We present Iris-mini and Iris-pro, two search agents trained at the 35B-A3B and 397B-A17B scales, together with the data pipeline and training recipe behind them. Tasks are reverse-constructed from the hyperlink structure of a web corpus: we author multi-hop chains over an entity graph distilled from a seed page and its out-links, rewrite every non-answer entity into a descriptive reference so that no clue can be resolved by string matching, and admit only questions that a reference model fails closed-book yet solves once the supporting evidence is supplied. These questions are then turned into trajectories, which are filtered at both the trajectory and the turn level before SFT. The policy is then optimized by RL against live search, with the reward judge and the observation summarizer served inside the training cluster, and with over-long rollouts interrupted at the request level and resumed from their committed prefix at the next step. We alternate the two stages in a procedure we call SFT-RL climbing, returning the hardest solved and most efficient rollouts of each RL round to the next supervised pass. Because inference-time context management is worth more on these benchmarks than most reported differences between systems, we evaluate every benchmark both with and without it, holding the tool set, the context limit, and the judge fixed. All results come from a single ReAct agent, with no sub-agents and no test-time verification. With management enabled, on BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE the two models reach $82.2/84.8/86.9/52.3$ and $88.6/85.1/92.9/56.4$, the strongest overall results among open-source search agents in their respective parameter ranges. We plan to release the model weights together with the complete recipe for data construction, training, and evaluation.
Ziyuan Liu, Hengqi Liu, Zichuan Wang et al.· 0 citations
Ensuring the security of the power system is essential for stability and reliability, especially in the event of disruption. Effective classification of contingency in power systems enables proactive decision-making and mitigates large-scale breakdowns and failures. This study explores the use of machine learning algorithms to classify security levels of contingencies in power systems into safe, moderate or severe classes. For this approach, Newton-Raphson load flow method extracts system data from contingency scenarios, using Overall Performance Index (OPI) as safety measure. For data pre-processing, Synthetic Minority Over-Sampling Technique (SMOTE) and Principal Component Analysis (PCA) is used to address class imbalance and reduce dimensionality, respectively. K-Nearest Neighbours (KNN), Random Forest (RF) and Support Vector Machines (SVM) is trained and evaluated on datasets generated through N-k contingency scenarios for k equal 1, 2, and 3 on IEEE-14 and IEEE-30 bus systems using four hybrid pre-processing configurations: normalized, SMOTE-balanced, PCA-transformed, and a combined SMOTE PCA-transformed. Performance is assessed by precision, recall and F1 score, with priority given to the severe contingency classes. The RF achieved the highest F1 scores of 0.97 in IEEE-30 and 0.86 in IEEE-14, SVM benefits significantly from PCA and improves the accuracy of the classification, while KNN is best suited for SMOTE and PCA conversion. The findings show that PCA contributes more than SMOTE to the overall performance of the model. However, SMOTE improves recall but can introduce false positives and is therefore a compromise of accuracy. This study highlights machine learning as a scalable and powerful alternative to traditional contingency analysis, which improves the assessment of security in real time.
Evaluating agents on the growing number of agentic benchmarks is challenging because they often require complex environments and agent integrations. We introduce Harbor Adapters, a unified evaluation infrastructure for agentic benchmarks. Our work makes three contributions. First, we develop benchmark adapters that port more than 80 benchmarks to evaluate arbitrary agents, and validate them through rigorous code review and parity experiments. Second, we conduct a large-scale evaluation of 8 models spanning capability tiers across 54 benchmarks; every model is run with Terminus-2 and with one of 3 native harnesses. This enables a broader analysis of agent capabilities and failure modes than was previously possible. Third, we introduce Harbor-Index, a curated set of 82 difficult, diverse, and high-quality tasks spanning 29 benchmarks, refined from the adapted suite through difficulty filtering, AI and human audit, and an audit-and-fix loop. Harbor-Index preserves the challenge and breadth of large-scale agentic evaluations while being affordable to run; no evaluated model-harness configuration exceeds 30% pass rate, and the strongest (GPT-5.5 with Codex) reaches 28.0%. We release the adapters, evaluation results, in-depth analysis, and Harbor-Index as open-source artifacts to support more reliable and comprehensive evaluation of language-model agents.
