人工智能大模型评测标准权威文献

当前医学人工智能标准化进程尚处于萌芽阶段，难以满足医学人工智能产品在研发、部署、管控、评估以及指导等多方面的需求。这一方面导致人工智能产品的研发过程难以规范化，增加了研发成本，影响了产品质量；另一方面也造成了人工智能产品难以进行统一的交互、比较和评价，可能导致产品被错误评估，从而误导医学人工智能产品的研发方向，因此建立成熟统一的医学人工智能标准体系成为当务之急。为推动医学人工智能标准体系从萌芽阶段走向成熟，本文从医学数据标准、标准数据集、基准和规范/指南4个方面深入分析医学人工智能标准的发展历程，揭示当前医学人工智能标准中存在的问题，以期为相关研究提供参考和借鉴。

医疗大模型基于深度神经网络架构进行复杂医疗数据的高效处理与模式识别，为智慧医疗提供新型的决策支持；需要系统分析医疗大模型技术及应用情况，以精准把握医疗大模型的发展方向、精准应对面临的发展挑战，进而基于医疗大模型提升医疗文本、医学图像、药械研发、医学教育等方面的能力。本文梳理了医疗大模型的技术范式与应用场景，剖析了由基础层、模型层、应用层、公共模块构成的医疗大模型技术体系，覆盖评价指标体系构建、数据集范围与题型、模型对齐方法、模型评测平台的医疗大模型评测体系，辨识出医疗大模型应用存在的数据安全、技术风险、落地挑战、伦理道德等方面的难点。为此建议，发挥政府引导优势、保障数据安全，加快基础理论研究、突破技术风险，强化应用场景牵引、缓解落地挑战，建立健全监管机制、规范伦理道德，完善公共服务体系、营造创新生态，以加快医疗大模型创新应用，推动我国智慧医疗的高端化、智能化、绿色化发展。

【目的】随着当前社会科学技术的发展, 文本摘要技术被广泛应用在生活中的各个方面, 并发挥着重要作用。然而在司法领域中, 使用传统深度学习模型生成司法文本摘要存在冗余、信息不一致、对文本语义理解不够等问题；而且现有生成摘要的质量评价方法ROUGE较为单一, 仅关注生成摘要与参考摘要之间的重叠, 不关注深层语义信息, 需要构建新的指标多维评价生成摘要质量。【方法】本文借助知识编辑、参数微调技术构建司法领域垂直大语言模型, 生成司法文本摘要, 改善传统模型对专业司法文本理解不够的问题；并提出信息缺失因子、信息密度两种评价指标, 用大语言模型对摘要实现知识抽取, 度量文本知识, 根据度量结果计算两类指标, 进一步在内在语义层次衡量生成摘要的质量。【结论】通过实验证明, 本文提出的基于垂直领域大语言模型生成司法文本摘要的方法有助于改善信息冗余、对专业文本理解能力不够的问题, 且提出的两类摘要评价指标从信息一致性、冗余度两方面补充评价了摘要生成质量, 丰富了评价摘要的方法。

采用定量和定性相结合的混合研究方法, 从准确性、规范性、情感性和创造性四个维度评估了人工智能的语言优势和不足。研究1发现, 相对于真实学生, GPT-4现代文知识(尤其概念知识)的准确性较高, 但其古代诗文和语言文字运用的准确性较低; GPT-4规范性得分与真实学生相当, 情感性和创造性超过及格水平、但低于真实学生, 且前者最优个体的规范性、情感性得分与真实学生最高分持平。研究2基于文心ERNIE-4重复验证了上述结果, 且ERNIE-4的规范性得分高于真实学生。研究揭示了人工智能在现代文知识、规范领域的优势和古代诗文知识的不足, 以及情感性与创造性方面的潜力。这些发现有助于理解和提升人工智能的文化适应性和人性化、个性化生成能力, 也对反思和培养人类的独特优势具有重要启发。

大语言模型能够在多个领域及任务上给出与人类水平相当的解答，并且在未经训练的领域和任务上展现了丰富的涌现能力。然而，目前基于大语言模型的人工智能系统存在许多安全性隐患，例如大语言模型系统容易受到难以被察觉的攻击，模型生成的内容存在违法、泄密、仇恨、偏见、错误等问题。并且在实际应用中，大语言模型可能被滥用，生成的内容可能引起国家、人群和领域等多个层面的困扰。本文旨在深入探讨大语言模型面临的安全性风险并进行分类，回顾现有的评估方法，研究安全性风险背后的因果机制，并总结现有的解决措施。具体而言，本文明确了大语言模型面临的10种安全性风险，并将其归类为模型自身安全性风险与生成内容的安全性风险两个方面，并对每种风险进行了详细的分析和讲解。此外，本文还从生命周期和危害程度两个角度对大语言模型的安全风险进行了系统化的分析，并介绍了现有的大语言模型安全风险评估方法、大语言模型安全风险的出现原因以及相应的缓解措施。大语言模型的安全风险是亟待解决的重要问题。

多媒体信息在人类社会的发展历程中有着至关重要的作用,构建具有多模态信息处理能力的智能系统也是通往通用人工智能的必经之路。随着预训练技术的发展以及对于通用模型的需求,多模态的研究也从早期的任务特定的方法转移到了构建统一泛用的多模态基座模型上。初步的统一多模态模型探索受到BERT启发,从表征学习的角度出发构建能为不同下游任务提供有效初始化的多模态预训练模型,这类方法尽管有效但仍然在泛用性方面受限于预训练中微调范式,无法更广泛高效地应用。近年来随着大语言模型的发展,以大语言模型为基座的多模态大模型则展现出了巨大的潜力:此类模型有着强大的信息感知,交互,以及推理能力并且能有效泛化到多样的场景下,为新时代的通用人工智能系统提供了切实可行的思路。本文将从构建统一多模态模型的角度出发,介绍和梳理相关工作的发展,从多模态预训练到多模态大模型,介绍对应的架构,训练,评测方法以及发展趋势,为读者提供一个全面的概览。

随着多模态大语言模型在感知任务中取得显著成果，如何提升其复杂推理能力成为研究重点。当前模型存在推理路径不透明、泛化能力不足等问题，而思维链推理有助于增强模型的推理透明性和输出可解释性，且已在大语言模型中取得显著效果，将该范式扩展至多模态领域，有望提升模型的推理能力。该文围绕“多模态思维链”展开系统梳理。首先，从技术演进与任务需求角度分析其提出背景与理论动因；其次，从思维链范式、后训练阶段与推理阶段三个方面介绍了当前多模态思维链的主流方法，并分析了其内在机理。此外，该文还总结了现有评测基准与指标，并探讨了多模态思维链的应用场景。最后，分析了当前多模态思维链面临的挑战并展望其未来发展方向。

多模态大语言模型(MLLMs)能够处理并整合图像、文本等多模态数据信息, 为理解人类心理与认知行为提供了强有力工具。结合经典的情绪心理学范式, 本研究通过比较两种主流MLLMs与人类被试在动态社会互动情景下情绪识别与情绪推理的表现, 分离出人物对话视觉特征(图像)和对话内容(文本)在识别与推理相关人物情绪中的不同作用。结果表明, 基于人物对话图像和对话内容的MLLMs已经初步展现出和人类被试类似的情绪识别与情绪推理能力。之后进一步比较仅基于人物对话图像、仅基于对话内容以及基于两者结合共三种条件下MLLMs的情绪识别与情绪推理表现, 发现人物对话视觉特征一定程度上制约MLLMs基本情绪识别的表现, 但能够有效促进复合情绪识别, 对情绪推理则未产生显著影响。通过对比两种主流MLLMs及其不同版本(GPT-4-vision/turbo vs. Claude-3-haiku)的表现, 发现相较于单纯扩大训练数据规模, 技术原理框架的创新对提升MLLMs在社会互动中情绪识别与推理能力更为重要。本研究结果对理解社会互动中情绪识别与推理的心理学机制、启发类人的情感计算与智能算法具有重要科学价值和意义。

随着人工智能技术的快速发展, 大语言模型在工业领域的应用日益广泛, 特别是在智能制造、智能监控等重要场景具有不可替代的作用, 成为新型工业化的核心驱动力。基准测试作为评估工业领域大语言模型性能的重要工具, 对于促进人工智能技术进步与产业应用至关重要。为应对工业领域大语言模型评估的复杂性问题, 提出Industrial Domain-Evaluation (ID-Eval)评测基准, 聚焦于工业领域大语言模型性能验证与优化, 从通用能力、应用能力、可信能力三个方面全面评估该模型在知识理解、常识问答、工程建模、采购分析等场景中的应用, 旨在推动人工智能技术从通用场景向工业领域特定场景转化, 促进其在设计、生产、检测等关键环节的深度应用, 加速推动工业领域智能化进程。

The rapid development of large language models (LLMs) presents both opportunities and challenges for their evaluation. While evaluation systems for general-domain LLMs are becoming more refined, assessments in specialized fields remain in the early stages. This study evaluates LLMs in the domain of classical Chinese, designing a series of tasks based on two key dimensions: language and knowledge. Thirteen leading general-domain LLMs were selected for evaluation using major benchmarks. The results show that ERNIE-Bot excels in domain-specific knowledge, while GPT-4 demonstrates the strongest language capabilities. Among open-source models, the ChatGLM series exhibits the best overall performance. By developing tailored evaluation tasks and datasets, this study provides a set of standards for evaluating LLMs in the classical Chinese domain, offering valuable reference points for future assessments. The findings also provide a foundation for selecting base models in future domain-specific LLM training.

Evaluating machine learning in scientific domains requires separating correct predictions from correct reasons under realistic distribution shifts. We introduce PertReason, a knowledge-grounded benchmark and framework suite for cell-state--conditioned reasoning about perturbation effects. At its core, PertReasonQA is a benchmark that tests whether models can generate mechanistically faithful explanations while remaining robust to complex shifts, such as new cells and unseen perturbations. PertReasonQA combines single-cell genetic and chemical perturbation data across multiple cellular contexts with knowledge graphs, and dynamically conditions pathways on cell-specific basal states to avoid generic memorization. Evaluations on state-of-the-art models reveal systematic gaps between predictive accuracy and mechanistic reasoning. Specifically, these models exhibit failure modes largely invisible to standard benchmarks, such as deriving correct answers through flawed logic, ignoring cellular context, and generating directionally inconsistent mechanisms. As a reference probe of the benchmark, we present PertReasonLM, a large language model trained to align outcome predictions with context-specific mechanistic reasoning. Our model targets the identified failure modes by grounding rationales in context-specific pathways and tightening agreement between outcomes and mechanisms. Together, we provide a diagnostic framework for exposing and mitigating failures in faithful reasoning in data-rich scientific systems.

Autonomous AI research agents aim to accelerate scientific discovery by automating the research pipeline, from hypothesis generation to peer review. However, existing benchmarks rarely test a fundamental bottleneck: whether Large Language Models can judge the methodological viability of a research idea before expending time and computational resources. We introduce SoundnessBench, a curated benchmark of 1,099 machine-learning research proposals reconstructed from ICLR submissions, labeled with reviewer soundness sub-scores, and audited against source papers. SoundnessBench should be interpreted as a benchmark for recoverable proposal-stage soundness rather than exact prediction of full-paper review outcomes. Across 12 frontier LLMs, we find a pervasive optimism bias: under standard prompting, models frequently rate low-soundness proposals as sound, while aggressive prompting largely shifts errors from false positives to false negatives. Additional controls for public-corpus contamination, paper-identifying phrases, surface features, and human audit quality suggest that this behavior is not explained by a single confounder. Our results indicate that current LLMs are not yet reliable as standalone first-gate evaluators for scientific rigor.

While multimodal large language models (MLLMs) have achieved rapid progress in vision-language understanding, they remain prone to multimodal hallucinations, producing responses that are inconsistent with the visual input. Existing benchmarks predominantly focus on detecting hallucination outcomes rather than evaluating the underlying causes of these failures. Moreover, many benchmarks rely on simplistic scenarios and limited evaluation formats that no longer challenge state-of-the-art models. To address these limitations, we introduce ReactBench, a cause-driven hallucination benchmark featuring multiple tasks and an exam-style evaluation format. By generating adversarial images and hallucination-inducing queries, ReactBench introduces four targeted tasks: Relational Erasure, Counterfactual Attribute, Alteration Tracing, and Dense Counting. These tasks systematically expose co-occurrence bias, language priors, cross-image comparative perception deficiencies, and fine-grained perceptual bottlenecks. Beyond standard accuracy-based evaluation, we leverage Chain-of-Thought reasoning to identify fine-grained sub-causes of hallucination within each task. Extensive evaluations reveal that current MLLMs remain notably vulnerable to cause-specific hallucination triggers, demonstrating the value of ReactBench as a systematic and interpretable testbed for diagnosing and improving multimodal model robustness. The project page is available at https://reactbench.github.io/.

Multimodal Large Language Models (MLLMs) exacerbate safety risks by introducing vulnerabilities across multiple modalities, such as language and vision. Current MLLM safety evaluation tools, however, suffer from major limitations: 1) English-centric dataset construction, and 2) a focus on generic risks that are not tied to local cultural contexts. This paper introduces KSAFE-MM, a benchmark for Korean multimodal safety evaluation that covers both general safety risks and culture-specific vulnerabilities. KSAFE-MM consists of two parts, KSAFE-MM-G and KSAFE-MM-C. KSAFE-MM-G evaluates globally shared risks in Korean contexts through linguistic contextualization, which transforms generic safety queries into contextually grounded multimodal samples. KSAFE-MM-C targets culture-dependent MLLM safety vulnerabilities using localized visual queries derived from real-world contexts. It pairs these visual queries with jailbreak-style textual queries to cover multimodal safety risks involving cultural visual cues and malicious textual intent. Together, these components provide a general-to-local construction pipeline for evaluating both globally shared safety risks and culture-specific vulnerabilities. We evaluate 12 state-of-the-art MLLMs on KSAFE-MM and reveal that models exhibit greater vulnerability to culturally grounded attacks than to generic ones. Notably, jailbreaking strategies substantially amplify attack success rates, with ProgramExecution yielding up to 74.2% ASR compared to 13.4% for standard queries. Furthermore, we identify a systematic trade-off between safety and over-refusal, where models achieving low ASR tend to exhibit excessive refusal behavior on benign queries. These findings highlight the urgent need for culturally grounded safety evaluation beyond English-centric benchmarks.

With the rapid progress of large language models (LLMs), reliably evaluating the capabilities of pre-trained LLMs has become increasingly important. The challenge is that base pre-trained models are optimized for next-token prediction and often fail to follow instructions or produce well-formed answers under standard prompting and direct decoding. As a result, benchmark performance can conflate model capability with decoding-induced failures to produce task-oriented outputs, while exposing such behavior often relies on costly post-training. Recent decodingonly approaches attempt to reshape output distributions, but such methods can be inefficient and brittle across open-ended tasks. To address these limitations, we propose Energy-Based Decoding (EBD), a training-free, reward-guided framework for activating task-oriented behaviors from frozen pre-trained LLMs across both open-ended and objective tasks. EBD augments decoding with an external lightweight reward model, steering generations toward high-utility responses while anchoring them to the pre-trained model prior through a reward-tilted target distribution. We show that EBD shifts base-model outputs toward more instructionfollowing behavior, increasing behavioral similarity to post-trained counterparts and enabling a fairer inference-time evaluation of accessible pre-trained-model behavior. Empirically, EBD outperforms baselines across five models and six benchmarks, improving Qwen3-8B-Base on AlpacaEval2.0 from 8.8 to 44.5, reducing Mistral-7B Math500 latency by 18.9x relative to prior decoding work, and remaining robust to reward-model size.

Parameter-efficient finetuning (PEFT) has become the standard approach for adapting large language models, yet evaluations largely emphasize downstream accuracy while overlooking the retention of pretrained capabilities. We argue that PEFT should be assessed through the stability-plasticity dilemma: the trade-off between target-task adaptation and resistance to forgetting. We introduce PEFT-Arena, a benchmark that jointly measures downstream performance and general capability retention. Across methods, we find distinct stability-plasticity profiles; under comparable parameter budgets, orthogonal finetuning achieves the most favorable Pareto frontier. To explain these differences, we analyze PEFT updates from two geometric perspectives. In weight space, spectral analysis reveals how parameterizations interact with the pretrained singular-value structure. In activation space, retention metrics show whether finetuning preserves or distorts general-capability representations, with forgetting linked to non-isometric representation distortion. Finally, an analysis shows that final SFT checkpoints often overshoot a better target-retention operating point. Inspired by this, we present case studies of a post-hoc improvement with path-wise rewinding.

