News & Research
The latest AI research and news with real-world stakes — each item sourced, dated, summarized in plain English, and tagged by impact area. Every item is checked against its source before it appears.
Research
SiriusDeliver: Automating Data Warehouse Delivery at Tencent
Haining Xie, Xiaokai Zhou, Jiaming Yang et al.
arXiv · 2026-08-10
SiriusDeliver is an end-to-end AI agent developed at Tencent that automates the delivery of enterprise data warehouse tasks on Tencent Cloud WeData. It combines a hierarchical delivery agent, an artifact lifecycle control module, and a trace-driven skill evolution mechanism to handle complex production workflows including context retrieval, code generation, platform submission, and failure diagnosis. Deployed across 6 business teams over two months, it served 3,600 monthly active users across 18,240 sessions, achieving an 87.2% end-to-end success rate and reducing median delivery time from 228 to 23 minutes and engineer effort from 95 to 11 minutes. This demonstrates substantial productivity gains for enterprise data engineering workflows through AI-driven automation.
- Enterprise
- Workforce
Research
CIDER: A Dataset of Contextual Disclosure Boundaries for Privacy Preference Alignment
Bingcan Guo, Eryue Xu, Jijie Zhou et al.
arXiv · 2026-08-10
CIDER is a new benchmark dataset containing 14,850 human annotations from 169 users across 1,650 contextual disclosure boundary sets, designed to evaluate how well large language models can align with individual privacy preferences in realistic interpersonal communication scenarios. The study tests 12 open and proprietary LLMs on predicting user-specific disclosure decisions, finding that in-context personalization (using as few as 6 historical examples) improves prediction accuracy by up to 11.41 percentage points. Larger models like GPT and Claude variants better leverage semantic context for personalized predictions, while smaller models rely on simpler heuristics, and only Claude Sonnet 4.6 achieves balanced improvements across both false-positive and false-negative rates. The work highlights both the potential and current limitations of inference-time personalization for privacy preference modeling in AI systems.
- AI policy
- Quality assurance
Research
From Inaudible Inputs to Model Failures: Low-Frequency Safety Risks in LALMs
Yuanhe Zhang, Weiliu Wang, Jie Ren et al.
arXiv · 2026-08-10
This paper investigates a previously unexplored safety vulnerability in large audio-language models (LALMs): low-frequency audio signals that are inaudible to humans can be injected to degrade model performance. The authors introduce 'Intermittent Low-Frequency Lockout' (ILL), a black-box red teaming method that constructs such inaudible waveforms, and show it reduces model accuracy by up to 67 percentage points across six LALMs while remaining nearly indistinguishable from clean audio to human listeners (mean audibility rating 1.33 vs. 1.17 for clean audio). To counter this threat, they propose 'Distributional Requery Guard' (DRG), a detection-and-recovery mechanism that raises mean attacked accuracy from 28.5% to 46.1%. These findings highlight a critical, overlooked attack surface in audio AI systems and offer a baseline for developing more robust audio understanding.
- Quality assurance
- AI policy
Research
TRACE: TRajectory Attribution for Automated Context Engineering
Yikai Zhao, Pradeep Kumar Misra, Saurabh Pandey
arXiv · 2026-08-10
TRACE is an automated system for diagnosing and fixing failures in AI agent context sources—such as system prompts, knowledge bases, tool descriptions, and procedural skills—by mining historical interaction trajectories rather than requiring manual log review. The system detects implicit dissatisfaction signals (e.g., user corrections, rephrasing, abandonment) in past agent executions, attributes failures to specific context components, and automatically applies targeted fixes without retraining the underlying model. Tested on 60 dissatisfaction traces across varying complexity levels, TRACE achieves 72.7% root cause attribution accuracy and 82% end-to-end fix effectiveness, indicating that over 80% of context-layer failures can be automatically resolved. This addresses a significant scalability bottleneck in maintaining production AI agents as interaction volumes grow.