Lin Shi (Audrey), Haowei Lin (Audrey), Zixuan Zhu (Audrey) et al.· 0 citations
Artificial intelligence in recruitment has shifted the object being automated from profile pairs and ranked lists to multi-stage workflows that retrieve evidence, compare candidates, and support or execute actions. This systematized narrative review traces that development from bilateral retrieval and behavioral ranking through neural person--job matching, large language model (LLM) components, and tool-using recruiting agents. Using a purposive search and coding protocol updated through 23 July 2026, plus targeted updates through 2 September 2026, we organize 40 representative works with supporting industrial and legal sources. This synthesis is not a prevalence estimate. We analyze three coupled transitions: from similarity to reciprocal suitability, from a model to a compound workflow, and from offline prediction to evidence- and productivity-aligned evaluation. Across document understanding, retrieval, ranking, assessment, interviewing, sourcing, and human handoff, we distinguish field-, pair-, list-, case-, trajectory-, and outcome-level evidence. Persistent gaps arise because behavioral labels confound exposure, preference, and qualification; private and synthetic data limit external validity; final-output scores conceal pipeline failures; and, within the coded set, privacy is not directly evaluated and no row jointly evaluates utility, fairness, privacy, and security. These observations describe the coded set rather than the field as a whole. We therefore introduce a staged mapping from evaluation evidence to the strongest defensible claim, together with an agenda for reciprocal, evidence-grounded, temporally controlled, selective, and auditable systems. Progress should be judged by whether workflows retrieve the right evidence, preserve uncertainty, support contestable decisions, and improve outcomes under explicit cost and risk constraints.
Ziyi Zhao, Guanzheng Wei· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
This technical report presents EXAONE Forecast for Finance (EXAONE Finance), a financial time series (TS) foundation model (TSFM) tailored to financial forecasting. Recent TSFMs achieve strong zero-shot performance through large-scale pretraining. However, they are primarily developed for general-domain TS and largely rely on self-attention backbones whose computational cost grows quadratically with sequence length and variate count. Moreover, they assume fully observed inputs and are pretrained on corpora that fail to capture the unique dynamics of financial markets. These limitations hinder their applicability to finance, where long, many-channel, intermittently observed panels are common. To address these challenges, EXAONE Finance adopts an attention-free architecture, replacing self-attention with two simple yet effective linear-time operators: 1) a causal 1D convolution for temporal mixing and 2) a group-aware pooling multi-layer perceptron (MLP) for variate mixing. Furthermore, a masked context augmentation exposes the model to contiguous missing spans during training, improving robustness to the missingness pervasive in financial markets. EXAONE Finance is pretrained on a large-scale financial corpus covering not only equities but also foreign exchange, commodities, crypto-assets, fixed income, and macroeconomic indicators. On FinVerse, a financial forecasting benchmark covering diverse asset classes, EXAONE Finance attains state-of-the-art performance, ranking first across all three evaluation tiers---point-forecast accuracy, cross-sectional asset ranking, and portfolio profitability.