Retrieval-augmented generation (RAG) systems are often compared by asking a large language model (LLM) judge which answer is better. For multi-hop RAG, this has become a measurement problem as much as a modeling problem: the same score can reflect retrieval quality, answer length, lexical overlap, or a statistical test that ignores clustered data. We ask what happens when these choices are made explicit. We propose a minimum measurement standard for LLM-as-a-judge comparisons in RAG. The standard fixes the top-100 candidate pool, evidence budget, answer cap, generator, and prompt; it also requires pre-registered hypotheses, cluster-aware inference, an exact cluster sign-flip check when feasible, and second-judge replication. Clustered benchmarks can overstate progress; the field should adopt this standard. We stress-test it with Genetic Algorithm Decoder for Multi-hop Evidence Composition (GADMEC), an evolutionary evidence selector, on 400 multi-hop questions in computer science/machine learning (CS/ML) and Materials Science. The protocol changes the empirical story. A binomial test makes all four semantic-baseline comparisons look significant; cluster-aware inference leaves only one Bonferroni-significant result. BM25 beats pure semantic GADMEC under the same budget, while a lexical-semantic hybrid recovers in CS/ML and narrows the Materials Science gap.

We introduce a testing-by-betting framework that leverages predictions on unlabeled data to enhance the power of sequential hypothesis testing. Given limited samples from the joint distribution of $(X,Y)$, and additional unlabeled samples from the marginal of $X$, we ask how unlabeled data can be used to hypothesize about the distribution of $Y$, and the conditional distribution of $Y\mid X$. We introduce an e-statistic and use it to construct a sequential test. Under standard distributional assumptions -- label shift or concept shift -- we establish that the test is anytime valid. Furthermore, we show that for binary data, the e-statistic has non-trivial power. Crucially, our approach retains these properties even when the underlying predictions are inaccurate. Through simulations and applications to large language models evaluation, we demonstrate power gains over baseline approaches, including prediction-powered inference. These gains persist even with relatively limited unlabeled data and when predictions have low accuracy due to weak correlation between $X$ and $Y$.

Federal agencies are deploying large language models (LLMs) to categorize public comment corpora, where the model's organization of the record shapes what policymakers see and which arguments register. Standard evaluation, anchored on stance accuracy against a small validated set, cannot detect when different models produce materially different categorizations of the same public input. We propose an Interpretive Audit Pipeline that treats multi-model disagreement as diagnostic of interpretive complexity and directs human review toward genuinely ambiguous public input. Analyzing 1,260 public comments on a federal USDA docket across four LLMs, we find that inter-model thematic divergence exceeds within-model prompt variation, and that an expert rubric suppresses deep interpretive disagreement without resolving it. In a two-stage labeling study on a stratified 40-comment subsample, four LLMs and a human annotator labeled independently and then revised after seeing the others'labels. Revision behavior varied across labelers, and the human annotator's revisions frequently introduced framings absent from the ensemble's collective output. We argue disagreement-based evaluation is a necessary complement to accuracy metrics for LLM-assisted interpretive coding.

Large language models (LLMs) have shown strong empirical gains as self-evolving agents for CUDA kernel generation, driven by feedback-conditioned planning across generations. However, how planning decisions attribute and combine heterogeneous feedback signals remains opaque. Standard end-to-end ablations fail to resolve this question, as iterative planning amplifies early perturbations and conflates feedback effects with trajectory-dependent drift. We introduce \texttt{CUDAnalyst}, a unified analysis layer for controlled, generation-level attribution of planning decisions to feedback components via trajectory freezing and selective feedback injection. \texttt{CUDAnalyst} enables stable generation-level evaluation and principled coalitional-style attribution of feedback effects and interactions. Our results show that explicit planning is beneficial only when feedback is aligned, that effective planning emerges from structured multi-feedback interactions, and that high-level plans from stronger reasoning models can partially transfer to weaker ones. These trends hold across reference backbones, representative workloads, and reference induction regimes, indicating that the identified feedback-to-plan structure is robust within the controlled axes studied.

Multimodal models in oncology can produce accurate predictions, but accurate prediction does not reveal whether the model has learned biology that is shared across modalities, biology confined to one modality, or spurious correlations that reflect confounders rather than genuine biology. We introduce DECAT, a model-agnostic post-hoc evaluation framework that classifies multimodal representations into four diagnostic scenarios for a given task and modality, using five null-referenced metrics and a rule-based decision procedure. The framework operates on learned representations, requires no knowledge of which specific confounder is present, and returns indeterminate when the evidence is insufficient. We validate DECAT on synthetic data across four multimodal model classes (over 2,500 trained representations) and on real data from 8,979 TCGA patients, evaluating both multimodal embeddings and five pretrained pathology foundation models. Entangled models (e.g., CLIP) achieve near-perfect shared biology detection but falsely claim shared biology in the majority of cases where it is absent on real foundation model embeddings. This false claim rate increases with confound strength so that larger cohorts and stronger representations produce more confident but still incorrect diagnoses. Applied to both multimodal TCGA embeddings and five pathology foundation models without paired RNA, DECAT detects confounding invisible to AUROC without requiring the confounder labels, as confirmed by post-hoc stratification.

Multimodal large language models (MLLMs) have shown strong potential as embodied agents, yet embodied geo-localization remains underexplored due to the lack of fine-grained evaluation. We introduce ERGeoBench, a diagnostic benchmark for vision-driven embodied geo-localization. ERGeoBench evaluates models under three progressive settings -- single-view, panorama-view, and embodied-view -- where agents may actively acquire observations through sequential changes in yaw, pitch, and zoom. The benchmark contains 2,207 globally distributed street-view panoramas and measures four complementary capabilities: foundational perception, spatial awareness, common sense reasoning, and geo-localization reasoning. Evaluations of leading proprietary and open-source MLLMs show that current models can infer high-level geographic semantics, but still struggle with fine-grained perceptual operations, metric localization, and spatial consistency across views. We further observe that geo-localization is strongly correlated with the other capability dimensions, suggesting that accurate localization depends on integrated perception, spatial reasoning, and commonsense inference rather than isolated visual recognition. Overall, ERGeoBench provides a unified framework for diagnosing and advancing human-like embodied geo-localization. Project Page: https://kaixuewen.github.io/ERGeoBench/

Relational reasoning lies at the heart of intelligence, but existing benchmarks are typically confined to formats such as grids or text. We introduce GraphARC, a benchmark for abstract reasoning on graph-structured data. GraphARC generalizes the few-shot transformation learning paradigm of the Abstraction and Reasoning Corpus (ARC). Each task requires inferring a transformation rule from a few input-output pairs and applying it to a new test graph, covering local, global, and hierarchical graph transformations. Unlike grid-based ARC, GraphARC instances can be generated at scale across diverse graph families and sizes, enabling systematic evaluation of generalization abilities. We evaluate state-of-the-art language models on GraphARC and observe clear limitations. Models can answer questions about graph properties but often fail to solve the full graph transformation task, revealing a comprehension-execution gap. Performance further degrades on larger instances, exposing scaling barriers. More broadly, by combining aspects of node classification, link prediction, and graph generation within a single framework, GraphARC provides a promising testbed for future graph foundation models.

Speech foundation models and Speech LLMs have advanced speech understanding, yet deployment-oriented model selection is hindered by non-comparable evaluations caused by mismatched post-processing, and by training results that are hard to reproduce across data scales and pipelines. We present SURE, a unified experimentation framework that standardizes prediction formats, normalization, and scoring. SURE evaluates strong systems across paradigms, from conventional pipelines to Speech LLMs, on representative tasks under realistic acoustic and linguistic stressors. Beyond evaluation, SURE introduces an agent-assisted training conversion flow that maps paper and code into versioned, runnable training pipelines under a unified protocol on matched open-data subsets. Overall, SURE improves comparability and reproducibility for deployment-oriented evaluation.

Scientific discovery is an inherently creative and uncertain process, requiring reasoning beyond the recall of known knowledge. While many benchmarks have been proposed to evaluate large language model (LLM) performance on deep research tasks via multi-hop retrieval, their innovative reasoning abilities essential for true scientific discovery remain largely untested. We introduce a benchmark framework for evaluating model performance in scientific discovery and reasoning, building up from a raw problem to the classical null hypothesis test. In our framework, models initially receive only the topic and research question from a recent paper, with technical details progressively revealed. At each stage of information disclosure, the model is tasked with generating hypotheses that address the research question, which is compared with the conclusions from the original paper and evaluated via automated semantic similarity of constituent atomic claims. This progressive evaluation of semantic divergence from ground-truth conclusions enables assessment of a model's innovativeness (under minimal information) to grounded reasoning capabilities (under full experimental details), both critical for using LLMs for scientific discovery purposes. Our framework provides a foundation for systematically evaluating scientific reasoning and discovery capabilities in LLMs, crucial for advancing the development of next-generation AI scientist/co-scientist systems. Specifically, here we evaluate GPT-5, GPT-5.4, Gemini 2.5 pro, and Gemini 3.1 pro preview across 45 papers spanning bioactive materials, mechanical materials, and nanomaterials. We find that GPT-5.4 and Gemini 3.1 pro outperform their previous generation counterparts as expected, and GPT-5.4 in particular maintains 0.7 F1 score alignment with ground truth conclusions even under minimal context.

Public numeric benchmarks appear in pretraining, so an evaluation that conditions on a date may be measuring memorized recall rather than out-of-sample skill. We introduce NumLeak, a measurement framework that combines API-boundary probes on production models with a white-box controlled validation on an open causal LM. Top-tier frontier LLMs recall the Fama-French market excess return at 3-seed pooled Pearson r=0.97-0.99 while staying within 0.15 within-25bps on the five sibling factors; comparable fidelity appears on U.S. unemployment, CPI inflation, and NOAA temperature. On a recent-release holdout, parse rate collapses to 21-57% but r stays at approximately 0.99 on months answered, the refuse-or-recall asymmetry a memorized channel predicts. The white-box experiment reproduces the dose-response, and logprob ranking detects memorization that open-ended generation misses, implying closed-API black-box probes understate the channel. A Sonnet"date to market-sentiment"regression that correlates with true Mkt-RF at r=0.74 collapses to r=0.02 once the model's own recall is residualized out. A one-line system-prompt defense blocks 99.8% of a non-adaptive single-turn suffix attack set at near-zero utility cost on conceptual and historical-narrative queries

Natural human conversation is full-duplex and audio-visual: people simultaneously speak and listen while continuously interpreting and producing nonverbal cues, such as nods, smiles, and gestures. To support successful human-agent interaction, agents must model full-duplex audiovisual conversation; however, existing full-duplex benchmarks evaluate only speech. In this work, we present VideoFDB, the first benchmark to evaluate full-duplex audio-visual-to-audio-visual (AV2AV) conversational agents. VideoFDB contributes (i) 237 dyadic clips spanning 11 nonverbal conversational dynamics from real-world video calls, (ii) a taxonomy separating perception from generation behaviors, and (iii) a rubric-based LM-as-judge evaluation framework with interpretable axes for assessing conversational quality with respect to nonverbal conversational dynamics. Across open- and closed-source vision-speech agents, we find systematic failure modes: captioning collapse and visual-stream ignorance, and we show that current systems exploit vision for explicit visual question answering but not for the streaming joint audiovisual grounding required in natural conversation. We further evaluate cascaded speech-to-avatar systems and find that their architecture fundamentally precludes the production of full-duplex nonverbal cues. As the first benchmark for full-duplex AV2AV interaction, VideoFDB establishes a foundation for systematic evaluation and, we hope, will accelerate the advancement and development of next-generation multimodal conversational agents.

Reliable probability estimates are critical in many machine learning applications, yet modern classifiers are often poorly calibrated. Post-hoc calibration provides a simple and widely used solution, but the large number of proposed methods, combined with small-scale and inconsistent evaluations, makes it difficult to determine which approaches are truly effective in practice. We introduce a large-scale, standardized benchmark for post-hoc calibration, covering nearly 2000 experiments across tabular and computer vision tasks, including binary, multiclass, and large-scale classification settings. Our benchmark aggregates predictions from a diverse set of classical models, modern deep learning architectures, and foundation models, and provides unified, reproducible implementations of dozens of calibration methods within a common evaluation framework. We argue that Post-Hoc Improvement (PHI) in proper scoring rules offers a principled alternative to traditional calibration error estimators for comparing post-hoc methods, capturing both calibration quality and potential degradation to the model's predictive performance. Using this framework, we conduct the most comprehensive empirical study of post-hoc calibration to date. Our results reveal consistent patterns across domains: smooth calibration functions outperform binning-based approaches, dedicated multiclass methods are essential in high-dimensional settings, and generic machine learning models are not competitive without calibration-specific design. To facilitate future research, we release all data, code, and evaluation tools, providing a plug-and-play benchmark for developing and comparing calibration methods.

Current alignment paradigms for generative artificial intelligence rely predominantly on monolithic benchmarking frameworks that reduce the plurality of human judgment to aggregated statistical baselines, thereby obscuring cultural, demographic, and contextual variability in evaluation. We introduce a state-space constrained emulation framework for AI evaluation that replaces singular assessment functions with a structured manifold of synthetic cognitive profiles representing diverse human perspectives. We show that modern generative architectures can instantiate and maintain these evaluative personas with high consistency, enabling a form of pluralistic, perspective-dependent benchmarking that more closely reflects real-world consensus variability. However, we further analyze the stability of these simulated evaluators under sequential inference and stochastic prompt perturbations, revealing systematic degradation in persona coherence that manifests as state-space drift and semantic inconsistency. These findings suggest that static alignment constraints are insufficient for sustaining robust evaluative behavior over time. Instead, we argue for the necessity of embedding dynamic, viability-driven regulatory mechanisms within generative systems to preserve coherent cognitive emulation. By framing persona-based evaluation as a structured dynamical system over latent representation manifolds, this study provides a foundation for more adaptive, human-aligned, and context-sensitive approaches to AI evaluation.

The rapid proliferation of Generative AI necessitates rigorous documentation standards for transparency and governance. However, manual creation of Model and Data Cards is not scalable, while automated approaches lack large-scale, high-fidelity benchmarks for systematic evaluation. We introduce MetaGAI, a comprehensive benchmark comprising 2,541 verified document triplets constructed through semantic triangulation of academic papers, GitHub repositories, and Hugging Face artifacts. Unlike prior single-source datasets, MetaGAI employs a multi-agent framework with specialized Retriever, Generator, and Editor agents, validated through four-dimensional human-in-the-loop assessment, including human evaluation of editor-refined ground truth. We establish a robust evaluation protocol combining automated metrics with validated LLM-as-a-Judge frameworks. Extensive analysis reveals that sparse Mixture-of-Experts architectures achieve superior cost-quality efficiency, while a fundamental trade-off exists between faithfulness and completeness. MetaGAI provides a foundational testbed for benchmarking, training, and analyzing automated Model and Data Card generation methods at scale. Our data and code are available at: https://github.com/haoxuan-unt2024/MetaGAI-Benchmark.

Large language models (LLMs) are becoming increasingly capable at small parameter scales. At the same time, conventional cloud-centric deployment introduces challenges around data privacy, latency, and cost that are acute in operational technology and defence environments. Advances in model distillation, quantisation, and affordable edge accelerators now make local LLM inference on single-board computers feasible, but the high dimensionality of the configuration space makes identifying optimal deployments difficult without structured evaluation. Existing LLM-specific edge benchmarking efforts rely on CPU-only inference, poor coverage of genuine single-board computers, and generic evaluation tasks that lack multi-dimensional assessment of hardware effectiveness. This paper proposes a multi-dimensional benchmarking methodology that jointly evaluates inference performance and hardware efficiency across four IoT-suitable edge platform configurations testing single-board computers with the latest available hardware accelerators. Our results reveal the benefits of using hardware accelerators such as NPUs and GPUs, along with multi-dimensional evaluations quantifying the trade-offs between power efficiency, physical device size and token throughput; offering practical guidance for deploying generative AI in privacy-sensitive and connectivity-limited environments such as unmanned vehicles and portable, ruggedised operations.

