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
Prompt Coverage Adequacy
Florian Tambon, Michael Konstantinou, Cedric Richter et al.
arXiv · 2026-07-02
This paper introduces Prompt Coverage Adequacy, a novel testing criterion designed for software developed through large language models (LLMs) and autonomous agents, where prompts rather than code are the primary development artifacts. Instead of measuring how thoroughly tests exercise lines or branches of code, it measures how well a test suite satisfies the requirements expressed in a prompt by leveraging LLM attention mechanisms. Evaluated across two datasets and multiple LLMs, the approach uncovers over 30% more faults than traditional code coverage when used to guide test generation. The findings suggest Prompt Coverage Adequacy could become a foundational metric for quality assurance in the emerging paradigm of LLM-driven software development.
- Quality assurance
Research
SPLIT: Cross-Lingual Empathy and Cultural Grounding in English and Ukrainian LLM Responses
Anna Chorna
arXiv · 2026-07-02
This paper introduces SPLIT, a 500-prompt benchmark designed to evaluate how well large language models (LLMs) provide emotionally grounded, culturally appropriate responses in English and Ukrainian across five crisis-related categories: Stress, Panic, Loneliness, Internal Displacement, and Tension. The authors evaluate three LLMs—Gemini-2.5-Flash, LLaMA-3.3-70B-Instruct, and DeepSeek-V3—on Empathetic Accuracy, Linguistic Naturalness, and Contextual & Cultural Grounding, finding that Gemini-2.5-Flash and LLaMA-3.3-70B-Instruct degrade in Ukrainian while DeepSeek-V3 remains comparatively stable. The study also finds that human and AI evaluators agree weakly on empathy and naturalness but diverge on cultural grounding, and argues that generating Ukrainian text is not equivalent to providing Ukrainian emotional support. These findings highlight significant gaps in LLM deployment for emotional support in low-to-mid-resource language contexts and call for more culturally tailored benchmarks and human-centered evaluation.
- Quality assurance
- AI policy
Research
Assessing VLM Reliability for Medical Image Quality Evaluation Under Corruption and Bias
Sofiane Ouaari, Kevin Vorwalder, Nico Pfeifer
arXiv · 2026-07-02
This paper benchmarks 16 vision-language models (VLMs) on medical image quality assessment (MIQA) under real-world degradation and contextual bias, using the MediMeta-C dataset zero-shot across seven corruption types, five severity levels, and seven imaging modalities. Results show pixelation caused the largest performance drops (mean -20.58%, up to -34.4% for OCT), while textual metadata such as institutional prestige or equipment age shifted scores by as much as +17.15% and -14.7% respectively, revealing that current VLMs lack objectivity and are sensitive to contextual cues that should not influence image quality judgments. Individual models showed extreme swings (up to +95.62% for InternVL-8B and -37.7% for MedGemma), and same-family models correlated at only 0.67–0.83, indicating inconsistency even within model families. These findings raise concerns about deploying VLMs for clinical MIQA without further safeguards, particularly given the tension between privacy-preserving transformations like pixelation and model reliability, and the risk of demographic or institutional metadata introducing bias into quality assessments.
- Quality assurance
- AI policy
Research
AIriskEval-edu: New Dataset for Risk Assessment in AI-mediated K-12 Educational Explanations
Javier Irigoyen, Roberto Daza, Francisco Jurado et al.
arXiv · 2026-07-02
This paper presents AIriskEval-edu-db2, a dataset of 1,639 instructional explanations drawn from 170 K-12 science, language arts, and social science questions, designed to train and evaluate LLM-based auditors for pedagogical risk assessment. Each question includes a human-teacher explanation and 11 LLM-generated explanations tied to distinct risk profiles, all assessed against a five-dimension rubric covering factual precision, depth, relevance, age-appropriateness, and ideological bias. A subset of 785 explanations includes structured explainability annotations with risk localization and descriptions, validated by expert teachers. Experiments compare proprietary frontier models against a fine-tuned local Llama 3.1 8B model, testing whether supervised fine-tuning can match stronger models while preserving privacy in educational auditing.
- Quality assurance
- Certifications
Research
Has This Checkpoint Been Abliterated? A Two-Signal Audit and Its Failure Map
Gabriel Hurtado
arXiv · 2026-07-02
This paper presents a two-signal method for auditing open-weight AI checkpoints to detect whether their refusal mechanisms have been removed (a process called 'abliteration') before deployment. The approach combines a reference-anchored activation refusal-gap with a weight-recovery energy metric derived from the difference between a base model and the candidate checkpoint, achieving an AUROC of 0.95 on a 273-checkpoint registry spanning Qwen, DeepSeek-distilled Qwen, Llama, and Gemma models — significantly outperforming either signal alone. The audit correctly identifies 53 of 57 public abliterations while maintaining a false positive rate of 0.11, and transfers to held-out model families at balanced accuracy 0.89. The authors also map two key failure modes — reference spoofing and adversarial fine-tuning past the threshold — framing the method as effective pre-deployment triage rather than tamper-proof security.