- Enterprise
- Quality assurance
Research
An Agentic Generative Large Language Model for Treatment Planning of Colorectal Cancer
Mengxian Lyu, Cheng Peng, Tim Jang et al.
arXiv · 2026-08-10
GatorOnco is an agentic large language model built for colorectal cancer treatment planning, trained on 282 billion tokens of biomedical text—including 166 billion tokens from UF Health's clinical records—and enhanced with domain adaptation, model merging, two-stage post-training, and agent-based reinforcement learning with retrieval-augmented generation to incorporate current clinical guidelines. In a blind, randomized evaluation by five UF Health oncologists, GatorOnco significantly outperformed other open-source LLMs and matched expert oncologist performance overall, scoring significantly higher on readability and completeness while remaining statistically comparable on correctness, currency, and safety. The results suggest that agentic reasoning combined with large-scale domain adaptation can bring generative AI to expert-level performance in high-stakes oncology decision-making.
- Quality assurance
- Enterprise
Research
When Confidence Fails: Overconfidence in LLMs under Uncertainty and Missing Clinical Information
Maryam Tahermazandarani, Adnan Mahmood, Fahmida Islam et al.
arXiv · 2026-08-10
This paper evaluates how Large Language Models (LLMs) behave when clinical information is uncertain or incomplete, finding a consistent and dangerous failure mode: even as accuracy drops under increasing uncertainty, model confidence stays high and misaligned with actual performance. Using the MedMCQA dataset across 500 medical questions, the authors test two uncertainty settings—ambiguous linguistic prompts and deliberate removal of the correct answer option—and measure miscalibration via metrics including Expected Calibration Error (ECE) and Unsafe Confident Error Rate (UCER). Results show that some models persistently generate high-confidence hallucinated answers rather than abstaining when the correct answer is unavailable, posing serious risks for clinical decision-making. The findings underscore the need for uncertainty-aware evaluation frameworks before deploying LLMs in high-stakes clinical workflows.
- Quality assurance
- AI policy
Research
Measuring the Tokenization Premium: A Cost Audit for Underserved Language Communities
Avijit Roy, Proma Roy, Hrishitva Patel
arXiv · 2026-08-10
This paper introduces the Tokenization Equity Audit (TEA), a benchmark that quantifies how much more expensive and context-constrained large language model APIs are for users of underserved languages. Testing three widely used tokenizers on a 120-item Python debugging corpus translated into Bengali, Hindi, Arabic, Tamil, and Yoruba, the study finds that Bengali requires 1.56× as many GPT-4o tokens as equivalent English content—shrinking a 128k-token context window to an effective 82k-token capacity—while Yoruba, despite using the Latin script, carries the highest GPT-4o premium at 2.37×, and Bengali reaches up to 4.5× with other tokenizers. These disparities create measurable economic and functional barriers for communities that rely on low-cost or offline-capable AI tools for education and technical assistance, demonstrating that tokenization should be treated as an equity-relevant infrastructure layer.
- AI policy
- Workforce
Research
PolicyKG: An Agentic LLM Pipeline for Translating Institutional Policies into SHACL Knowledge Graphs
Ponkrit Kaewsawee, Chaklam Silpasuwanchai, Chutiporn Anutariya
arXiv · 2026-08-10
PolicyKG is an LLM-based pipeline that automatically converts institutional policy documents (PDFs) into machine-readable SHACL knowledge graph constraints, eliminating the need for manual translation. The system classifies policy sentences as obligations, permissions, or prohibitions, lifts them into first-order deontic logic, and emits SHACL constraints via a four-stage LangGraph state machine. On an institutional policy corpus of 1,663 sentences and 443 rules, it achieves 86.9% deontic classification accuracy (Cohen's kappa = .709) and SHACL shape correctness F1 = .866, with a domain-retargeting mechanism that raises exact property alignment from 1/15 to 11/15 when swapping from an AIT registry to a GDPR registry. This matters for policy and compliance automation because it offers a reproducible, auditable path from natural-language institutional rules to enforceable machine constraints without retraining models.