Seunghan Lee, Jaehoon Lee, Jun Seo et al.· 0 citations
Abstract Introduction Artificial intelligence (AI) systems increasingly mediate online speech, shaping visibility, participation, and recognition in digital public spaces. For LGBTQI+ communities, AI-based content moderation operates within contexts marked by persistent homophobia and transphobia, raising concerns about bias, exclusion, and uneven protection from harm. This article examines the emerging scientific literature on AI-based moderation of LGBTQI+phobic discourse through the lens of sociotechnical governance. Methods The study employs a bibliometric and sociotechnical mapping approach guided by the PRISMA 2021 framework. A systematic search of the Scopus database identified 23 peer-reviewed publications published between 2020 and 2025. Bibliometric indicators, co-authorship and keyword network analyses were combined with a critical sociotechnical perspective to examine the intellectual structure, epistemic authority, methodological trends, and themes characterising sociotechnical governance in this emerging research field. Results Findings indicate a highly concentrated research field dominated by a small number of countries, institutions, and computational disciplines, reinforcing geopolitical and linguistic asymmetries in scientific production. The literature indicates an increasing research focus on large language models, explainable AI, and counter-narrative generation, reflecting a shift in scholarly attention from hate speech detection to broader approaches to AI-enabled content governance. However, links between AI moderation, mental health, stigma, and well-being remain weakly integrated, and participatory approaches are scarce. Conclusions The bibliometric evidence indicates that the emerging scientific literature increasingly conceptualises AI-based moderation as a form of digital governance with potential implications for LGBTQI+ communities. The mapped literature also highlights persistent concerns regarding accountability, inclusiveness, and interdisciplinary integration, suggesting important directions for future research and policy development. Policy Implications The mapped literature suggests that future policies should promote transparent, auditable, and participatory AI governance, support multilingual and culturally inclusive datasets, and integrate content moderation into broader digital public health and anti-discrimination strategies.
David Ruiz‐Muñoz, Ana María Sánchez Sánchez, Francisca J. Sánchez‐Sánchez· Sexuality Research and Socia...· 0 citations
Minimally invasive D2 lymphadenectomy for gastric cancer is technically demanding, and its quality varies across patients, surgeons, platforms, and institutions. Conventional endpoints, including lymph node yield, margin status, operative time, blood loss, postoperative morbidity, and survival, remain essential but do not consistently capture intraoperative procedural fidelity, safety-critical deviations, or case complexity. This narrative framework review synthesized evidence from MEDLINE, Embase, and Web of Science from inception to May 10, 2026, focusing on technical difficulty, surgical quality assessment, pathology-centered oncologic adequacy, risk-adjusted outcomes, and artificial intelligence (AI)-enabled audit. Technical difficulty was conceptualized as a case- and context-dependent risk-adjustment layer shaped by anatomical complexity, vascular variation, therapy-altered tissue planes, visceral adiposity, operative platform, team workflow, and learning stage. Surgical quality was defined as a multidomain construct integrating process metrics, pathology metrics, and risk-adjusted clinical outcomes. The proposed framework provides audit-oriented guidance for identifying a simplified minimum dataset, prioritizing high-risk D2 segments for selective process review, and interpreting process, pathology, and outcome indicators together after adjustment for technical difficulty. AI may support scalable audit through video indexing, phase and step recognition, extraction of high-risk operative segments, assisted event logging, and structured feedback. However, AI outputs should be treated as candidate measurement signals requiring human confirmation, expert surgical review, pathology-based assessment, external validation, governance, and post-deployment monitoring. Future validation should proceed stepwise, from feasibility testing and inter-rater reliability assessment to prospective workflow evaluation and multicenter assessment of audit efficiency, benchmarking validity, and process- or patient-level outcomes. Quality assessment in minimally invasive D2 lymphadenectomy should shift from isolated surrogate endpoints toward an auditable, difficulty-adjusted framework that makes “D2 achieved” more measurable, reviewable, and clinically meaningful.