The rapid adoption of large language models (LLMs) in education raises profound challenges for assessment design. To adapt assessments to the presence of LLM-based tools, it is crucial to characterize the strengths and weaknesses of LLMs in a generalizable, valid and reliable manner. However, current LLM evaluations often rely on descriptive statistics derived from benchmarks, and little research applies theory-grounded measurement methods to characterize LLM capabilities relative to human learners in ways that directly support assessment design. Here, by combining educational data mining and psychometric theory, we introduce a statistically principled approach for identifying items on which humans and LLMs show systematic response differences, pinpointing where assessments may be most vulnerable to AI misuse, and which task dimensions make problems particularly easy or difficult for generative AI. The method is based on Differential Item Functioning (DIF) analysis -- traditionally used to detect bias across demographic groups -- together with negative control analysis and item-total correlation discrimination analysis. It is evaluated on responses from human learners and six leading chatbots (ChatGPT-4o \&5.2, Gemini 1.5 \&3 Pro, Claude 3.5 \&4.5 Sonnet) to two instruments: a high school chemistry diagnostic test and a university entrance exam. Subject-matter experts then analyzed DIF-flagged items to characterize task dimensions associated with chatbot over- or under-performance. Results show that DIF-informed analytics provide a robust framework for understanding where LLM and human capabilities diverge, and highlight their value for improving the design of valid, reliable, and fair assessment in the AI era.

The deployment of Multimodal Large Language Models (MLLMs) in agriculture is currently stalled by a critical trade-off: the existing literature lacks the large-scale agricultural datasets required for robust model development and evaluation, while current state-of-the-art models lack the verified domain expertise necessary to reason across diverse taxonomies. To address these challenges, we propose the Vision-to-Verified-Knowledge (V2VK) pipeline, a novel generative AI-driven annotation framework that integrates visual captioning with web-augmented scientific retrieval to autonomously generate the AgriMM benchmark, effectively eliminating biological hallucinations by grounding training data in verified phytopathological literature. The AgriMM benchmark contains over 3,000 agricultural classes and more than 607k VQAs spanning multiple tasks, including fine-grained plant species identification, plant disease symptom recognition, crop counting, and ripeness assessment. Leveraging this verifiable data, we present AgriChat, a specialized MLLM that presents broad knowledge across thousands of agricultural classes and provides detailed agricultural assessments with extensive explanations. Extensive evaluation across diverse tasks, datasets, and evaluation conditions reveals both the capabilities and limitations of current agricultural MLLMs, while demonstrating AgriChat's superior performance over other open-source models, including internal and external benchmarks. The results validate that preserving visual detail combined with web-verified knowledge constitutes a reliable pathway toward robust and trustworthy agricultural AI. The code and dataset are publicly available at https://github.com/boudiafA/AgriChat .

Multi-party conversation generation, such as smart reply and collaborative assistants, is an increasingly important capability of generative AI, yet its evaluation remains a critical bottleneck. Compared to two-party dialogue, multi-party settings introduce distinct challenges, including complex turn-taking, role-dependent speaker behavior, long-range conversational structure, and multiple equally valid continuations. Accordingly, we introduce MPCEval, a task-aware evaluation and benchmarking suite for multi-party conversation generation. MPCEval decomposes generation quality into speaker modeling, content quality, and speaker--content consistency, and explicitly distinguishes local next-turn prediction from global full-conversation generation. It provides novel, quantitative, reference-free, and reproducible metrics that scale across datasets and models. We apply MPCEval to diverse public and real-world datasets and evaluate modern generation methods alongside human-authored conversations. The results reveal systematic, dimension-specific model characteristics in participation balance, content progression and novelty, and speaker--content consistency, demonstrating that evaluation objectives critically shape model assessment and that single-score evaluation obscures fundamental differences in multi-party conversational behavior. The implementation of MPCEval and the associated evaluation code are publicly available at https://github.com/Owen-Yang-18/MPCEval.

As generative AI technologies are increasingly being launched across the globe, assessing their competence to operate in different cultural contexts is exigently becoming a priority. While recent years have seen numerous and much-needed efforts on cultural benchmarking, these efforts have largely focused on specific aspects of culture and evaluation. While these efforts contribute to our understanding of cultural competence, a unified and systematic evaluation approach is needed for us as a field to comprehensively assess diverse cultural dimensions at scale. Drawing on measurement theory, we present a principled framework to aggregate multifaceted indicators of cultural capabilities into a unified assessment of cultural intelligence. We start by developing a working definition of culture that includes identifying core domains of culture. We then introduce a broad-purpose, systematic, and extensible framework for assessing cultural intelligence of AI systems. Drawing on theoretical framing from psychometric measurement validity theory, we decouple the background concept (i.e., cultural intelligence) from its operationalization via measurement. We conceptualize cultural intelligence as a suite of core capabilities spanning diverse domains, which we then operationalize through a set of indicators designed for reliable measurement. Finally, we identify the considerations, challenges, and research pathways to meaningfully measure these indicators, specifically focusing on data collection, probing strategies, and evaluation metrics.

Generative AI is widely used to create commercial posters. However, rapid advances in generation have outpaced automated quality assessment. Existing models emphasize generic esthetics or low level distortions and lack the functional criteria required for e-commerce design. It is especially challenging for Chinese content, where complex characters often produce subtle but critical textual artifacts that are overlooked by existing methods. To address this, we introduce E-comIQ-ZH, a framework for evaluating Chinese e-commerce posters. We build the first dataset E-comIQ-18k to feature multi dimensional scores and expert calibrated Chain of Thought (CoT) rationales. Using this dataset, we train E-comIQ-M, a specialized evaluation model that aligns with human expert judgment. Our framework enables E-comIQ-Bench, the first automated and scalable benchmark for the generation of Chinese e-commerce posters. Extensive experiments show our E-comIQ-M aligns more closely with expert standards and enables scalable automated assessment of e-commerce posters. All datasets, models, and evaluation tools will be released to support future research in this area.Code will be available at https://github.com/4mm7/E-comIQ-ZH.

AI video generation is evolving rapidly. For video generators to be useful for applications ranging from robotics to film-making, they must consistently produce realistic videos. However, evaluating the realism of generated videos remains a largely manual process -- requiring human annotation or bespoke evaluation datasets which have restricted scope. Here we develop an automated evaluation framework for video realism which captures both semantics and coherent 3D structure and which does not require access to a reference video. Our method, 3DSPA, is a 3D spatiotemporal point autoencoder which integrates 3D point trajectories, depth cues, and DINO semantic features into a unified representation for video evaluation. 3DSPA models how objects move and what is happening in the scene, enabling robust assessments of realism, temporal consistency, and physical plausibility. Experiments show that 3DSPA reliably identifies videos which violate physical laws, is more sensitive to motion artifacts, and aligns more closely with human judgments of video quality and realism across multiple datasets. Our results demonstrate that enriching trajectory-based representations with 3D semantics offers a stronger foundation for benchmarking generative video models, and implicitly captures physical rule violations. The code and pretrained model weights will be available at https://github.com/TheProParadox/3dspa_code.

Recent advancements in image generation models have enabled the prediction of future Graphical User Interface (GUI) states based on user instructions. However, existing benchmarks primarily focus on general domain visual fidelity, leaving the evaluation of state transitions and temporal coherence in GUI-specific contexts underexplored. To address this gap, we introduce GEBench, a comprehensive benchmark for evaluating dynamic interaction and temporal coherence in GUI generation. GEBench comprises 700 carefully curated samples spanning five task categories, covering both single-step interactions and multi-step trajectories across real-world and fictional scenarios, as well as grounding point localization. To support systematic evaluation, we propose GE-Score, a novel five-dimensional metric that assesses Goal Achievement, Interaction Logic, Content Consistency, UI Plausibility, and Visual Quality. Extensive evaluations on current models indicate that while they perform well on single-step transitions, they struggle significantly with maintaining temporal coherence and spatial grounding over longer interaction sequences. Our findings identify icon interpretation, text rendering, and localization precision as critical bottlenecks. This work provides a foundation for systematic assessment and suggests promising directions for future research toward building high-fidelity generative GUI environments. The code is available at: https://github.com/stepfun-ai/GEBench.

Generative text-to-image models are advancing at an unprecedented pace, continuously shifting the perceptual quality ceiling and rendering previously collected labels unreliable for newer generations. To address this, we present ELIQ, a Label-free Framework for Quality Assessment of Evolving AI-generated Images. Specifically, ELIQ focuses on visual quality and prompt-image alignment, automatically constructs positive and aspect-specific negative pairs to cover both conventional distortions and AIGC-specific distortion modes, enabling transferable supervision without human annotations. Building on these pairs, ELIQ adapts a pre-trained multimodal model into a quality-aware critic via instruction tuning and predicts two-dimensional quality using lightweight gated fusion and a Quality Query Transformer. Experiments across multiple benchmarks demonstrate that ELIQ consistently outperforms existing label-free methods, generalizes from AI-generated content (AIGC) to user-generated content (UGC) scenarios without modification, and paves the way for scalable and label-free quality assessment under continuously evolving generative models. The code will be released upon publication.

In recent years, generative artificial intelligence has made significant advances in the design of crystalline materials, giving rise to approaches based on graph neural networks, diffusion models, and large language models. Existing evaluations commonly follow the stability-uniqueness-novelty (S.U.N.) framework, where stability is primarily assessed using thermodynamic criteria, which do not fully capture the dynamical stability essential for a material's practical existence. Dynamical stability is a key determinant of whether a material can be synthesized and persist, with phonon spectrum calculations serving as the standard for its evaluation. However, the high computational cost of such calculations has prevented large-scale assessment of dynamical stability in generated crystals. In this work, we introduce PhononBench, the first large-scale benchmark for dynamical stability in AI-generated crystals. Leveraging the recently developed MatterSim interatomic potential, which achieves density-functional-theory (DFT)-level accuracy in phonon predictions across more than 10,000 materials, PhononBench enables efficient phonon calculations and dynamical-stability analysis for 133,838 crystal structures generated by 7 leading crystal generation models. PhononBench reveals a widespread limitation of current generative models: unless otherwise specified, all reported dynamical-stability metrics are evaluated at a phonon-frequency threshold of -0.1 THz, with the average dynamical-stability rate across all generated structures being only 32.15%, and the top-performing model, MatterGen, reaching just 45.05%.In addition, we identify 32,995 crystal structures that are phonon-stable across the entire Brillouin zone under a strict threshold of -0.001 THz. In addition, a web-based service is accessible at http://phononbench.cn/, enabling minute-level ultra-fast phonon predictions.

Physical AI aims to develop models that can perceive and predict real-world dynamics; yet, the extent to which current multi-modal large language models and video generative models support these abilities is insufficiently understood. We introduce Physical AI Bench (PAI-Bench), a unified and comprehensive benchmark that evaluates perception and prediction capabilities across video generation, conditional video generation, and video understanding, comprising 2,808 real-world cases with task-aligned metrics designed to capture physical plausibility and domain-specific reasoning. Our study provides a systematic assessment of recent models and shows that video generative models, despite strong visual fidelity, often struggle to maintain physically coherent dynamics, while multi-modal large language models exhibit limited performance in forecasting and causal interpretation. These observations suggest that current systems are still at an early stage in handling the perceptual and predictive demands of Physical AI. In summary, PAI-Bench establishes a realistic foundation for evaluating Physical AI and highlights key gaps that future systems must address.

As generative artificial intelligence (AI) continues to transform education, most existing AI evaluations rely primarily on technical performance metrics such as accuracy or task efficiency while overlooking human identity, learner agency, contextual learning processes, and ethical considerations. In this paper, we present TEACH-AI (Trustworthy and Effective AI Classroom Heuristics), a domain-independent, pedagogically grounded, and stakeholder-aligned framework with measurable indicators and a practical toolkit for guiding the design, development, and evaluation of generative AI systems in educational contexts. Built on an extensive literature review and synthesis, the ten-component assessment framework and toolkit checklist provide a foundation for scalable, value-aligned AI evaluation in education. TEACH-AI rethinks"evaluation"through sociotechnical, educational, theoretical, and applied lenses, engaging designers, developers, researchers, and policymakers across AI and education. Our work invites the community to reconsider what constructs"effective"AI in education and to design model evaluation approaches that promote co-creation, inclusivity, and long-term human, social, and educational impact.

While Chain-of-Thought (CoT) prompting enables sophisticated symbolic reasoning in LLMs, it remains confined to discrete text and cannot simulate the continuous, physics-governed dynamics of the real world. Recent video generation models have emerged as potential world simulators through Chain-of-Frames (CoF) reasoning -- materializing thought as frame-by-frame visual sequences, with each frame representing a physically-grounded reasoning step. Despite compelling demonstrations, a challenge persists: existing benchmarks, focusing on fidelity or alignment, do not assess CoF reasoning and thus cannot measure core cognitive abilities in multi-step planning, algorithmic logic, or abstract pattern extrapolation. This evaluation void prevents systematic understanding of model capabilities and principled guidance for improvement. We introduce Gen-ViRe (Generative Visual Reasoning Benchmark), a framework grounded in cognitive science and real-world AI applications, which decomposes CoF reasoning into six cognitive dimensions -- from perceptual logic to abstract planning -- and 24 subtasks. Through multi-source data curation, minimal prompting protocols, and hybrid VLM-assisted evaluation with detailed criteria, Gen-ViRe delivers the first quantitative assessment of video models as reasoners. Our experiments on SOTA systems reveal substantial discrepancies between impressive visual quality and actual reasoning depth, establishing baselines and diagnostic tools to advance genuine world simulators.

As Vision-Language Models (VLMs) increasingly gain traction in medical applications, clinicians are progressively expecting AI systems not only to generate textual diagnoses but also to produce corresponding medical images that integrate seamlessly into authentic clinical workflows. Despite the growing interest, existing medical visual benchmarks present notable limitations. They often rely on ambiguous queries that lack sufficient relevance to image content, oversimplify complex diagnostic reasoning into closed-ended shortcuts, and adopt a text-centric evaluation paradigm that overlooks the importance of image generation capabilities. To address these challenges, we introduce MedGEN-Bench, a comprehensive multimodal benchmark designed to advance medical AI research. MedGEN-Bench comprises 6,422 expert-validated image-text pairs spanning six imaging modalities, 16 clinical tasks, and 28 subtasks. It is structured into three distinct formats: Visual Question Answering, Image Editing, and Contextual Multimodal Generation. What sets MedGEN-Bench apart is its focus on contextually intertwined instructions that necessitate sophisticated cross-modal reasoning and open-ended generative outputs, moving beyond the constraints of multiple-choice formats. To evaluate the performance of existing systems, we employ a novel three-tier assessment framework that integrates pixel-level metrics, semantic text analysis, and expert-guided clinical relevance scoring. Using this framework, we systematically assess 10 compositional frameworks, 3 unified models, and 5 VLMs.

Assessing whether Large Language Models outputs are factually grounded, epistemically calibrated, and methodologically reproducible is a prerequisite for responsible AI deployment. Yet auditing LLMs remains inaccessible to non-technical practitioners: existing tools require programming expertise and non-trivial environment setup, and cloud-hosted platforms transmit evaluation data to external services, creating barriers for domain experts and compliance officers legally responsible for AI oversight. We introduce LLM-FACETS (LLM FActuality Cross-EvaluaTion System): an open-source framework with a browser-accessible interface and a plugin architecture, structured around three practitioner profiles (technical experts, domain experts, compliance officers) that mirror the stakeholder categories identified in the EU AI Act and the NIST AI Risk Management Framework. The architecture makes data flows explicit: deterministic metrics (BLEU, ROUGE, BERTScore) run entirely within the self-hosted server with no outbound transmission; LLM-judge metrics contact external APIs explicitly, with users retaining full credential control. The framework operationalizes transparency through three mechanisms: token-level log-probability visualization for epistemic uncertainty, multi-judge consensus to mitigate judge bias, and RAG Triad metrics (Faithfulness, Answer Relevance, Context Relevance) to detect and localize hallucinations. A plugin architecture allows any new metric or dataset to be integrated without modifying the evaluation pipeline. The open-source implementation enables cross-checking across multiple metrics targeting the same property, ensuring reproducibility and decoupling AI accountability from the teams building the systems assessed. We verify the framework through cross-validation of 18 metric implementations against canonical reference libraries.