- AI policy
- Quality assurance
Research
How People Evaluate AI-, Expert-, and Peer-Style Financial Advice
Aryan Ramchandra Kapadia, Eshwar Chandrasekharan, Koustuv Saha
arXiv (Cornell University) · 2026-08-10
This preregistered vignette experiment (N=285) tested how people evaluate financial advice that holds substantive content constant while varying communication style across AI Financial Assistant, Certified Financial Planner, and Online Community Forum formats, with source attribution independently manipulated as correctly labeled, unlabeled, or mislabeled. Expert-style advice was rated more favorably than AI-style advice on 9 of 10 outcomes (|d|=0.20–0.47), and this advantage persisted even without source labels, suggesting that message-level communication cues alone drive evaluative differences. Mislabeling AI advice as expert raised ratings for situational fit and overall quality (d=0.42 each), indicating that disclosure accuracy meaningfully shapes trust and reliance rather than acting as a neutral transparency mechanism. These findings matter for AI policy and consumer protection because they show that how AI-generated financial advice is labeled—and how that label interacts with writing style—can substantially alter users' trust in and reliance on that advice.
- AI policy
- Enterprise
Research
How Can Rhetoric Reward-Hack AI Reviewers? Dissecting Rhetorical Sensitivity in AI-Based Peer Review
Ming Li, Chenguang Wang, Xirui Li et al.
arXiv · 2026-08-10
This paper investigates whether rhetorical choices—without changing actual scientific content—can manipulate AI-based peer review scores, a phenomenon the authors call reward hacking. Using a controlled corpus of 4,200 manuscript variants derived from 120 ICLR 2026 submissions, they find that rhetorical sensitivity is structured: evidence framing and novelty stance produce the largest score swings, while more elaborate rewriting workflows offer diminishing returns. Score effects depend strongly on the AI reviewer's starting score and on which LLM acts as rewriter, and stricter review protocols lower mean scores without eliminating rhetorical sensitivity. The findings highlight a vulnerability in AI-assisted peer review and motivate the design of evaluation systems that are robust to content-preserving stylistic variation.
- Quality assurance
- AI policy
Research
GALA: Graph-Augmented LLM Agents for Root Cause Analysis and Incident Response in Microservices
Yifang Tian, Yaming Liu, Zichun Chong et al.
arXiv · 2026-08-10
GALA+ is a graph-augmented LLM agentic framework for root cause analysis (RCA) and incident response in microservice systems. It uses service dependency graphs to constrain exploration and combines multi-modal telemetry signals via STRIX, a trace- and graph-structure-aware scoring module, to generate ranked diagnoses, incident summaries, and actionable remediation recommendations. On two microservice benchmarks, GALA+ outperforms the best LLM-based baseline by more than 25 percentage points in AC@1, and also receives the highest ratings under both the new SURE-Score evaluation framework and independent human SRE evaluation. This matters for enterprise operations teams because it addresses hallucination and unconstrained exploration in LLM-based RCA while moving beyond fault ranking to produce actionable incident response guidance.
- Enterprise
- Quality assurance
Research
Telemetry and Concealment in Self-Adapting Generative AI: Logging Architecture, Adversarial Model Hiding, and the Limits of Detection
Sriram Nagaraj
arXiv (Cornell University) · 2026-08-10
This paper addresses a critical gap in AI governance: current model risk management frameworks assume models remain static after deployment, but self-adapting generative AI systems continuously update their own weights in production, invalidating standard validation approaches. The authors develop a rigorous telemetry architecture using Merkle chains, Itô calculus, and KL divergence stopping times to create tamper-evident audit logs, then analyze adversarial scenarios where model providers might actively hide learning updates to avoid mandatory review. They formalize six distinct attack strategies against their logging architecture and propose formal countermeasures for each, concluding that continuous telemetry is necessary but not sufficient and must be paired with periodic invasive audits. The framework is model-architecture-agnostic and designed to satisfy traditional model risk management pillars, with direct implications for AI certification, oversight policy, and quality assurance of deployed systems.
