cost efficient LLM for sentiment analysis has the potential to provide for product evaluation with results that are rich in terms of sentiment scores (both numerical and classified sentiment) and in terms of the high-level user impressions of the product that can be used to identify ideas for product development and improvement, as well as marketing ideas for target audiences.

Appropriate reliance on AI advice has become a central research theme in human-AI collaboration. Existing frameworks have focused exclusively on point predictions as AI advice. However, set-valued AI advice (e.g., discrete sets or continuous intervals) is increasingly being used to communicate uncertainty and improve human decision making. In this paper, we develop the first formal framework for measuring appropriate reliance on set-valued AI advice within the sequential judge-advisor paradigm, spanning both classification and regression tasks. For classification, we first introduce the dimensions that are necessary for evaluating set-valued AI advice. We then define two metrics: correct reliance rate on AI and correct reliance rate on self, which jointly characterize appropriate reliance in this setting. For regression, we introduce quantity of AI reliance and quality of AI reliance, which respectively measure whether a decision maker utilized the AI advice and whether their reliance helped them get closer to the ground truth relative to their initial estimate. Through the application of our framework, we demonstrate how these metrics capture important nuances in human-AI collaboration that existing measures overlook.

Evaluating large language models (LLMs) for education requires measuring how models teach, not only what they know. Existing benchmarks emphasize domain-general correctness or depend on manually designed rubrics that scale poorly to long-tail pedagogical scenarios. We introduce Elmes*, an end-to-end framework for constructing, refining, and applying fine-grained scenario-specific rubrics. Elmes* combines a declarative multi-agent engine for teacher--student--judge interactions with SceneGen, a self-evolving module that co-optimizes evaluation criteria and test data from expert-defined pedagogical dimensions. Using Elmes*, we build Edu-330, covering 330 scenarios across 11 subjects, 3 grade bands, and 10 task types, with over 1{,}000 second-level indicators. Experiments on Edu-330 and four expert-authored gold-standard scenarios show that educational capability is multidimensional: top-tier LLMs differ mainly in creativity and values integration, knowledge-strong models may fail at Socratic scaffolding, and the education-specialized InnoSpark achieves the best human-evaluated average score. LLM judges preserve human-comparable rankings with much lower scoring variance, but exhibit judge-specific biases such as self-preference. Ablations show that expert-scored few-shot anchoring improves human--LLM alignment, while reasoning enforcement and greedy decoding are model-dependent. Elmes* thus provides scalable diagnostic infrastructure for pedagogically grounded LLM evaluation.