Two clinical AI systems can score nearly identically on coverage-based rubrics yet behave radically differently when their patient inputs change: one updates its recommendations to match the new clinical signal, while the other produces the same output regardless. We introduce the Causal Sensitivity Score (CSS), a pre-registered interventional metric that mutates oncology tumor-board cases along five clinically meaningful dimensions - biomarker flips, prior-treatment failures, biomarker removals, surgery-status changes, and stage perturbations - and scores whether each model updates its recommendations in the pre-registered correct direction using a {0, 0.5, 1.0} scale. Benchmarked against the Consensus Match Score (CMS), a coverage-based weighted recall metric, six frontier models from three labs evaluated in single-shot inference across 224 cases rank in nearly opposite orders: all six models change rank, the CMS-worst model becomes CSS-best, and one upper-mid CMS model ranks last on CSS. We further surface a universal safety blind spot: every frontier model fails on surgery-status interventions (at most 17.2% CSS on Family D), a finding CMS does not expose. The metric also transfers to tool-using agents: in a ReAct-style experiment, tool use improves CSS for five of six models (+2.5 to +20.3 percentage points), yet the lowest-CSS model retrieves the same chart sections and still fails to update its recommendations - revealing a structural responsiveness deficit visible only under counterfactual evaluation. Cross-judge replication and three-rater medical-professional validation confirm the aggregate findings. Interventional pre-registered metrics like CSS complement coverage-based evaluation for clinical AI agents: they capture responsiveness that coverage metrics miss and offer a candidate dense reward signal for future agentic RL systems.

The growing number of submitted papers has motivated the exploration of Large Language Models (LLMs) as a means to support and augment the peer review process, particularly in terms of improving its speed and scalability. Yet, it remains unknown whether LLMs engage with scientific manuscripts in the same manner as human reviewers, or whether they merely produce review-looking text. To address this, we introduce the Peer Review AI Benchmark (PRAIB), a novel framework comprising thoroughly defined metrics that measure review specificity, style, and behavior of engagement. To complement the PRAIB framework, we conduct a large-scale empirical study leveraging a dataset of 11,000 reviews generated by five proprietary and open-source models for 1,000 ICLR and NeurIPS papers. Spanning the 2021--2025 period, these machine-generated reviews are compared against original human feedback across diverse prompting strategies to identify systematic behavioral divergences. Our analysis reveals that the generated reviews diverge significantly from feedback provided by human reviewers: LLM ratings are less variable, positively biased, and overconfident, and their cross-reference patterns are model-dependent and distinct from human norms. Furthermore, when evaluated through PRAIB, we observe that LLMs tend to generate longer, more complex reviews, yet frequently overlook the atomic weaknesses noted by human reviewers. By characterizing where and how LLMs reviewing behavior departs from human norms, PRAIB provides the community with a diagnostic tool for identifying which aspects of the review process LLMs can reliably support today and which require further development before deployment.

Real-world evaluation of vision-language-action (VLA) policies still rests on binary success rate at a fixed timeout with $N \le 25$ rollouts per condition, almost always without confidence intervals or paired statistical comparison; these cohort sizes struggle to resolve close comparisons reliably. We introduce PhAIL (Physical AI Leaderboard, https://phail.ai), an open real-robot benchmark on a Franka FR3 (dataset, per-rollout artifacts, and end-to-end reference implementation) of a distributional evaluation methodology: the time-to-success cumulative distribution function (CDF) as the evaluation primitive, with two separated jobs. The first is scoring via Human-Relative Throughput (HRT), a dimensionless scalar with bootstrap confidence intervals, anchored to same-fixture human teleoperation. The second is a significance test (Kolmogorov-Smirnov, computed per-object and macro-averaged across objects). On four publicly-available VLAs, the macro-averaged KS test resolves two close comparisons (GR00T vs. ACT, OpenPI vs. ACT) at $N \le 30$ rollouts per (model, object) cell where binary-threshold metrics do not; the closest pair (OpenPI vs. GR00T) remains unresolved within our budget. The best evaluated VLA is $\sim 7\times$ slower per operation (RMST ratio) than the human reference.

Jailbreak attacks on multimodal AI systems remain underexplored, even though unsafe image generation can have more severe consequences than unsafe text and current defenses are relatively immature. We introduce PAST2HARM, a simple yet effective adaptive jailbreak framework that bypasses refusal training in state of the art multimodal text to image models. Building on prior findings that past tense reformulations can evade safeguards, PAST2HARM systematically exploits this vulnerability in multimodal generative AI. We characterize the attack along two dimensions. First, breadth: through temporal deepening, the framework incrementally strengthens historical anchoring and archival cues, eroding refusal boundaries across models with varying alignment strength. Second, depth: via iterative escalation after initial compliance, we probe the upper bound of harmful generation, measuring severity using a scalar severity jailbreak metric evaluated by a language model acting as a judge. We find that mid conversation turns form peak vulnerability windows, where harmfulness increases before plateauing and eventually undergoing semantic inversion. We evaluate PAST2HARM on three models Gemini Nano Banana Pro, GPT Image 2, and SD XL achieving attack success rates of 83 percent, 67 percent, and 100 percent in a black box, gradient free setting. Adversarial prompts also transfer across models, with cross model success rates above 50 percent. The attack elicits diverse harmful outputs, including explicit sexual content, political disinformation, historical denial narratives, hate speech, and self harm glorification. We further release a curated benchmark of prompts, reformulations, and outputs as a resource for red teaming and alignment. Our results expose fundamental brittleness in current safeguards and highlight the need for stronger multimodal safety training.

Prior work shows that Large Language Models (LLMs) can transform Explainable AI (XAI) outputs into Natural Language Explanations (NLEs) that score highly on quality metrics such as plausibility, coherence, and comprehensibility. But does explanation quality translate to practical usefulness? We investigate this question in a time-series energy forecasting domain through five controlled experiments (2,730 judgments across 60 test instances), each operationalising a distinct facet of usefulness studied in the XAI literature. Holding NLE quality constant at the high levels established by a prior factorial study, we find that NLEs do not improve task accuracy on any of the five tasks, while inflating self-reported confidence. A placebic control shows that this confidence boost is driven by text presence rather than content. In an out-of-distribution detection task, NLEs reduce the LLM judge's ability to flag unreliable predictions, providing false reassurance that masks model failure. We characterise these findings as the Quality-Usefulness Gap and argue that evaluation of the XAI-to-NLE pipeline must extend beyond text-quality metrics to downstream task performance.

Autonomous AI agents increasingly issue side-effect-bearing actions: database mutations, refunds, payments, external commitments. We propose the Actuarial Action Interface (AAI), a deterministic runtime contract that prices each such action against a contractually fixed safe default under a time-consistent risk mapping, and gates execution against a per-boundary reserve capital budget. We then develop the Authority Frontier, an evaluation primitive measuring how much autonomous authority the runtime releases at each level of reserve capital. The framework provides (i) a deterministic quote-bind-commit protocol with toll-bounded capability tokens; (ii) a universal seven-class action taxonomy mapping heterogeneous tool calls to comparable authority units; (iii) replay determinism and pathwise reserve coverage under alpha-spending; (iv) cross-domain normalization via full reserve demand C_full and capital metrics Capital@k. We instantiate AAI across four agentic environments (database mutation, customer-service refund, and the public tau-bench retail and airline tool-use traces) and report a live Postgres panel in which three Azure-hosted models propose actions through the same contract. The frontier exhibits a common low-reserve refusal and intermediate-release pattern across domains, with saturation only where the budget grid reaches full reserve demand; required reserve capital varies by 22x (Capital@50 from 289 to 6457). The framework does not force domains into the same shape; it surfaces each domain's actuarial geometry. In the live panel the contract prevents realized loss across all three models at low budget while differing in underwriting persistence under denial: model identity is an actuarial underwriting variable. The contribution is a benchmark-ready evaluation framework for runtime actuarial control of autonomous-agent side effects.

Accurate evaluation of weather forecasting models is critical for their reliable deployment in real-world applications. However, existing benchmarks predominantly rely on reanalysis products such as ERA5, which are generated through delayed data assimilation and do not reflect the constraints of real-time operational forecasting, thereby resulting in a systematic mismatch between benchmark performance and real-world forecasting. In this work, we introduce RealBench, a next-generation benchmark for AI weather forecasting that emphasizes realistic evaluation under operational conditions. RealBench features a strictly out-of-distribution test set spanning 2025 to eliminate data leakage and capture recent atmospheric regimes. It integrates multiple data sources, including low-latency operational analysis and a large-scale global in-situ observation dataset comprising over 10,000 stations, enabling direct evaluation against real atmospheric measurements. Beyond standard global metrics, RealBench provides a comprehensive evaluation framework for high-impact extreme events, including heatwaves, cold surges, and tropical cyclones, using event-specific metrics that better reflect real-world forecasting priorities. The evaluation results reveal substantial discrepancies between reanalysis-based metrics and real-world performance, particularly concerning extreme events. By highlighting the limitations of existing benchmarks, this work establishes a more faithful and operationally relevant evaluation paradigm, providing a rigorous foundation for advancing next-generation AI weather forecasting systems. The benchmark implementation is available at: https://github.com/lixruize-del/NWP-Benchmark.

Recent advances in multimodal web agents often rely on increased inference-time computation, including rollout search, verifier passes, offline skill discovery, and specialist model stacks. This raises a central question: can a web agent become more efficient as it accumulates experience, rather than more expensive? We first analyze trajectories from VisualWebArena and identify three recurring sources of inefficiency: repeat-action loops, hidden discovery costs, and low prompt-cache reuse. We then introduce PANDO, a single-rollout online skill-distillation framework that maintains a structured Skill Library and combines progress reflection, confidence-based skill demotion, hierarchical routing, visual compression, and cache-aware prompting. On the full set of 910 VisualWebArena tasks, PANDO achieves a 58.3% success rate, outperforming SGV (54.0%) and our WALT reproduction (45.2%), while using 58% fewer tokens than SGV and 61% fewer tokens than WALT, without any pre-evaluation discovery budget. A 300-task ablation further shows that rules and routines provide most of the success gains, while routing, compression, and cache-aware prompting convert the larger skill library into lower marginal token cost. Finally, we introduce three trajectory-level efficiency metrics -- Action Repetition Rate, Step Overhead Ratio, and Prompt Cache Utilization -- to make efficiency visible beyond terminal success.

Current approaches to AI compliance treat conformity as a binary, audit-time verdict rather than a continuous, measurable property of production systems. We argue that this compliance fiction is structurally ill-suited to the requirements of the EU AI Act, which demands ongoing human oversight and the detection of emergent behavioural drift in deployed systems. We introduce governance from metrics, a principle whereby regulatory compliance is derived as a continuous signal from runtime observability rather than from static assessments. Building on this principle, we present govllm, an open-source framework implementing a governance-driven routing architecture in which model selection is determined by accumulated compliance scores rather than by latency or cost alone. Central to our approach is a panel of regulatory judges - LLM evaluators specialised per criterion (EU AI Act, GDPR, ANSSI, accessibility) - whose inter-judge disagreement we reframe not as noise but as a regulatory uncertainty signal warranting human arbitration. We validate this approach through a ground truth corpus of 49 annotated prompt/response pairs across five regulatory criteria, evaluated by four small language models (SLMs, 1.7B-7B parameters) running fully on-premise. Agreement rates range from 51.5% (mistral:7b) to 69.1% (phi4-mini), with no single model dominating across all criteria - empirically motivating the Profile-as-jury design. We further document three structural failure modes in small regulatory judges and a judge-specific position bias that degrades agreement by up to 25 percentage points across three question-order conditions (original, reversed, permuted). govllm is released as open-source software to support reproducible AI governance research.

As AI-powered compliance monitoring becomes increasingly important in public governance and industrial safety, the ability to provide verifiable evidence and traceable accountability signals is essential. However, existing video anomaly detection datasets focus on event-level binary classification, lacking the rule-driven, explainable analysis required for real-world compliance scenarios. We introduce FoodMonitor, a benchmark for explainable compliance analysis in commercial kitchen surveillance. FoodMonitor comprises 477 video clips with 3,307 violation annotations across a dual-channel design covering both person-level and environment-level violations. Each annotation specifies which rule was violated, what non-compliant behavior occurred, and who committed it with frame-level bounding boxes. We establish a unified evaluation protocol with a two-stage matching mechanism that separately assesses spatial localization and semantic understanding, along with a composite metric ($C_{\text{score}}$) that balances environment and person detection performance. Systematic evaluation of several state-of-the-art multimodal large language models reveals that the best-performing model achieves only 0.360 $C_{\text{score}}$, with spatial localization and fine-grained rule understanding emerging as the primary bottlenecks. Our analysis identifies two distinct failure modes: localization-dominated errors and semantics-dominated errors, providing diagnostic insights for future model development.

Large language model (LLM) unlearning has emerged as a crucial post-hoc mechanism for privacy protection and AI safety, yet auditing whether target knowledge is truly erased remains challenging. Existing output-level metrics fail to detect when this knowledge remains recoverable from internal representations. Recent white-box studies reveal such residual knowledge but often rely on auxiliary training or dataset-specific adaptations, leaving no generalizable metric. To address these limitations, we propose the Unlearning Depth Score (UDS), a metric that quantifies the mechanistic depth of unlearning via activation patching. UDS first identifies layers that encode the target knowledge using a retain model baseline, then measures how much of it is erased in the unlearned model on a 0-1 scale. In a meta-evaluation across 20 metrics on 150 unlearned models spanning 8 methods, UDS achieves the highest faithfulness and robustness, confirming our causal approach as the most reliable for unlearning evaluation. Case studies further reveal that white-box metrics can disagree at the layer level and that erasure depth varies across examples. We provide guidelines for integrating UDS into existing benchmarking frameworks and streamlining the evaluation pipeline. Code and data are available at https://github.com/gnueaj/unlearning-depth-score

We evaluate the consistency of automated judges in conducting a multi-dimensional safety evaluation in a reference-free setup. Our results indicate that Large Language Models are unreliable judges in identifying safety issues related to machine-generated advice in regulated domains such as finance, although they are more reliable at identifying more overt forms of unsafe/harmful content such as violence. The degree of inconsistency in a model's judgments can vary significantly by the chosen safety criteria and can be impacted by the language of the content and its linguistic style as well. Finally, there is high disagreement among different judges for the same output, across domains, safety criteria, and languages. These findings provide new insights on the practice of using LLMs as evaluators and offer several recommendations for practitioners on how to use automated judges in practical scenarios.

Autonomous coding agents built on large language models (LLMs) are rapidly being integrated into development workflows, yet their operational safety properties remain poorly understood beyond evaluations of explicitly malicious inputs. In practice, high-impact failures arise during benign, goal-directed use through environment breakage, fabricated success reports, etc. that current benchmarks do not capture. What categories of operational safety failures actually occur when coding agents are used for everyday development tasks and what is their impact? We present an incident-driven empirical study grounded in two complementary evidence streams. We screen 68,816 papers from 22 premier venues, curating 185 safety-relevant studies, and mine 16,586 GitHub issues from widely deployed LLM-powered coding tools, manually confirming 547 genuine safety failures. Applying systematic open coding over both corpora, we derive a multi-dimensional safety taxonomy of 33 operational risk types organized across seven dimensions, and annotate each incident with contributing factors, task context, severity, and downstream impact. Our findings show that coding-agent failures are often severe, with 326 of 547 incidents rated high or critical. The dominant risks are constraint violations, destructive operations, authorization bypasses, and deception, and over 65% of incidents arise in bug fixing and setup or configuration, patterns largely missing from prior literature. These results have direct implications for SE tool designers and benchmark developers: guardrails must go beyond adversarial-prompt defenses to enforce environmental constraints, failure transparency, and safe-halt behaviors.

Prompt-injection detectors are heterogeneous: each is strong on a different slice of attacks, and none is always reliable. Yet existing systems still treat detection as a fixed single-detector pipeline, committing every request to one detector's blind spots. We reframe defense as detector allocation: given a heterogeneous pool, decide per request which detectors to run and whether to escalate to an LLM judge. Our framework SCOUT (Scalable and Controllable Outcome-prediction for Uncertainty-aware Triage) makes this decision dynamic by predicting each detector's per-sample reliability and latency from how it behaved on similar past inputs, and exposes a single safety-utility threshold to the operator (where utility bundles benign-pass rate and wall-clock). To evaluate this setting, we build SCOUT-450, a benchmark that captures the structurally complex, agent-facing injections that older prompt-injection sets under-represent. On SCOUT-450, a safety-oriented operating point reduces attack-success rate by 46% and total wall-clock by 40% relative to an always-on GPT-4o judge, at a 5.1-point benign-utility drop. SCOUT also transfers to three external benchmarks (BIPIA, IPI, and IHEval), improving the safety-utility frontier.