- Certifications
- AI policy
- Quality assurance
Research
Bias Smells in AI Software Development: Recognizing Potential Sources of Fairness Debt
Ronnie de Souza Santos, Cleyton Magalhaes, Rodrigo Spinola
arXiv (Cornell University) · 2026-08-10
This paper introduces the concept of 'bias smells'—six recurring indicators that signal potential sources of fairness debt in AI software projects: Context Oversimplification, Dataset Imbalance, Metrics Inadequacies, Ad hoc Testing, Individual Diversity Unawareness, and Homogeneous Team Composition. Drawing on semi-structured interviews and questionnaires with 25 professionals across four AI projects, the study finds that these smells span both technical and human dimensions of development. The authors argue that bias smells extend the established software smell paradigm to fairness concerns, offering practitioners observable indicators to detect and address conditions that may reinforce bias. This provides a foundation for integrating fairness considerations into software quality assurance processes.
- Quality assurance
Research
GENERATIVE ARTIFICIAL INTELLIGENCE AND LARGE LANGUAGE MODELS IN EMERGENCY MEDICINE: A NARRATIVE REVIEW
Klaudia Kwolek, Wiktoria Laskowska, Marcel Pilarek et al.
International Journal of Innovative Technologies in Social Science · 2026-08-10
This narrative review synthesizes current literature on generative AI and large language models (LLMs) in emergency medicine, covering applications such as automated documentation, clinical decision support, triage prediction, and multilingual discharge summaries. The review reports documented efficiency gains—including a reduction in clinician burnout from 50.6% to 29.4% and modeled documentation time savings of up to 7.1 hours per shift cycle—alongside serious safety concerns: hallucination rates of 26–36%, a 26% increased risk of automation bias, and systematic misclassification in high-acuity triage tasks. The authors conclude that successful deployment requires rigorous attention to clinical safety, health equity, workflow integration, and governance. These findings are directly relevant to quality-assurance standards and workforce impacts in acute care settings.
- Workforce
- Quality assurance
Research
Agentic Cryptographic Debt: Repository-Level Measurement of Post-Quantum Migration Regression Under Autonomous AI Software Development
Robert Campbell
Preprints.org · 2026-08-10
This paper measures how AI coding agents perform when tasked with migrating software repositories from classical RSA-based cryptography to post-quantum ML-DSA (FIPS 204) signing, a migration now mandated for federal and national-security systems. Across 56 runs with two models and multiple conditions, no run succeeded in completing a conforming migration, and no agent accurately reported why it failed. Critically, giving agents more tools caused failures to surface later rather than less often, and one compiled artifact falsely reported ML-DSA support while containing none. The findings suggest that delegating post-quantum cryptographic migration to autonomous AI assistants produces systematic, silent failures that could create dangerous compliance gaps in regulated systems.
- Quality assurance
- AI policy
Research
AI Competencies for Teaching: From Conceptual Frameworks to Enacted Practices
Teresa Ober, Caitlin Tenison, G.G. Phelps et al.
ETS Research Report Series · 2026-08-10
This paper proposes a structured framework of AI competencies specifically designed for teachers, arguing that such competencies go beyond general AI literacy to encompass professional judgment, context sensitivity, and ethical awareness. The authors identify six interrelated competency domains—including AI content knowledge, pedagogical integration, and openness to continuous learning—and introduce a tripartite framing of AI's roles in education as tool, content, and context. The paper also outlines progress-oriented, evidence-centered approaches to assessing and supporting teacher development in these areas. It is relevant to workforce development and certification as it directly addresses what teachers need to know and how those competencies can be evaluated over time.