Yasheng Xue, Xiaoyun Dai, Xi Wang et al.· World Journal of Surgical On...· 0 citations
Title: Generative AI-Powered Pedagogical Agents in Immersive Environments for Social Sciences and Humanities Education: A Scoping Review Purpose This research project investigates how generative artificial intelligence (GenAI)-powered pedagogical agents and virtual instructors are being used within immersive virtual reality (VR) and augmented reality (AR) environments, specifically in the context of social sciences and humanities education. While research on generative-AI agents in immersive learning environments has grown rapidly since 2023, no existing systematic or umbrella review has yet mapped this specific intersection — most prior reviews either predate the generative-AI/large-language-model (LLM) era or focus predominantly on STEM, medical, and engineering education. Given that social sciences and humanities education (e.g., history, geography, civics) involves distinctive pedagogical and ethical demands — such as historical empathy, multi-perspective reasoning, and open-ended interpretive dialogue — that differ meaningfully from technical or procedural training domains, the project aims to determine what is currently known about this intersection, how mature the field is methodologically, and where meaningful gaps remain. To address this aim, the study was designed as a scoping review, following Arksey and O'Malley's (2005) five-stage methodological framework and reported according to the PRISMA-ScR (PRISMA Extension for Scoping Reviews) guideline. A scoping-review design was chosen deliberately over a systematic review or meta-analysis because the objective is to map the breadth, characteristics, and trends of an emerging body of literature — rather than to statistically synthesize effect sizes or assess a narrow effectiveness question — which is appropriate given how new and heterogeneous this specific research area still is. Research Questions The project is guided by four research questions: RQ1: What are the design features (embodiment, mode of interaction, role assumed) of GenAI-based pedagogical agents/virtual instructors used in immersive VR/AR environments in social and humanities education? RQ2: What learning outcomes have been reported in studies on these agents, and in which direction do the findings trend? RQ3: At which educational levels and in which social sciences/humanities subfields have these studies been conducted? RQ4: What are the methodological trends and limitations in the field, and what directions are recommended for future research? Methodology A systematic search was conducted across Scopus, Web of Science, and ERIC (August 2026), combining terms related to pedagogical agents/virtual instructors, generative AI/LLMs, immersive VR/AR/XR technologies, and education. The search was restricted to English-language, peer-reviewed journal articles published between 2023 and 2026 — a window chosen to capture the generative-AI/LLM era specifically. Of 105 records initially identified, a multi-stage screening and eligibility process (title/abstract screening, full-text assessment, and data-charting verification) resulted in 9 studies meeting all inclusion criteria. Data extracted from each study included agent design characteristics, technology used, research design, educational level and subject area, reported learning outcomes, and author-stated limitations and future-research recommendations. Findings were synthesized narratively (rather than statistically) around the four research questions and subsequently interpreted through the theoretical lenses of Presence Theory and Embodied Cognition. Expected/Actual Outcomes The review's findings indicate that the included agents are predominantly designed as embodied 3D characters built on GPT-family models, most often assuming peer or mentor roles within VR environments. Reported effects on learning outcomes (motivation, engagement, partner perception, and, in some cases, academic performance) trend positive overall, though effect sizes vary considerably across studies and are notably smaller in the few studies employing control-group comparisons than in single-group, pre-/post-test designs. A key substantive finding is that the existing literature is concentrated almost entirely at the higher-education level and clusters around language education and AI ethics/literacy — it has not yet reached classic social-studies subfields such as history, geography, or civics education, despite the conceptual gap the study set out to address. Interpreted through Presence Theory and Embodied Cognition, the findings further suggest that an agent's educational impact depends less on its technical sophistication (e.g., visual realism) than on whether an appropriate balance between presence and embodiment has been achieved relative to the nature of the learning task — an "embodiment paradox" identified across several included studies. The project's broader contribution is threefold: (1) it provides the field's first dedicated mapping of the generative-AI/LLM generation of pedagogical agents within the social sciences/humanities education context, filling a gap left by earlier, pre-generative-AI-era reviews; (2) it offers a theoretically grounded interpretive lens (Presence Theory/Embodied Cognition) for understanding why and how these agents affect learning, rather than only cataloguing whether they do; and (3) it identifies concrete directions for future research — including extending investigation to K-12 contexts, directly targeting classic social-studies content, adopting more rigorous control-group designs, and incorporating physiological/multimodal measures alongside self-report data. The review also transparently documents its own methodological limitations (a single-researcher screening stage, no prior protocol registration, no formal quality/risk-of-bias appraisal, and a modest final sample of nine studies), consistent with the exploratory nature of scoping reviews and intended to guide readers in appropriately weighing the strength of the evidence presented.