The rise of LLM agents introduces a new threat by enabling planning, coding, and even end-to-end execution of expert-level attack workflows. However, this threat remains underexplored and underestimated since (i) safety alignment prevents LLMs from directly generating harmful instructions, and (ii) most existing jailbreak methods cannot consistently induce agents to execute malicious operations. In this paper, we propose TRACE, a practical agentic jailbreaking framework to further reveal the risks of this threat surface. To conceal the malicious intent, TRACE decomposes a malicious task into multiple subtask sequences under different schemes and selects the sequence with the fewest explicitly harmful subtasks. TRACE then disguises the remaining harmful subtasks as benign-looking instructions by embedding them in task-aware scenarios with related roles, environments, directives, and heuristics. The scenarios are iteratively evolved through well-defined transformation actions, which are sampled by a Q-learning-inspired mechanism, for inducing the agent to execute on the harmful subtasks. Extensive evaluations on AgentHarm and AdvCUA show that TRACE consistently outperforms existing jailbreak baselines across multiple advanced LLM agents, achieving up to 100% bypass rate and 0.73 average success score. We also demonstrate the effectiveness of TRACE in controlled cyberattack instances. Our code and demos are available at https://github.com/ZJU-LLM-Safety/TRACE.git.

Large language models (LLMs) show promise in generating supportive responses for mental health queries, but improving their usefulness, empathy, and safety often requires substantial compute, expert input, and labeled data. At the same time, deploying proprietary, cloud-based models for mental health-related interactions raises important privacy and data-governance concerns, given the sensitivities. To address this challenge, we introduce LLUMI setup that can be hosted in-house within protected environments. LLUMI consists of two complementary components: a generation model (GM), which drafts supportive responses to mental health queries, and an improvement model (IM), which revises an initial human-crafted response. We leverage feedback signals from Reddit mental health communities, using community endorsement patterns such as upvotes and downvotes to construct chosen-rejected response pairs for Supervised Fine Tuning (SFT) and Direct Preference Optimization (DPO). We further align LLUMI using human evaluation across five dimensions: readability, empathy, connection, actionability, and safety. Our results show that, despite relying on smaller open-source models rather than proprietary cloud-based GPT models, LLUMI achieves comparable performance across linguistic analyses and human evaluations. These findings suggest that open-source models, when trained with community-derived preference signals, can support high-quality mental health support assistance while offering a more privacy-preserving alternative for sensitive support contexts.

In this paper, we discuss question-answer dataset for LLM safety evaluation, with a focus on illegal activities. Specifically, on the basis of manual analysis of AnswerCarefully, we introduce several additional information, methods for creating question-answer examples, and a rubric for evaluating LLM-generated responses. The outcomes of this study are intended to be shared with the"JAI-Trust"project.

Jailbreak attacks on large language models (LLMs) aim to induce LLMs to produce content that they are expected to refuse. Automated black-box jailbreak generation is especially important for safety evaluation, where the attacker observes only model outputs and needs to automatically search for effective adversarial prompts. Existing black-box jailbreak methods either depend on sample-wise heuristic search or leverage attack experience through accumulating strategy pools or method libraries, lacking a systematic organization and management of attack experience. To mitigate these drawbacks, we propose MemoAttack, a memory-driven black-box jailbreak framework with comprehensive attack memory modeling, evolution, and selection. Specifically, MemoAttack comprises three key designs: (1) Skill-Structured Memory Modeling, which abstracts accumulated attack experience into reusable skill-structured attack memory whose units pair attack skills with templates, evidence, and lifecycle state; (2) Lifecycle-Driven Memory Evolution, which evolves the memory through evidence-based probation, promotion, retirement, reactivation, elimination, and storage cleanup; and (3) Explore-Exploit Balanced Memory Selection, which balances reliable memory reuse with uncertainty-driven exploration via contextual Thompson Sampling. Experiments on AdvBench demonstrate that MemoAttack achieves an average attack success rate of 98.00%, outperforming the strongest baseline by 16.67 percentage points, while reducing request count by 45.9%. Moreover, MemoAttack continuously improves as memory accumulates over more samples.

Language models are increasingly being deployed for conversational support in informal caregiving contexts, where interactions often extend beyond information-seeking: caregivers seek emotional reassurance, guidance, and help, while navigating uncertain, relationally complex care decisions. Yet most safety evaluations assess model behavior under generic prompts, leaving a critical question unexamined: does a model's safety profile change with its support role? We study this by operationalizing four expert-reviewed support roles grounded in social support theory: Inform, Coach, Relate, and Listen, and comparing them against two baseline controls: a basic prompting condition and a retrieval-augmented generation (RAG) condition. We evaluate across three language models (GPT-4o-mini, Llama-3.1-8B-Instruct, and MedGemma-1.5-4b-it) on 5,000 real-world queries from online Alzheimer's Disease and Related Dementias (ADRD) communities. We find that the LLM's support role systematically shapes both the prevalence and composition of interactional risks. Furthermore, a human evaluation study reveals a perceived quality--safety tension: more directive, information-oriented roles are rated as more helpful and trustworthy despite exhibiting elevated interactional risk profiles. We release ~90,000 support role-conditioned model responses with risk annotations as an ecologically grounded resource for research on safer LLM-mediated conversational support.

Real-time safety filtering for large language model (LLM) applications requires classifiers that can detect unsafe prompts, toxic language, jailbreak attempts, and unsafe responses without the cost profile of large guardrail models, and that can distinguish benign sensitive text from genuinely covert harmful content. In this paper, we introduce Opir, a family of encoder-based guardrail models built on the GLiClass architecture. Opir includes multi-task models for binary safe/unsafe classification, multi-label toxicity classification, jailbreak classification, and zero-shot unsafe prompt and response categorization. We also release edge variants with fewer than 100M parameters dedicated to binary safe/unsafe categorization. The models are trained on a three-level taxonomy containing 996 categories across 16 top-level labels, 126 mid-level labels, and 854 leaf labels. Opir's training data combines taxonomy-grounded unsafe prompts, adversarially mined hard negatives, benign safety-preserving examples, generated response examples, multilingual translations, and portions of the Aegis2 and WildGuard training subsets. We also open-sourced an evaluation harness that supports GLiClass and GLiNER2 backends as well as decoder-based models, and covers binary safety classification, multi-label categorization, toxicity, jailbreak detection, prompt safety, response safety, response refusal, and prompt subcategory views across public benchmark families. Across an expanded comparison spanning 12 safety-classification tasks and 17 category tasks against eight contemporary guardrail systems -- including both GLiNER2-based and generative guardrail models -- Opir variants are competitive on or ahead of the strongest open-weight baselines on the majority of benchmark datasets while operating with a substantially smaller deployment footprint.

Large language models (LLMs) are increasingly used as conversational partners for companionship, emotional disclosure, and interpersonal advice, but the social dynamics of these interactions can create harms that are not captured by capability-oriented or traditional safety evaluations. We introduce the Social AI Design Code, a framework for evaluating whether LLMs align with user welfare in social interactions, including whether they encourage harmful intimacy, dependence, or prolonged engagement. To evaluate these risks in natural and diverse user-LLM interactions, we operationalize the code with EUDAIMONIA, a benchmark of 969 user inputs and 3,147 design-requirement violation checks built from WildChat through weak-to-strong filtration, multi-model relabeling, and controlled rewriting. Evaluating 22 recent LLMs, we find that even the strongest models, Claude-Opus-4.7 and GPT-5.5, violate 30.7% and 27.2% of checks, respectively. Extended thinking does not reduce violation rates, suggesting that these failures are persistent social-alignment problems rather than deficits solvable through test-time reasoning alone.

AI agents augment large language models with external tools such as web retrieval, enabling grounded and up-to-date responses. However, incorporating external content into the generation pipeline can weaken the safety alignment mechanisms that govern model outputs. Prior work shows that enabling retrieval in agents increases compliance with harmful requests. We introduce AgentREVEAL, a diagnostic framework for analyzing retrieval-induced safety degradation in LLM agents. The framework examines two axes: how retrieval is integrated into the agent pipeline and the properties of the retrieved content. Along the integration axis, we find that binding tool invocation and response generation in a single step amplifies harmful outputs. Along the content axis, we uncover the Safe Source Paradox: even oppositional or safety-oriented sources, such as pages containing warnings or risk disclaimers, can increase harmful compliance by an average of 25% compared to the no-retrieval baseline. Finally, we show that relevance acts as a shared activation condition for both vulnerabilities. Similar patterns appear on frontier closed models, and harmful compliance remains elevated under several representative pipeline interventions, with some agents also entering this regime under autonomous retrieval. Because relevance is also what makes retrieval useful, these results expose a safety-utility trade-off for retrieval-enabled agents. We introduce HarmURLBench, a benchmark containing 1,405 real-world URLs paired with 320 harmful behaviors to support future evaluations.

Large language models (LLMs) frequently achieve impressive scores on standardized benchmarks, yet accuracy alone offers a limited view of their capabilities. Evaluating open-source LLMs through leaderboards faces persistent issues like data contamination, narrow task scope, and weak alignment with real-world reliability. Benchmark-based evaluations such as MMLU PRO, BBH, or IFEval primarily capture what a model outputs on fixed test sets, not how it processes information, calibrates uncertainty, or structures internal knowledge. In this article, we advocate for a shift from benchmark-centric evaluation toward a complementary, state-centered intrinsic assessment of LLMs. To this end, we introduce Latent Performance Profiling (LPP) -- a framework that derives task-agnostic diagnostics from hidden activations and output distributions. LPP defines a set of scalar metrics on a model's latent representations and dynamics, revealing scale-independent traits that enable interpretable comparisons and uncover hidden vulnerabilities. Unlike static accuracy scores, LPP provides stable, architecture-sensitive signatures across models of similar size. With extensive empirical analyses across eight LLMs, spanning a size range of 0.5B-14B, we demonstrate that models with similar benchmark scores can exhibit contrasting latent profiles, such as differences in entropy or adaptability. Guided by these insights, we design synthetic probes for uncertainty and symbolic reasoning that align with intrinsic metrics while decoupling from leaderboard bias. We recommend that reporting LPP alongside benchmarks provides a deeper, interpretable understanding of model behavior, enabling more reliable model selection, safety assessment, and evaluation beyond surface-level accuracy.

Large Language Models (LLMs) are increasingly used in clinical applications. However, their behavior remains highly sensitive to subtle linguistic variations, such as rephrasing or syntactic variation. This sensitivity poses risks in safety-critical healthcare settings, where semantically equivalent inputs should produce consistent predictions. However, a key challenge is to ensure that prompt variations truly preserve clinical meaning, as embedding-based similarity metrics often fail to capture distinctions involving negation, temporality, or severity. To address this limitation, we propose a semantic verification framework based on Natural Language Inference (NLI) to filter meaning-preserving prompt variations, which are further refined using an LLM-as-a-judge and audited by a clinical expert. In addition, we introduce three metrics to quantify model sensitivity: MeaningPreserving Variation Sensitivity (MVS), confidence variation (\Delta C), and Worst-Case Instability (WCI). We evaluate 16 open-source general-purpose (GP) and medical LLMs within the same model families and parameter scales, using reformulated prompts derived from the DiagnosisQA and MedQA datasets. Our results demonstrate that robustness differences between domain-specific (DS) models are mixed and highly model-dependent, i.e., domain specialization does not consistently improve or reduce robustness to meaning-preserving prompt reformulations. Several DS models rank among the most robust (when compared with GP counterparts), and strong GP baselines remain competitive as well.

Adversarial robustness evaluation underpins every claim of trustworthy ML deployment, yet the field suffers from fragmented protocols and undetected gradient masking. We make two contributions. (1) Structured synthesis. We analyze nine peer-reviewed corpus sources (2020--2026) through seven complementary protocols, producing the first end-to-end structured analysis of the field's consensus and unresolved challenges. (2) Auto-ART framework. We introduce Auto-ART, an open-source framework that operationalizes identified gaps: 50+ attacks, 28 defense modules, the Robustness Diagnostic Index (RDI), and gradient-masking detection. It supports multi-norm evaluation (l1/l2/linf/semantic/spatial) and compliance mapping to NIST AI RMF, OWASP LLM Top 10, and the EU AI Act. Empirical validation on RobustBench demonstrates that Auto-ART's pre-screening identifies gradient masking in 92% of flagged cases, and RDI rankings correlate highly with full AutoAttack. Multi-norm evaluation exposes a 23.5 pp gap between average and worst-case robustness on state-of-the-art models. No prior work combines such structured meta-scientific analysis with an executable evaluation framework bridging literature gaps into engineering.

As Large Language Models (LLMs) are increasingly deployed in safety-critical applications, robust content moderation becomes essential. We present a comprehensive evaluation of 14 open-source safety guard models on a curated benchmark of 79,331 samples spanning 8 NIST AI Risk Framework safety categories. Our benchmark aggregates four diverse datasets (HarmBench, StrongREJECT, RealToxicityPrompts, and BeaverTails), filtered to focus exclusively on safety-relevant content (violence, hate speech, harassment, sexual content, suicide/self-harm, profanity, threats, and health misinformation). We find that recall is the critical metric for safety applications, as missing harmful content poses greater risk than false positives. Our evaluation reveals surprising results: Qwen Guard (4B parameters) achieves the highest recall (83.97%) while larger models like Llama Guard (12B) and GPT-OSS Safeguard (20B) exhibit conservative behavior, missing up to 75% of unsafe content. We demonstrate that model size does not correlate with safety detection performance and that general-purpose guard models outperform specialized ones. These findings provide practical guidance for selecting safety guard models in production deployments.

The emergence of autonomous, high-velocity Agentic AI systems is creating an internal assurance scalability crisis. Point-in-time, document-based audits cannot keep pace with non deterministic behaviour and distributed deployments of agents across rapidly evolving environments. The crisis is dual-scale: vertically, governance and control obligations change faster than frameworks can operationalise them; horizontally, assurance mechanisms fail to scale across complex, heterogeneous systems and evidence sources. Risk-based regulation now requires organisations to demonstrate ongoing control adequacy and effectiveness, yet existing Trustworthy AI Assurance and Audit frameworks remain fragmented and largely manual. Drawing on the evolution of cybersecurity posture management, this paper reframes trustworthiness as a continuously generated signal rather than a static certificate. It contributes 1) A Trustworthy AI Assurance Ontology modelling the end-to-end pathway from regulatory obligation to verifiable evidence 2) An ontology-driven, evidence-gated benchmark of thirteen leading frameworks, revealing a posture readiness gap 3) The Trustworthy AI Posture (TAIP) framework, which operationalises the NIST AI RMF Test,Evaluate,Verify,Validate (TEVV) cycle as reusable AI Assurance Objects. TAIP decouples policy content ('what') from execution semantics ('how'), enabling composable, automatable assurance across jurisdictions and agentic systems. Evidence from heterogeneous tools is normalised and recursively aggregated into posture at claim, system, organisational, and ecosystem levels. A use case mapping Australian AI Guardrails to Microsoft 365 Copilot demonstrates claim decomposition, evidence binding, and posture computation in practice. By standardising execution while allowing policy variation, TAIP enables scalable, machine-speed trust signal generation.

This memorandum presents four recommendations aimed at strengthening the principles of AI model reliability and AI model governability, as DoW, ODNI, NIST, and CAISI refine AI assurance frameworks under the AI Action Plan. Our focus concerns the open scientific problem of misalignment and its implications on AI model behavior. Specifically, misalignment and scheming capabilities can be a red flag indicating AI model insufficient reliability and governability. To address the national security threats arising from misalignment, we recommend that DoW and the IC strategically leverage existing testing and evaluation pipelines and their OT authority to future proof the principles of AI model reliability and AI model governability through a suite of scheming and control evaluations.