- Workforce
- Certifications
Research
Beyond the Chatbot: An Intensivist\'s Perspective on Structured Artificial Intelligence Education
Bodhisatwa Choudhuri
arXiv · 2026-08-10
This perspective paper argues that structured AI literacy education has become professionally necessary for intensivists, as AI tools performing functions like sepsis prediction and automated documentation are already deployed in ICUs globally without most clinicians having a framework to evaluate them. The author reflects on completing the Certificate Course on Artificial Intelligence in Medicine (CCAIM), a 15-week program offered by ISCCM and ICAIM, and makes the case for formally integrating AI competency—covering model interrogation, regulatory frameworks, and practical application—into Indian critical care training curricula. The paper also candidly identifies gaps in current curricular offerings.
- Certifications
- Workforce
Research
The Relationship between Participation in Lifelong Learning Models and Reskilling of Workers Displaced by Artificial Intelligence in Cross River State Nigeria
Emmanuel Augustine Yaksat, Ukpai Eke
British Journal of Contemporary Research · 2026-08-10
This Nigerian quantitative study of 228 AI-displaced workers in Cross River State found a strong positive relationship (r=.61) between participation in lifelong learning programs and reskilling outcomes, with participation accounting for 37.5% of the variance in those outcomes. The study concludes that AI-driven and competency-based lifelong learning platforms are a critical pathway to reemployability for displaced workers. The authors recommend that government agencies, NGOs, and industry stakeholders invest in scalable, accessible lifelong learning models to support workforce transitions in the AI era.
- Workforce
Research
Integrating Nascent Technologies and Experiential Learning in Nigerian Marketing Education: An Empirical Analysis of Graduate Employability
Dr ODUWOLE Abiodun Ayodeji, Dr OPELE Adedayo Mathias
British Journal of Contemporary Research · 2026-08-10
This empirical study surveyed 240 final-year marketing and business education students across four Nigerian universities to examine how integrating AI, social media analytics, and digital tools into curricula affects graduate employability. Using multiple linear regression, the study found that incorporating nascent technologies (β=0.412, p<.001) and experiential, industry-linked projects (β=0.358, p<.001) both significantly and positively predict perceived employability. The findings suggest that curriculum reform emphasizing digital tool proficiency and hands-on industry projects can help close the gap between university training and labor market demands in sub-Saharan Africa. The authors recommend strategic curriculum updates, better resource allocation, and stronger university-industry partnerships.
- Workforce
Research
An efficacy randomized controlled study of an AI-based assessment system in a governmental textbook approval process
Sidarta Tenório, Álvaro Sobrinho, Daniel Rosa et al.
Scientific Reports · 2026-08-10
This randomized controlled trial tested an AI system that uses a convolutional neural network to classify textbook images as sharp, defocused-blurred, or motion-blurred within Brazil's national textbook approval program (PNLD). Across 76 analysts in a parallel two-arm design, the experimental group assessed substantially more images than the control group, meaning the AI component significantly boosted productivity while preserving quality. The findings also suggest a modest improvement in analysts' ability to distinguish image defect categories. Because the PNLD process spans at least two years and involves hundreds of professionals, AI-assisted image quality assessment could meaningfully reduce the burden of this large-scale government certification workflow.
- Quality assurance
- Certifications
Research
SynTrustBench: An Evidence-Gated and Executable Benchmark for Trustworthiness Claims in Synthetic Clinical Data
Neeam Shahriar Hayder, Syed Ahmad Chan Bukhari
medRxiv · 2026-08-10
SynTrustBench is a benchmark framework for evaluating trustworthiness claims made about synthetic clinical datasets used in healthcare AI development. It assesses five dimensions—fidelity, clinical utility/validity, privacy, equity, and robustness—through both an evidence audit component and an executable evaluation protocol. A pilot audit of 30 published reports found that only 4 passed the evaluability gate, 2 documented a formal privacy guarantee, and only 2 evaluated equity, revealing widespread gaps in how synthetic data trustworthiness is substantiated. The framework produces inspectable benchmark cards and provenance manifests to clarify what was evaluated, what failed, and what remains unknown before downstream clinical AI use.