Veysel Dağdemir, Erhan Görmez· Open Science Framework· 0 citations
This systematic literature review examines research on artificial intelligence (AI) in mathematics education published between January 1, 2021, and August 1, 2025. Searches of Scopus and Google Scholar identified 922 records; after deduplication, screening, and full-text eligibility assessment, 42 peer-reviewed journal articles and conference proceedings were included. The review used descriptive quantitative summaries and a deductive-inductive thematic synthesis. Two independent reviewers conducted screening (Cohen's kappa = 0.88), and methodological quality was assessed using the Mixed Methods Appraisal Tool (MMAT). The included literature indicates increasing attention to generative AI, personalised support, feedback, teacher practice, academic integrity, and equity. Evidence for educational benefits varies substantially across study designs and contexts, and technical capability should not be equated with demonstrated classroom effectiveness. Geographic patterns in the selected sample are described without attributing them to regulatory, economic, or infrastructural causes that were not directly tested. Because the 2025 search covered only January through August 2025, publication counts are treated as partial-year data and are not directly comparable to complete prior years. Key limitations include reliance on two databases, English- and Russian-language restrictions, reproducibility constraints in Google Scholar, methodological heterogeneity, and limited long-term evidence. Overall, AI shows potential to support mathematics teaching and learning, but stronger longitudinal and comparative evidence is needed to establish effectiveness, equity, and sustainable implementation.
Fariza Omirzakova, Sarsenkul Sh. Tleubai· European Journal of STEM Edu...· 0 citations
The fields of human-robot interaction (HRI) and embodied conversational agents (ECAs) have long studied how empathy could be implemented in machines. One of the major drivers has been the goal of giving multimodal social and emotional intelligence to these artificially intelligent agents, which interact with people through facial expressions, body, gesture, and speech. What empathic behaviors and models have these fields implemented by mimicking human and animal behavior? In what ways have they explored creating machine-specific analogies? This chapter aims to review the knowledge from these studies toward applying the lessons learned to today’s ubiquitous, language-based agents such as ChatGPT.
Angelica Lim, Özge Nilay Yalçın· Cambridge University Press e...· 0 citations
Advances in the performance of large language models (LLMs) have led some researchers to propose the emergence of theory of mind (ToM) in artificial intelligence (AI). LLMs can attribute beliefs, desires, intentions, and emotions. But, rather than employing the characteristically human method of empathy, they learn to attribute mental states by recognizing behavioral and linguistic patterns in a dataset. We ask whether LLMs’ inability to empathize precludes them from honoring an individual’s right to be an exception, from assessing character with appropriate sensitivity to a person’s individuality. We defend the following claims: (1) Both humans and LLMs can honor the right to be an exception. (2) Humans can honor that right through empathy, and LLMs cannot empathize. (3) There is distinctive moral value in honoring the right to be an exception through empathy.
Will Kidder, Jason D’Cruz, Kush R. Varshney· Cambridge University Press e...· 0 citations
Artificial Intelligence systems increasingly match or surpass human performance across a wide range of cognitive tasks. Recent advances in large language models have yielded conversational abilities often indistinguishable from human dialogue. In this chapter, we argue that whether artificial agents should be considered empathetic depends on how empathy itself is defined and how it relates to subjective experience. We show that different philosophical schools produce distinct, and often incompatible, answers to the question of whether such a system should be considered genuinely empathetic. On one hand, empathy is a functional trait that can be fully characterized by observable behavior; on the other hand, empathy is inseparable from subjective experience and conscious feeling. Drawing on the philosophical notion of the “zombie” we introduce a thought experiment involving a hypothetical chatbot that exhibits perfect empathic behavior across all conceivable benchmarks while lacking any subjective experience. This framing allows us to disentangle functional performance from conceptual attribution and to examine whether empathy judgments depend on observable behavior alone or on assumptions about inner experience. As artificial agents become increasingly integrated into emotionally contexts, understanding how and why humans attribute empathy to machines becomes a pressing scientific, ethical, and societal question.
Ariel Goldstein, Gabriel Stanovsky· Cambridge University Press e...· 0 citations