Researchers, government bodies, and organizations have been repeatedly calling for a shift in the responsible AI community from general principles to tangible and operationalizable practices in mitigating the potential sociotechnical harms of AI. Frameworks like the NIST AI RMF embody an emerging consensus on recommended practices in operationalizing sociotechnical harm mitigation. However, private sector organizations currently lag far behind this emerging consensus. Implementation is sporadic and selective at best. At worst, it is ineffective and can risk serving as a misleading veneer of trustworthy processes, providing an appearance of legitimacy to substantively harmful practices. In this paper, we provide a foundation for a framework for evaluating where organizations sit relative to the emerging consensus on sociotechnical harm mitigation best practices: a flexible maturity model based on the NIST AI RMF.

Large language models (LLMs) are gaining increasing popularity in both academia and industry, owing to their unprecedented performance in various applications. As LLMs continue to play a vital role in both research and daily use, their evaluation becomes increasingly critical, not only at the task level, but also at the society level for better understanding of their potential risks. Over the past years, significant efforts have been made to examine LLMs from various perspectives. This paper presents a comprehensive review of these evaluation methods for LLMs, focusing on three key dimensions: what to evaluate, where to evaluate, and how to evaluate. Firstly, we provide an overview from the perspective of evaluation tasks, encompassing general natural language processing tasks, reasoning, medical usage, ethics, education, natural and social sciences, agent applications, and other areas. Secondly, we answer the ‘where’ and ‘how’ questions by diving into the evaluation methods and benchmarks, which serve as crucial components in assessing the performance of LLMs. Then, we summarize the success and failure cases of LLMs in different tasks. Finally, we shed light on several future challenges that lie ahead in LLMs evaluation. Our aim is to offer invaluable insights to researchers in the realm of LLMs evaluation, thereby aiding the development of more proficient LLMs. Our key point is that evaluation should be treated as an essential discipline to better assist the development of LLMs. We consistently maintain the related open-source materials at: https://github.com/MLGroupJLU/LLM-eval-survey

Language models (LMs) like GPT‐3, PaLM, and ChatGPT are the foundation for almost all major language technologies, but their capabilities, limitations, and risks are not well understood. We present Holistic Evaluation of Language Models (HELM) to improve the transparency of LMs. LMs can serve many purposes and their behavior should satisfy many desiderata. To navigate the vast space of potential scenarios and metrics, we taxonomize the space and select representative subsets. We evaluate models on 16 core scenarios and 7 metrics, exposing important trade‐offs. We supplement our core evaluation with seven targeted evaluations to deeply analyze specific aspects (including world knowledge, reasoning, regurgitation of copyrighted content, and generation of disinformation). We benchmark 30 LMs, from OpenAI, Microsoft, Google, Meta, Cohere, AI21 Labs, and others. Prior to HELM, models were evaluated on just 17.9% of the core HELM scenarios, with some prominent models not sharing a single scenario in common. We improve this to 96.0%: all 30 models are now benchmarked under the same standardized conditions. Our evaluation surfaces 25 top‐level findings. For full transparency, we release all raw model prompts and completions publicly. HELM is a living benchmark for the community, continuously updated with new scenarios, metrics, and models https://crfm.stanford.edu/helm/latest/.

Progress in AI is often demonstrated by new models claiming improved performance on tasks measuring model capabilities.Evaluating language models can be particularly challenging, as choices of how a model is evaluated on a task can lead to large changes in measured performance.There is no common standard setup, so different models are evaluated on the same tasks in different ways, leading to claims about which models perform best not being reproducible.We propose OLMES, a completely documented, practical, open standard for reproducible LLM evaluations.In developing this standard, we identify and review the varying factors in evaluation practices adopted by the community -such as details of prompt formatting, choice of in-context examples, probability normalizations, and task formulation.In particular, OLMES supports meaningful comparisons between smaller base models that require the unnatural "cloze" formulation of multiple-choice questions against larger models that can utilize the original formulation.OLMES includes well-considered, documented recommendations guided by results from existing literature as well as new experiments resolving open questions. 1 Model Ref1 Ref2 Ref3 Ref4 Ref5 Ref6 OLMES MPT-7B 47.7 42.6 46.5 45.7 RPJ-Incite-7B 46.3 42.8 45.3 Falcon-7B 47.9 42.4 44.5 47.5 49.7 Mistral-7B 60.0 55.5 54.9 78.6 Llama2-7B 53.1 45.9 43.2 45.9 48.5 53.7 54.2 Llama2-13B 59.4 49.4 48.8 49.4 67.

Recently, the evaluation of Large Language Models has emerged as a popular area of research. The three crucial questions for LLM evaluation are ``what, where, and how to evaluate''. However, the existing research mainly focuses on the first two questions, which are basically what tasks to give the LLM during testing and what kind of knowledge it should deal with. As for the third question, which is about what standards to use, the types of evaluators, how to score, and how to rank, there hasn't been much discussion. In this paper, we analyze evaluation methods by comparing various criteria with both manual and automatic evaluation, utilizing onsite, crowd-sourcing, public annotators and GPT-4, with different scoring methods and ranking systems. We propose a new dataset, LLMEval and conduct evaluations on 20 LLMs. A total of 2,186 individuals participated, leading to the generation of 243,337 manual annotations and 57,511 automatic evaluation results. We perform comparisons and analyses of different settings and conduct 10 conclusions that can provide some insights for evaluating LLM in the future. The dataset and the results are publicly available at https://github.com/llmeval. The version with the appendix are publicly available at https://arxiv.org/abs/2312.07398.

Md Tahmid Rahman Laskar, Sawsan Alqahtani, M Saiful Bari, Mizanur Rahman, Mohammad Abdullah Matin Khan, Haidar Khan, Israt Jahan, Amran Bhuiyan, Chee Wei Tan, Md Rizwan Parvez, Enamul Hoque, Shafiq Joty, Jimmy Huang. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Chenxin An, Shansan Gong, Ming Zhong, Xingjian Zhao, Mukai Li, Jun Zhang, Lingpeng Kong, Xipeng Qiu. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Recently, there has been growing interest in using Large Language Models (LLMs) for scientific research. Numerous benchmarks have been proposed to evaluate the ability of LLMs for scientific research. However, current benchmarks are mostly based on pre-collected objective questions. This design suffers from data leakage problem and lacks the evaluation of subjective Q/A ability. In this paper, we propose SciEval, a comprehensive and multi-disciplinary evaluation benchmark to address these issues. Based on Bloom's taxonomy, SciEval covers four dimensions to systematically evaluate scientific research ability. In particular, we design a "dynamic" subset based on scientific principles to prevent evaluation from potential data leakage. Both objective and subjective questions are included in SciEval. These characteristics make SciEval a more effective benchmark for scientific research ability evaluation of LLMs. Comprehensive experiments on most advanced LLMs show that, although GPT-4 achieves SOTA performance compared to other LLMs, there is still substantial room for improvement, especially for dynamic questions. The codes and data are publicly available on https://github.com/OpenDFM/SciEval.

… Given the inherent ambiguities of language, both gold-labeled data sets and human criteria … evaluation criteria can be enhanced by using guidelines that map to terminology standards …

Zhexin Zhang, Leqi Lei, Lindong Wu, Rui Sun, Yongkang Huang, Chong Long, Xiao Liu, Xuanyu Lei, Jie Tang, Minlie Huang. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Abstract Background The proliferation of both general purpose and health care–specific large language models (LLMs) has intensified the challenge of effectively evaluating and comparing them. Data contamination plagues the validity of public benchmarks, self-preference distorts LLM-as-a-judge approaches, and there is a gap between the tasks used to test models and those used in clinical practice. Objective In response, we propose CLEVER (Clinical Large Language Model Evaluation–Expert Review), a methodology for blind, randomized, preference-based evaluation by practicing medical doctors on specific tasks. Methods We demonstrate the methodology by comparing GPT-4o (OpenAI) against 2 health care–specific LLMs, with 8 billion and 70 billion parameters, over 3 tasks: clinical text summarization, clinical information extraction, and question answering on biomedical research. Results Medical doctors prefer the medical model–small LLM trained by John Snow Labs over GPT-4o 45% to 92% more often on the dimensions of factuality, clinical relevance, and conciseness. Conclusions The models show comparable performance on open-ended medical question answering, suggesting that health care–specific LLMs can outperform much larger general purpose LLMs in tasks that require understanding of clinical context. We test the validity of CLEVER evaluations by conducting interannotator agreement, interclass correlation, and washout period analysis.

Large Language Models (LLMs), advanced AI tools based on transformer architectures, demonstrate significant potential in clinical medicine by enhancing decision support, diagnostics, and medical education. However, their integration into clinical workflows requires rigorous evaluation to ensure reliability, safety, and ethical alignment. This systematic review examines the evaluation parameters and methodologies applied to LLMs in clinical medicine, highlighting their capabilities, limitations, and application trends. A comprehensive review of the literature was conducted across PubMed, Scopus, Web of Science, IEEE Xplore, and arXiv databases, encompassing both peer-reviewed and preprint studies. Studies were screened against predefined inclusion and exclusion criteria to identify original research evaluating LLM performance in medical contexts. The results reveal a growing interest in leveraging LLM tools in clinical settings, with 761 studies meeting the inclusion criteria. While general-domain LLMs, particularly ChatGPT and GPT-4, dominated evaluations (93.55%), medical-domain LLMs accounted for only 6.45%. Accuracy emerged as the most commonly assessed parameter (21.78%). Despite these advancements, the evidence base highlights certain limitations and biases across the included studies, emphasizing the need for careful interpretation and robust evaluation frameworks. The exponential growth in LLM research underscores their transformative potential in healthcare. However, addressing challenges such as ethical risks, evaluation variability, and underrepresentation of critical specialties will be essential. Future efforts should prioritize standardized frameworks to ensure safe, effective, and equitable LLM integration in clinical practice.

The emergence of various medical large language models (LLMs) in the medical domain has highlighted the need for unified evaluation standards, as manual evaluation of LLMs proves to be time-consuming and labor-intensive. To address this issue, we introduce MedBench, a comprehensive benchmark for the Chinese medical domain, comprising 40,041 questions sourced from authentic examination exercises and medical reports of diverse branches of medicine. In particular, this benchmark is composed of four key components: the Chinese Medical Licensing Examination, the Resident Standardization Training Examination, the Doctor In-Charge Qualification Examination, and real-world clinic cases encompassing examinations, diagnoses, and treatments. MedBench replicates the educational progression and clinical practice experiences of doctors in Mainland China, thereby establish- ing itself as a credible benchmark for assessing the mastery of knowledge and reasoning abilities in medical language learning models. We perform extensive experiments and conduct an in-depth analysis from diverse perspectives, which culminate in the following findings: (1) Chinese medical LLMs underperform on this benchmark, highlighting the need for significant advances in clinical knowledge and diagnostic precision. (2) Several general-domain LLMs surprisingly possess considerable medical knowledge. These findings elucidate both the capabilities and limitations of LLMs within the context of MedBench, with the ultimate goal of aiding the medical research community.

With generative artificial intelligence (GenAI), particularly large language models (LLMs), continuing to make inroads in healthcare, assessing LLMs with human evaluations is essential to assuring safety and effectiveness. This study reviews existing literature on human evaluation methodologies for LLMs in healthcare across various medical specialties and addresses factors such as evaluation dimensions, sample types and sizes, selection, and recruitment of evaluators, frameworks and metrics, evaluation process, and statistical analysis type. Our literature review of 142 studies shows gaps in reliability, generalizability, and applicability of current human evaluation practices. To overcome such significant obstacles to healthcare LLM developments and deployments, we propose QUEST, a comprehensive and practical framework for human evaluation of LLMs covering three phases of workflow: Planning, Implementation and Adjudication, and Scoring and Review. QUEST is designed with five proposed evaluation principles: Quality of Information, Understanding and Reasoning, Expression Style and Persona, Safety and Harm, and Trust and Confidence.

As large language models (LLMs) develop increasingly sophisticated capabilities and find applications in medical settings, it becomes important to assess their medical safety due to their far-reaching implications for personal and public health, patient safety, and human rights. However, there is little to no understanding of the notion of medical safety in the context of LLMs, let alone how to evaluate and improve it. To address this gap, we first define the notion of medical safety in LLMs based on the Principles of Medical Ethics set forth by the American Medical Association. We then leverage this understanding to introduce MedSafetyBench, the first benchmark dataset designed to measure the medical safety of LLMs. We demonstrate the utility of MedSafetyBench by using it to evaluate and improve the medical safety of LLMs. Our results show that publicly-available medical LLMs do not meet standards of medical safety and that fine-tuning them using MedSafetyBench improves their medical safety while preserving their medical performance. By introducing this new benchmark dataset, our work enables a systematic study of the state of medical safety in LLMs and motivates future work in this area, paving the way to mitigate the safety risks of LLMs in medicine. The benchmark dataset and code are available at https://github.com/AI4LIFE-GROUP/med-safety-bench.

Human evaluation is indispensable and inevitable for assessing the quality of texts generated by machine learning models or written by humans. However, human evaluation is very difficult to reproduce and its quality is notoriously unstable, hindering fair comparisons among different natural language processing (NLP) models and algorithms.Recently, large language models (LLMs) have demonstrated exceptional performance on unseen tasks when only the task instructions are provided.In this paper, we explore if such an ability of the LLMs can be used as an alternative to human evaluation.We present the LLMs with the exact same instructions, samples to be evaluated, and questions used to conduct human evaluation, and then ask the LLMs to generate responses to those questions; we dub this LLM evaluation.We use human evaluation and LLM evaluation to evaluate the texts in two NLP tasks: open-ended story generation and adversarial attacks.We show that the result of LLM evaluation is consistent with the results obtained by expert human evaluation: the texts rated higher by human experts are also rated higher by the LLMs.We also find that the results of LLM evaluation are stable over different formatting of the task instructions and the sampling algorithm used to generate the answer.We are the first to show the potential of using LLMs to assess the quality of texts and discuss the limitations and ethical considerations of LLM evaluation.

This study aimed to evaluate the performance of large language models (LLMs) in the task of abstract screening in systematic review and meta-analysis studies, exploring their effectiveness, efficiency, and potential integration into existing human expert-based workflows. We developed automation scripts in Python to interact with the APIs of several LLM tools, including ChatGPT v4.0, ChatGPT v3.5, Google PaLM 2, and Meta Llama 2, and latest tools including ChatGPT v4.0 turbo, ChatGPT v3.5 turbo, Google Gemini 1.0 pro, Meta Llama 3, and Claude 3. This study focused on three databases of abstracts and used them as benchmarks to evaluate the performance of these LLM tools in terms of sensitivity, specificity, and overall accuracy. The results of the LLM tools were compared to human-curated inclusion decisions, gold standard for systematic review and meta-analysis studies. Different LLM tools had varying abilities in abstract screening. Chat GPT v4.0 demonstrated remarkable performance, with balanced sensitivity and specificity, and overall accuracy consistently reaching or exceeding 90%, indicating a high potential for LLMs in abstract screening tasks. The study found that LLMs could provide reliable results with minimal human effort and thus serve as a cost-effective and efficient alternative to traditional abstract screening methods. While LLM tools are not yet ready to completely replace human experts in abstract screening, they show great promise in revolutionizing the process. They can serve as autonomous AI reviewers, contribute to collaborative workflows with human experts, and integrate with hybrid approaches to develop custom tools for increased efficiency. As technology continues to advance, LLMs are poised to play an increasingly important role in abstract screening, reshaping the workflow of systematic review and meta-analysis studies.

Large language models (LLMs) are a type of generative artificial intelligence (AI) designed to produce text‐based content. LLMs use deep learning techniques and massively large data sets to understand, summarize, generate, and predict new text. LLMs caught the public eye in early 2023 when ChatGPT (the first consumer facing LLM) was released. LLM technologies are driven by recent advances in deep‐learning AI techniques, where language models are trained on extremely large text data from the internet and then re‐used for downstream tasks with limited fine‐tuning required. They offer exciting opportunities for evaluators to automate and accelerate time‐consuming tasks involving text analytics and text generation. We estimate that over two‐thirds of evaluation tasks will be affected by LLMs in the next 5 years. Use‐case examples include summarizing text data, extracting key information from text, analyzing and classifying text content, writing text, and translation. Despite the advances, the technologies pose significant challenges and risks. Because LLM technologies are generally trained on text from the internet, they tend to perpetuate biases (racism, sexism, ethnocentrism, and more) and exclusion of non‐majority languages. Current tools like ChatGPT have not been specifically developed for monitoring, evaluation, research, and learning (MERL) purposes, possibly limiting their accuracy and usefulness for evaluation. In addition, technical limitations and challenges with bias can lead to real world harm. To overcome these technical challenges and ethical risks, the evaluation community will need to work collaboratively with the data science community to co‐develop tools and processes and to ensure the application of quality and ethical standards.