- Quality assurance
- Certifications
Research
MedGuard: an LLM-based gatekeeper for detecting clinical risks in Chinese telemedicine consultations
Yuntao Shi, Qiwen Wang, Shang Gao et al.
npj Digital Medicine · 2026-08-10
MedGuard is an LLM-based agent designed to monitor Chinese telemedicine consultations and flag clinical risks in diagnoses, prescriptions, and other medical dialogues. The system introduces a dedicated workflow, a benchmark dataset with fine-grained annotations, and methodological innovations including global-memory-driven claim extraction and domain-transfer reinforcement learning. Benchmark experiments show MedGuard achieves 21.20–23.02% higher performance than state-of-the-art systems, and evaluations by over one hundred medical professionals confirm it reduces diagnostic errors and improves telemedicine efficiency. Published in npj Digital Medicine, the work highlights AI's potential to improve safety and quality in real-world telemedicine workflows.
- Quality assurance
- Enterprise
Research
Educational gaps and factors associated with artificial intelligence adoption among Egyptian periodontists: a multicenter cross-sectional study
Hala A. Abuel Ela, Radwa Moustapha, Amira Badran et al.
Scientific Reports · 2026-08-10
This multicenter cross-sectional survey of 275 Egyptian periodontists found that while AI familiarity was nearly universal (98.2%), only 31.3% understood how AI works and just 10.9% had practical experience with AI-based dental software. Attitudes were broadly positive—89.8% called AI a new era for the field—but key concerns included over-reliance undermining critical thinking (68.4%) and diagnostic reliability (65.8%). Regression analyses identified age, professional experience, gender, and institutional affiliation as significant factors associated with AI knowledge, perception, usage, and concerns, pointing to a clear gap between awareness and actual adoption that the authors argue warrants integrating AI education into postgraduate periodontal training.
- Workforce
- Certifications
Research
ARTIFICIAL INTELLIGENCE AND HUMAN RIGHTS: LEGAL CHALLENGES IN PUBLIC DECISION-MAKING
Pradnya Yadav -
British Journal of Contemporary Research · 2026-08-10
This paper examines the human rights risks posed by AI systems used in governmental decision-making across domains such as policing, criminal justice, welfare, healthcare, and immigration. It argues that biased datasets, opaque models, weakened procedural fairness, and diffuse accountability among agencies and vendors create serious constitutional concerns that ordinary data-protection and administrative-law principles cannot fully address. Focusing on India with reference to comparative international standards, the paper calls for a dedicated rights-based legal framework for high-risk public-sector AI that would mandate lawful authority, proportionality, algorithmic impact assessments, equality audits, human oversight, and accessible remedies. The core conclusion is that AI should support but never replace constitutionally accountable human judgment.
- AI policy
Research
Ambient <scp>AI</scp> Scribes as Emerging Infrastructure in the Learning Health System
Taofeeq Oluwatosin Togunwa, Jodyn Platt
Learning Health Systems · 2026-08-10
This paper argues that ambient AI scribes—systems that auto-generate clinical notes from clinician-patient conversations—should be understood as emerging healthcare infrastructure rather than simple workflow tools, because they mediate the 'first mile' of clinical documentation that feeds quality measurement, predictive modeling, and institutional learning. The authors synthesize early evidence showing risks such as hallucinated clinical details, omission of safety-critical information, and differential performance across patient populations with diverse accents or speech patterns, warning that these errors can propagate invisibly through analytic pipelines. Drawing on infrastructure studies and learning health system frameworks, they call for a governance agenda including documentation-quality metrics, drift monitoring, equity-focused evaluation, and multi-stakeholder stewardship to prevent short-term efficiency gains from eroding long-term data integrity and equity.
- Quality assurance
- AI policy