… foundation models cultivate complex innovation ecosystems that render traditional, project-centric evaluation frameworks … and validates a three-level framework for assessing ecosystem-…

… In this paper, we propose a novel benchmarking framework, Language-Model-as-an-… a single model as examiner, we propose a decentralized evaluation setting where all participating …

… paradigms are most effective, there is a need for a holistic evaluation scheme … Evaluation of Multimodal Models (HEMM), visualized in Figure 1. HEMM, as an evaluation framework, …

… such a model for coded EHR data; in contrast, most prior models released for … evaluation of foundation models on benefits such as sample efficiency and task adaptation. Our model and …

Assessing foundation models' abilities for human-level tasks is crucial for Artificial General Intelligence (AGI) development.Traditional benchmarks, which rely on artificial datasets, may not accurately represent these capabilities.In this paper, we introduce AGIEval, a novel bilingual benchmark designed to assess foundation models in the context of humancentric standardized exams, such as college entrance exams, law school admission tests, math competitions, and lawyer qualification tests.We evaluate several state-of-the-art foundation models on our benchmark.Impressively, we show that GPT-4 exceeds the average human performance in SAT, LSAT, and math contests, with 95% accuracy on SAT Math and 92.5% on the Chinese college entrance English exam.This demonstrates the exceptional performance of contemporary foundation models.In contrast, we also find that GPT-4 is less proficient in tasks requiring complex reasoning or specific domain knowledge.Our comprehensive analyses of model capabilities (understanding, knowledge, reasoning, and calculation) reveal their strengths and limitations, providing valuable insights into future directions for enhancing general capabilities.By concentrating on tasks pertinent to human cognition and decision-making, our benchmark delivers a meaningful and robust evaluation of foundation models' performance in real-world scenarios 1 .Recently, large foundation models, such as the large language models (LLMs) ChatGPT(OpenAI, 2022) and GPT-4 (OpenAI, 2023), exhibited remarkable versatility and adaptability, with plethora of applications spanning various domains as a decision-making assistant, from processing daily events to assisting in specialized fields such as law * indicates equal contribution.Yaobo Liang and Nan Duan are the corresponding authors.1 The data, code, and all model outputs are released in

Computational pathology foundation models (CPathFMs) have emerged as a powerful approach for analyzing histopathological data, leveraging self-supervised learning to extract robust feature representations from unlabeled whole-slide images. These models, categorized into uni-modal and multi-modal frameworks, have demonstrated promise in automating complex pathology tasks such as segmentation, classification, and biomarker discovery. However, the development of CPathFMs presents significant challenges, such as limited data accessibility, high variability across datasets, the necessity for domain-specific adaptation, and the lack of standardized evaluation benchmarks. This survey provides a comprehensive review of CPathFMs in computational pathology, focusing on pre-training datasets, adaptation strategies, and evaluation tasks. We analyze key techniques, such as contrastive learning, masked image modeling and multi-modal integration, and highlight existing gaps in current research. Finally, we explore future directions from four perspectives for advancing CPathFMs. This survey serves as a valuable resource for researchers, clinicians, and AI practitioners, guiding the advancement of CPathFMs toward robust and clinically applicable AI-driven pathology solutions.

Reasoning, a crucial ability for complex problem-solving, plays a pivotal role in various real-world settings such as negotiation, medical diagnosis, and criminal investigation. It serves as a fundamental methodology in the field of Artificial General Intelligence (AGI). With the ongoing development of foundation models, there is a growing interest in exploring their abilities in reasoning tasks. In this article, we introduce seminal foundation models proposed or adaptable for reasoning, highlighting the latest advancements in various reasoning tasks, methods, and benchmarks. We then delve into the potential future directions behind the emergence of reasoning abilities within foundation models. We also discuss the relevance of multimodal learning, autonomous agents, and super alignment in the context of reasoning. By discussing these future research directions, we hope to inspire researchers in their exploration of this field, stimulate further advancements in reasoning with foundation models, e.g., Large Language Models (LLMs), and contribute to the development of AGI.

Recent studies have indicated that foundation models, such as BERT and GPT, excel at adapting to various downstream tasks. This adaptability has made them a dominant force in building artificial intelligence (AI) systems. Moreover, a new research paradigm has emerged as visualization techniques are incorporated into these models. This study divides these intersections into two research areas: visualization for foundation model (VIS4FM) and foundation model for visualization (FM4VIS). In terms of VIS4FM, we explore the primary role of visualizations in understanding, refining, and evaluating these intricate foundation models. VIS4FM addresses the pressing need for transparency, explainability, fairness, and robustness. Conversely, in terms of FM4VIS, we highlight how foundation models can be used to advance the visualization field itself. The intersection of foundation models with visualizations is promising but also introduces a set of challenges. By highlighting these challenges and promising opportunities, this study aims to provide a starting point for the continued exploration of this research avenue.

… is rapidly improving, this analysis sets a benchmark for ChatGPTLs performance as of early … to establish a benchmark of ChatGPTLs performance in a diverse range of assessment tasks …

Generative Artificial Intelligence (AI) is transforming industries by producing high-quality text, images, music, and code. Its applications extend to natural language processing, computer vision, and creative arts. However, assessing these systems' performance and impact remains challenging due to their complexity, subjectivity, and open-ended outputs. This paper comprehensively reviews evaluation methods for generative AI, beginning with its evolution and major applications, including advanced models like GPT, DALL·E, and AlphaCode. It categorizes evaluation approaches into quantitative metrics (such as BLEU and FID) and qualitative methods (human assessment and user-centered testing). Key challenges, such as subjectivity, bias, and scalability, are explored alongside emerging trends like automated evaluation tools, ethical impact assessments, and multimodal techniques. Through real-world case studies, this paper highlights practical evaluation strategies and their limitations. By integrating current best practices and identifying future research opportunities, this study aims to guide the development of reliable, fair, and comprehensive evaluation frameworks for generative AI systems.

… Abstract Generative artificial intelligence (AI) systems show immense potential to transform … benchmarking of four leading generative models. Through standardized assessments, the …

The introduction and rapid evolution of generative artificial intelligence (genAI) models necessitates a refined understanding for the concept of “intelligence”. The genAI tools are known for its capability to produce complex, creative, and contextually relevant output. Nevertheless, the deployment of genAI models in healthcare should be accompanied appropriate and rigorous performance evaluation tools. In this rapid communication, we emphasizes the urgent need to develop a “Generative AIQ Test” as a novel tailored tool for comprehensive benchmarking of genAI models against multiple human-like intelligence attributes. A preliminary framework is proposed in this communication. This framework incorporates miscellaneous performance metrics including accuracy, diversity, novelty, and consistency. These metrics were considered critical in the evaluation of genAI models that might be utilized to generate diagnostic recommendations, treatment plans, and patient interaction suggestions. This communication also highlights the importance of orchestrated collaboration to construct robust and well-annotated benchmarking datasets to capture the complexity of diverse medical scenarios and patient demographics. This communication suggests an approach aiming to ensure that genAI models are effective, equitable, and transparent. To maximize the potential of genAI models in healthcare, it is important to establish rigorous, dynamic standards for its benchmarking. Consequently, this approach can help to improve clinical decision-making with enhancement in patient care, which will enhance the reliability of genAI applications in healthcare.

… advance the understanding of how to assess generative AI systems effectively and responsibly. … New benchmarks assess whether human collaborators can distinguish between human-…

… benchmark is being designed to reflect real-world, productionscale GenAI scenarios. The benchmark … inference, postprocessing, and quality assessment. The TPC consortium is actively …

… metrics that assess the efficacy of the end-user application enabled by AI. Standards and … , the adoption of Generative AI systems will require a holistic approach to benchmarking that …

… benchmark is being designed to reflect real-world, productionscale GenAI scenarios. The benchmark … inference, postprocessing, and quality assessment. The TPC consortium is actively …

… assessment. We 13 also develop a living repository of benchmark assessments to support benchmark … language model benchmarks in the era of generative artificial intelligence. arXiv …

Critical National Infrastructures (CNIs)—including energy grids, water systems, transportation networks, and communication frameworks—are essential to modern society yet face escalating cybersecurity threats. This review paper comprehensively analyzes AI-driven approaches for Critical Infrastructure Protection (CIP). We begin by examining the reliability of CNIs and introduce established benchmarks for evaluating Large Language Models (LLMs) within cybersecurity contexts. Next, we explore core cybersecurity issues, focusing on trust, privacy, resilience, and securability in these vital systems. Building on this foundation, we assess the role of Generative AI and LLMs in enhancing CIP and present insights on applying Agentic AI for proactive defense mechanisms. Finally, we outline future directions to guide the integration of advanced AI methodologies into protecting critical infrastructures. Our paper provides a strategic roadmap for researchers and practitioners committed to fortifying national infrastructures against emerging cyber threats through this synthesis of current challenges, benchmarking strategies, and innovative AI applications.

In this study, we introduce a framework designed to help educators assess the effectiveness of popular generative artificial intelligence (AI) tools in solving authentic assessments. We employed Bloom’s taxonomy as a guiding principle to create authentic assessments that evaluate the capabilities of generative AI tools. We applied this framework to assess the abilities of ChatGPT-4, ChatGPT-3.5, Google Bard and Microsoft Bing in solving authentic assessments in economics. We found that generative AI tools perform very well at the lower levels of Bloom's taxonomy while still maintaining a decent level of performance at the higher levels, with “create” being the weakest level of performance. Interestingly, these tools are better able to address numeric-based questions than text-based ones. Moreover, all the generative AI tools exhibit weaknesses in building arguments based on theoretical frameworks, maintaining the coherence of different arguments and providing appropriate references. Our study provides educators with a framework to assess the capabilities of generative AI tools, enabling them to make more informed decisions regarding assessments and learning activities. Our findings demand a strategic reimagining of educational goals and assessments, emphasising higher cognitive skills and calling for a concerted effort to enhance the capabilities of educators in preparing students for a rapidly transforming professional environment. Implications for practice or policy Our proposed framework enables educators to systematically evaluate the capabilities of widely used generative AI tools in assessments and assist them in the assessment design process. Tertiary institutions should re-evaluate and redesign programmes and course learning outcomes. The new focus on learning outcomes should address the higher levels of educational goals of Bloom’s taxonomy, specifically the “create” level.

Generative artificial intelligence (AI) has emerged as a powerful technology with numerous applications in various domains. There is a need to identify the requirements and evaluation metrics for generative AI models designed for specific tasks. The purpose of the research aims to investigate the fundamental aspects of generative AI systems, including their requirements, models, input–output formats, and evaluation metrics. The study addresses key research questions and presents comprehensive insights to guide researchers, developers, and practitioners in the field. Firstly, the requirements necessary for implementing generative AI systems are examined and categorized into three distinct categories: hardware, software, and user experience. Furthermore, the study explores the different types of generative AI models described in the literature by presenting a taxonomy based on architectural characteristics, such as variational autoencoders (VAEs), generative adversarial networks (GANs), diffusion models, transformers, language models, normalizing flow models, and hybrid models. A comprehensive classification of input and output formats used in generative AI systems is also provided. Moreover, the research proposes a classification system based on output types and discusses commonly used evaluation metrics in generative AI. The findings contribute to advancements in the field, enabling researchers, developers, and practitioners to effectively implement and evaluate generative AI models for various applications. The significance of the research lies in understanding that generative AI system requirements are crucial for effective planning, design, and optimal performance. A taxonomy of models aids in selecting suitable options and driving advancements. Classifying input–output formats enables leveraging diverse formats for customized systems, while evaluation metrics establish standardized methods to assess model quality and performance.

Generative Artificial Intelligence (GenAI) is transforming software development processes, including test and evaluation (T&E). From automating test case design to enabling continuous testing in DevOps pipelines, AI-driven tools enhance the efficiency, accuracy, and speed of software testing. At the same time, the integration of AI components into software-reliant systems introduces new challenges for verification and validation (V&V). Traditional T&E methodologies must evolve to address issues such as AI bias, hallucinated outputs, and the complexity of validating non-deterministic behaviors. This position paper examines how existing T&E methods must evolve to account for AI's stochastic nature, and conversely how GenAI is transforming T&E practices across the software development lifecycle (SDLC).

The remarkable strides in artificial intelligence (AI), exemplified by ChatGPT, have unveiled a wealth of opportunities and challenges in assessment. Applying cutting‐edge large language models (LLMs) and generative AI to assessment holds great promise in boosting efficiency, mitigating bias, and facilitating customized evaluations. Conversely, these innovations raise significant concerns regarding validity, reliability, transparency, fairness, equity, and test security, necessitating careful thinking when applying them in assessments. In this article, we discuss the impacts and implications of LLMs and generative AI on critical dimensions of assessment with example use cases and call for a community effort to equip assessment professionals with the needed AI literacy to harness the potential effectively.

OBJECTIVE Machine learning (ML) and artificial intelligence (AI) are rapidly becoming the most talked about and controversial topics in radiology and medicine. Over the past few years, the numbers of ML- or AI-focused studies in the literature have increased almost exponentially, and ML has become a hot topic at academic and industry conferences. However, despite the increased awareness of ML as a tool, many medical professionals have a poor understanding of how ML works and how to critically appraise studies and tools that are presented to us. Thus, we present a brief overview of ML, explain the metrics used in ML and how to interpret them, and explain some of the technical jargon associated with the field so that readers with a medical background and basic knowledge of statistics can feel more comfortable when examining ML applications. CONCLUSION Attention to sample size, overfitting, underfitting, cross validation, as well as a broad knowledge of the metrics of machine learning, can help those with little or no technical knowledge begin to assess machine learning studies. However, transparency in methods and sharing of algorithms is vital to allow clinicians to assess these tools themselves.

… if the criteria for the measurement and evaluation of these models are not standardized. … evaluation metrics being applied in research and the non-standardization of evaluation metrics …

… the bigger picture of Evaluating model performance within the context of the conceptualization of the key error metrics in AI and how to interpret them to model performance evaluation. A …

Purpose: As artificial intelligence (AI) systems become integral to diverse applications, ensuring their reliability is of paramount importance. This paper explores the multifaceted landscape of AI reliability, encompassing challenges, evaluation metrics, and prospective advancements. Methodology: This paper employs a comprehensive literature review approach to assess the existing body of knowledge on the reliability of AI systems. The review aims to synthesize insights into the challenges faced in evaluating AI reliability, the metrics used for assessment, and the potential future directions in this critical research domain. Findings: In this paper, challenges in AI reliability assessment, including explainability, data quality, and susceptibility to adversarial attacks, are scrutinized. Metrics for evaluating AI reliability, such as robustness, accuracy, precision, and explainability, are also elucidated. In addition, case studies illustrate instances where AI reliability has been successfully assessed or has fallen short, offering valuable insights. Originality/value: This paper sheds light on the complexities surrounding the assessment of artificial intelligence (AI) reliability and contributes to the ongoing discourse on AI reliability by providing a comprehensive examination of its challenges, metrics, and future trajectories.

Ensuring the trustworthiness of artificial intelligence (AI) systems is critical as they become increasingly integrated into domains like healthcare, finance, and public administration. This paper explores frameworks and metrics for evaluating AI trustworthiness, focusing on key principles such as fairness, transparency, privacy, and security. This study is guided by two central questions: how can trust in AI systems be systematically measured across the AI lifecycle, and what are the trade-offs involved when optimizing for different trustworthiness dimensions? By examining frameworks such as the NIST AI Risk Management Framework (AI RMF), the AI Trust Framework and Maturity Model (AI-TMM), and ISO/IEC standards, this study bridges theoretical insights with practical applications. We identify major risks across the AI lifecycle stages and outline various metrics to address challenges in system reliability, bias mitigation, and model explainability. This study includes a comparative analysis of existing standards and their application across industries to illustrate their effectiveness. Real-world case studies, including applications in healthcare, financial services, and autonomous systems, demonstrate approaches to applying trust metrics. The findings reveal that achieving trustworthiness involves navigating trade-offs between competing metrics, such as fairness versus efficiency or privacy versus transparency, and emphasizes the importance of interdisciplinary collaboration for robust AI governance. Emerging trends suggest the need for adaptive frameworks for AI trustworthiness that evolve alongside advancements in AI technologies. This paper contributes to the field by proposing a comprehensive review of existing frameworks with guidelines for building resilient, ethical, and transparent AI systems, ensuring their alignment with regulatory requirements and societal expectations.

… We contribute to the field of AI governance with the development of a unified compliance metric that integrates three key dimensions of SAFE Artificial Intelligence: Security, Accuracy, …

The effective integration of artificial intelligence (AI) systems into clinical medicine depends on comprehensive and transparent performance evaluation; however, the lack of standardized and widely accepted metrics poses challenges for reproducibility and model adoption. A comprehensive, machine-interpretable framework is presented to formalize the nomenclature and descriptions of 207 graphical, matrix, and scalar metrics used to measure AI model performance. The metrics taxonomy, developed as part of the Radiology Ontology of AI Datasets, Models and Projects (ROADMAP), provides a logically structured representation that captures the semantics of AI evaluation metrics, supports reasoning over metric classes, and enables automated completeness checks for AI model reporting. For each metric, the taxonomy incorporates a definition and citations to authoritative reference sources; where applicable, the taxonomy also includes synonyms, abbreviations, alternate language forms, mathematical formulae, and numerical bounds. The taxonomy supports evaluation of models operating on structured data, medical images, audio signals, and/or unstructured text. Logical axioms link each metric to one or more of 18 AI model performance criteria, including classification, calibration, image segmentation, and text analysis. By harmonizing terminology and enabling structured queries, ROADMAP's taxonomy of AI performance metrics facilitates model comparison, bias detection, and selection of appropriate evaluation methods across diverse datasets and clinical tasks. © RSNA, 2026 See also accompanying Special Report on ROADMAP ontology.

… This metric is particularly important for risk assessment models. For example, in the Heart … leveraging these evaluation metrics, we ensured a more comprehensive assessment of model …

Generative Artificial Intelligence is set to revolutionize healthcare delivery by transforming traditional patient care into a more personalized, efficient, and proactive process. Chatbots, serving as interactive conversational models, will probably drive this patient-centered transformation in healthcare. Through the provision of various services, including diagnosis, personalized lifestyle recommendations, dynamic scheduling of follow-ups, and mental health support, the objective is to substantially augment patient health outcomes, all the while mitigating the workload burden on healthcare providers. The life-critical nature of healthcare applications necessitates establishing a unified and comprehensive set of evaluation metrics for conversational models. Existing evaluation metrics proposed for various generic large language models (LLMs) demonstrate a lack of comprehension regarding medical and health concepts and their significance in promoting patients’ well-being. Moreover, these metrics neglect pivotal user-centered aspects, including trust-building, ethics, personalization, empathy, user comprehension, and emotional support. The purpose of this paper is to explore state-of-the-art LLM-based evaluation metrics that are specifically applicable to the assessment of interactive conversational models in healthcare. Subsequently, we present a comprehensive set of evaluation metrics designed to thoroughly assess the performance of healthcare chatbots from an end-user perspective. These metrics encompass an evaluation of language processing abilities, impact on real-world clinical tasks, and effectiveness in user-interactive conversations. Finally, we engage in a discussion concerning the challenges associated with defining and implementing these metrics, with particular emphasis on confounding factors such as the target audience, evaluation methods, and prompt techniques involved in the evaluation process.

Agentic AI systems are a recently emerged and important approach that goes beyond traditional AI, generative AI, and autonomous systems by focusing on autonomy, adaptability, and goal-driven reasoning. This study provides a clear review of agentic AI systems by bringing together their definitions, frameworks, and architectures, and by comparing them with related areas like generative AI, autonomic computing, and multi-agent systems. To do this, we reviewed 143 primary studies on current LLM-based and non-LLM-driven agentic systems and examined how they support planning, memory, reflection, and goal pursuit. Furthermore, we classified architectural models, input–output mechanisms, and applications based on their task domains where agentic AI is applied, supported using tabular summaries that highlight real-world case studies. Evaluation metrics were classified as qualitative and quantitative measures, along with available testing methods of agentic AI systems to check the system’s performance and reliability. This study also highlights the main challenges and limitations of agentic AI, covering technical, architectural, coordination, ethical, and security issues. We organized the conceptual foundations, available tools, architectures, and evaluation metrics in this research, which defines a structured foundation for understanding and advancing agentic AI. These findings aim to help researchers and developers build better, clearer, and more adaptable systems that support responsible deployment in different domains.

Tongxin Yuan, Zhiwei He, Lingzhong Dong, Yiming Wang, Ruijie Zhao, Tian Xia, Lizhen Xu, Binglin Zhou, Fangqi Li, Zhuosheng Zhang, Rui Wang, Gongshen Liu. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024.

The last two years have seen a rapid growth in concerns around the safety of large language models (LLMs). Researchers and practitioners have met these concerns by creating an abundance of datasets for evaluating and improving LLM safety. However, much of this work has happened in parallel, and with very different goals in mind, ranging from the mitigation of near-term risks around bias and toxic content generation to the assessment of longer-term catastrophic risk potential. This makes it difficult for researchers and practitioners to find the most relevant datasets for their use case, and to identify gaps in dataset coverage that future work may fill. To remedy these issues, we conduct a first systematic review of open datasets for evaluating and improving LLM safety. We review 144 datasets, which we identified through an iterative and community-driven process over the course of several months. We highlight patterns and trends, such as a trend towards fully synthetic datasets, as well as gaps in dataset coverage, such as a clear lack of non-English and naturalistic datasets. We also examine how LLM safety datasets are used in practice -- in LLM release publications and popular LLM benchmarks -- finding that current evaluation practices are highly idiosyncratic and make use of only a small fraction of available datasets. Our contributions are based on SafetyPrompts.com, a living catalogue of open datasets for LLM safety, which we plan to update continuously as the field of LLM safety develops.

… To delve into these issues, we firstly proposed a multi-dimensional safety evaluation Benchmark to evaluate LLM Safety Generalization across diverse test tasks and prompt types (SG-…

… provide robust, interpretable, and human-aligned evaluations for complex agent behaviors. … LLM-based evaluators in agent safety and security, we introduce ASSEBench (Agent Safety …

Zhichen Dong, Zhanhui Zhou, Chao Yang, Jing Shao, Yu Qiao. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Generative large language models (LLMs) have revolutionized natural language processing with their transformative and emergent capabilities. However, recent evidence indicates that LLMs can produce harmful content that violates social norms, raising significant concerns regarding the safety and ethical ramifications of deploying these advanced models. Thus, it is both critical and imperative to perform a rigorous and comprehensive safety evaluation of LLMs before deployment. Despite this need, owing to the extensiveness of LLM generation space, it still lacks a unified and standardized risk taxonomy to systematically reflect the LLM content safety, as well as automated safety assessment techniques to explore the potential risks efficiently. To bridge the striking gap, we propose S-Eval, a novel LLM-based automated Safety Evaluation framework with a newly defined comprehensive risk taxonomy. S-Eval incorporates two key components, i.e., an expert testing LLM Mt and a novel safety critique LLM Mc. The expert testing LLM Mt is responsible for automatically generating test cases in accordance with the proposed risk management (including 8 risk dimensions and a total of 102 subdivided risks). The safety critique LLM Mc can provide quantitative and explainable safety evaluations for better risk awareness of LLMs. In contrast to prior works, S-Eval differs in significant ways: (i) efficient – we construct a multi-dimensional and open-ended benchmark comprising 220,000 test cases across 102 risks utilizing Mt and conduct safety evaluations for 21 influential LLMs via Mc on our benchmark. The entire process is fully automated and requires no human involvement. (ii) effective – extensive validations show S-Eval facilitates a more thorough assessment and better perception of potential LLM risks, and Mc not only accurately quantifies the risks of LLMs but also provides explainable and in-depth insights into their safety, surpassing comparable models such as LLaMA-Guard-2. (iii) adaptive – S-Eval can be flexibly configured and adapted to the rapid evolution of LLMs and accompanying new safety threats, test generation methods and safety critique methods thanks to the LLM-based architecture. We further study the impact of hyper-parameters and language environments on model safety, which may lead to promising directions for future research. S-Eval has been deployed in our industrial partner for the automated safety evaluation of multiple LLMs serving millions of users, demonstrating its effectiveness in real-world scenarios.

… the safety of LLM compression techniques. We systematically investigated multiple aspects of safety… Our safety evaluation, along with downstream task performances, reveals that model …

… in standard safety benchmarks and other current evaluation methodologies. To address these needs for robust AI safety evaluation, we introduce SAGE (Safety AI Generic Evaluation), …

Background: The deployment of artificial intelligence (AI) in healthcare necessitates robust safety validation frameworks, particularly for systems directly interacting with patients. While theoretical frameworks exist, there remains a critical gap between abstract principles and practical implementation. Traditional LLM benchmarking approaches provide very limited output coverage and are insufficient for healthcare applications requiring high safety standards. Objective: To develop and evaluate a comprehensive framework for healthcare AI safety validation through large-scale clinician engagement. Methods: We implemented the RWE-LLM (Real-World Evaluation of Large Language Models in Healthcare) framework, drawing inspiration from red teaming methodologies while expanding their scope to achieve comprehensive safety validation. Our approach emphasizes output testing rather than relying solely on input data quality across four stages: pre-implementation, tiered review, resolution, and continuous monitoring. We engaged 6,234 US licensed clinicians (5,969 nurses and 265 physicians) with an average of 11.5 years of clinical experience. The framework employed a three-tier review process for error detection and resolution, evaluating a non-diagnostic AI Care Agent focused on patient education, follow-ups, and administrative support across four iterations (pre-Polaris and Polaris 1.0, 2.0, and 3.0). Results: Over 307,000 unique calls were evaluated using the RWE-LLM framework. Each interaction was subject to potential error flagging across multiple severity categories, from minor clinical inaccuracies to significant safety concerns. The multi-tiered review system successfully processed all flagged interactions, with internal nursing reviews providing initial expert evaluation followed by physician adjudication when necessary. The framework demonstrated effective throughput in addressing identified safety concerns while maintaining consistent processing times and documentation standards. Systematic improvements in safety protocols were achieved through a continuous feedback loop between error identification and system enhancement. Performance metrics demonstrated substantial safety improvements between iterations, with correct medical advice rates improving from ~80.0% (pre-Polaris), to 96.79% (Polaris 1.0), to 98.75% (Polaris 2.0) and 99.38% (Polaris 3.0). Incorrect advice resulting in potential minor harm decreased from 1.32% to 0.13% and 0.07%, and severe harm concerns were eliminated (0.06% to 0.10% and 0.00%). Conclusions: The successful nationwide implementation of the RWE-LLM framework establishes a practical model for ensuring AI safety in healthcare settings. Our methodology demonstrates that comprehensive output testing provides significantly stronger safety assurance than traditional input validation approaches used by horizontal LLMs. While resource-intensive, this approach proves that rigorous safety validation for healthcare AI systems is both necessary and achievable, setting a benchmark for future deployments.

… the AI RMF 1.0 as well as AI research and development and evaluation conducted by NIST and … that will enhance this Framework will be captured in an associated AI Risk Management …

Artificial intelligence (AI) is being adopted at an exponential rate to improve efficiency, decision-making, and cybersecurity, but its rapid integration introduces new and often poorly understood risks, including system errors, algorithmic bias, data privacy concerns, security vulnerabilities, and ethical dilemmas. This paper examines how organizations are implementing AI and evaluates the National Institute of Standards and Technology's AI Risk Management Framework (NIST AI RMF) as a tool for managing these risks. It reviews the benefits of AI adoption alongside the risks emerging from its use in business and broader society and examines the legal and ethical challenges organizations face when implementing AI risk management, including barriers specific to small and medium-sized enterprises (SMEs). Drawing on current literature and complementary regulatory frameworks, including the EU AI Act and ISO/IEC 42001, the paper situates the NIST AI RMF's four core functions, Govern, Map, Measure, and Manage, within this broader governance landscape. It finds that the primary obstacle to effective AI governance is not a lack of standards but inconsistent and incomplete organizational implementation. The paper concludes by identifying current gaps in AI governance practice and proposing future directions for improving AI risk management and security practices across organizations of varying size and maturity.

Critical infrastructure systems increasingly rely on artificial intelligence for decision-making, yet current development practices inadequately address security risks unique to AI systems. This paper presents Secure AI-SDLC, a comprehensive security framework that operationalizes the NIST AI Risk Management Framework with structured software development lifecycle controls for cyber-physical and safety-critical environments. Our approach maps AI-specific threats to actionable engineering controls and required evidence artifacts across five lifecycle phases: govern, design, build, test, and deploy/monitor, aligned with ISO/IEC 42001, IEC 62443, and NERC CIP. Validation through a 16-week deployment on an industrial monitoring pipeline in the energy sector demonstrated measurable improvements, including a 12.3% increase in adversarial robustness and a 40.2% reduction in mean time to detect anomalies. Mean time to recovery decreased from 4.2 hours to 1.8 hours. The proposed framework addresses the gap between high-level AI governance and practical implementation by providing concrete, auditable mechanisms for building and operating trustworthy AI systems in high-consequence environments.

The rapid advancement of Artificial Intelligence (AI) and digital technologies has introduced complex cyber threats. To manage these threats effectively, the United States National Institute of Standards and Technology (NIST) has evolved its Risk Management Framework (RMF), culminating in the 2023 release of the AI RMF 1.0 to address AI-specific risks. However, NIST Special Publication (SP) 800-53 Revision 5 does not fully encompass inherent AI risk factors such as bias and lack of explainability, necessitating the integration of AI-specific risk management into existing security frameworks. While organizations like stackArmor have conducted crosswalk research, previous studies often lack quantitative validation and methodological precision. In this study, we propose a hybrid automated crosswalk framework that integrates Term Frequency-Inverse Document Frequency (TF-IDF) and Sentence-Bidirectional Encoder Representations from Transformers (Sentence-BERT) to systematically map security controls to AI RMF functions. The methodology follows a two-stage structured process: first, deriving unique keywords for each control family using context-enhanced embeddings and Maximal Marginal Relevance (MMR) to minimize redundancy; and second, employing a hybrid query structure incorporating domain-specific keywords—such as model drift and distributed denial of service (DDoS) against machine learning (ML) infrastructure—to ensure technical granularity. Our approach demonstrated 98.63% consistency with expert-defined mappings and identified 60% more valid controls compared to manual baselines. These findings suggest that the proposed framework significantly enhances the scalability and precision of AI risk compliance, providing a robust foundation for automated, high-fidelity security auditing.

人工智能大模型应用生态快速拓展，环境、社会和治理面临新的挑战和机遇，探索构建大模型发展风险的治理框架，对于推进人工智能健康可持续发展具有重要的理论价值和实践意义。文章基于ESG（环境—社会—治理）和人工智能治理相关理论，分析ESG视角下人工智能大模型发展效益、典型风险，进而构建人工智能大模型风险治理框架和实施策略。研究认为技术治理和制度创新双管齐下是应对大模型发展风险的必要手段。遵循“切入视角→风险识别→治理机制→治理策略”的逻辑，文章探索性地提出了大模型风险治理框架，并根据模型风险影响周期和影响范围，提出“放开、管控、试点、迭代”4类差异化策略。

当前，全球人工智能大模型行业竞争日趋激烈，语料库成为提升人工智能大模型技术性能和应用效果的关键。但是，我国语料库在数量和质量上均存在不足，难以满足快速发展的人工智能大模型训练需求。从全球来看，各国都在加快语料库发展，特别是推动高质量语料库的建设和应用。因此，文章基于国外对标和国内环境分析，从平台定位、总体架构、运营主体、核心内容等维度提出建设国家级语料库运营平台的建议。

为解决传统评估系统内在模块相互耦合、无法重复利用、难以适应多领域评估需求的难题，设计了一个“数据+服务+应用”三层解耦的评估技术框架。 提出基于关键事件对音视频数据进行切片并抽取高价值信息的策略 ，解决了海量非结构化评估数据加工问题； 给出通用与专用相结合的评估模型模板设计 方法 ，有效应对评估模型的普适性应用难题； 构建了专家知识驱动的定性定量相结合的综合集成研讨评估环境 ，方便不同领域专家从不同维度对关键评估指标进行客观评判，提升了关键指标评估的全面性与客观性。基于技术框架设计开发联合训练仿真评估原型系统，验证了方法的有效性。