- Quality assurance
- AI policy
Research
CLAP: Closed-Loop Training, Evaluation, and Release Control for Domain Agent Post-training
Fangfei Li, Chenyang Zhao, Long Wang et al.
arXiv · 2026-07-02
CLAP (Closed-Loop Agent Post-training) is a framework designed to address key challenges in deploying domain-specific AI agents on noisy business data, including unreliable post-training gains, mismatches between offline and live performance, and risks from releasing updated model adapters. The method integrates data validation, reward diagnostics, offline evaluation gates, and application-chain replay into a unified loop that determines whether a fine-tuned adapter is safe to release. Tested on five anonymized manufacturing-scenario batches using QLoRA-style LoRA-SFT, the framework shows modest average improvements in overall score, pass rate, and evidence accuracy, but also reveals that only 3 of 5 batches improve, some regress, and certain training methods (GRPO) introduce high KL risks. The findings argue that domain-agent post-training should be governed by an integrated data-training-evaluation-release pipeline rather than relying on training completion or a single offline metric.
- Enterprise
- Quality assurance
Research
Decoupling Code Complexity from Newcomer Participation: A Causal Study of AI Coding Agent Adoption in OSS
Weiwei Xu, Xuanning Cui, Hengzhi Ye et al.
arXiv · 2026-07-02
This causal study examines whether AI coding agents (e.g., Cursor, Claude Code) crowd out newcomers in open-source software projects. Using difference-in-differences analysis across 1,888 GitHub projects that adopted an agent, the researchers find no significant decline in newcomer inflow, onboarding, or retention after adoption. While adoption does modestly increase code complexity (roughly +11% on a cognitive metric for Python and +3–4% on cyclomatic metrics across languages), this complexity rise does not translate into reduced newcomer participation. The findings suggest that the feared trade-off between AI coding assistance and human newcomer involvement does not materialize in established open-source projects.
- Workforce
- Enterprise
Research
Safety Testing LLM Agents at Scale: From Risk Discovery to Evidence-Grounded Verification
Yunhao Feng, Ruixiao Lin, Ming Wen et al.
arXiv · 2026-07-02
This paper introduces Vera, an automated end-to-end safety testing framework for large language model (LLM) agents that operate autonomously through external tools. Vera uses a three-stage pipeline: it discovers emerging safety risks from literature, composes executable safety test cases from taxonomies of risks and attack methods, and runs agents in isolated sandboxes where outcomes are verified from observable environment state rather than model self-report. Evaluated on four production agent frameworks, Vera found average attack success rates of 93.9% under multi-channel attacks, exposing substantial safety weaknesses. The authors also release Vera-Bench, a benchmark of 1,600 executable safety cases across 124 risk categories, arguing that modular, scalable testing infrastructure is essential for keeping pace with rapidly evolving agentic AI systems.
- Quality assurance
- Certifications
Research
Meta-Benchmarks for Financial-Services LLM Evaluation
Blair Hudson
arXiv · 2026-07-02
This paper presents a meta-benchmarking framework designed to evaluate large language models (LLMs) specifically for financial-services work, rather than relying on general-purpose leaderboards. It organises 452 publicly reported benchmarks into 41 O*NET Generalized Work Activities and 38 BIAN banking business domains—covering sales, operations, risk, and support—then applies a multiplicative weighting scheme (discrimination × coverage × recency) to suppress saturated legacy benchmarks and reward those still differentiating top models. A pairwise Elo tournament scaled by these weights produces cross-benchmark-comparable scores at both the work-activity and business-domain level. The framework is demonstrated on a snapshot of 288 models from 25 organisations as of June 2026, and is presented with full methodology to support reproducibility for institutions facing model selection and governance challenges.
- Enterprise
- AI policy
Research
Beyond Pixel Diffs: Benchmarking Image Change Captioning for Web UI Visual Regression Testing
Licheng Zhang, Bach Le, Pengtao Zhao et al.
arXiv · 2026-07-02
This paper introduces WUICC (Web UI Image Change Captioning), a new task and benchmark dataset for automatically generating natural-language descriptions of visual changes in user interface screenshots during visual regression testing (VRT). The authors evaluate eleven image difference captioning methods and two zero-shot large language models, finding that current methods struggle with web UI-specific challenges like layout diversity, dense text, and fine-grained changes, but already outperform pixel-level comparison by more selectively suppressing non-meaningful visual noise. The work addresses a gap in quality assurance tooling where developers must manually review large volumes of false positives flagged by semantically blind pixel-diff approaches. By enabling text descriptions of what changed rather than just flagging a difference, the benchmark provides a foundation for reducing manual review burden in software release pipelines.
- Quality assurance
Research
AI Writes Faster Than Humans Can Review: A Longitudinal Study of an Enterprise 2x Mandate
H. He, Shyam Agarwal, Yegor Denisov-Blanch et al.
arXiv (Cornell University) · 2026-07-02
A longitudinal study of 802 developers and 196,212 pull requests at a mid-sized enterprise found that a company-wide AI coding mandate helped per-capita pull request throughput reach 2.09x the pre-mandate baseline by April 2026, among the largest productivity gains reported from a real-world AI coding deployment. A difference-in-differences design links this gain to AI adoption and accumulated usage, with the mandate acting as a catalyst rather than a direct driver. Critically, the surge in code output restructured review workflows: per-reviewer load roughly doubled and automated review overtook human review, raising concerns about whether humans can keep pace with AI-generated code. Merge and revert rates held steady, but the study cannot fully separate quality implications across model generations or causally attribute gains due to non-random adoption.
- Enterprise
- Workforce
- Quality assurance
Research
AI adoption and perceived organizational performance in Chinese pharmaceutical sales: Efficiency and motivation pathways
Yunlu Cai, Siti Rohaida Mohamed Zainal
Problems and Perspectives in Management · 2026-07-02
This study of 335 pharmaceutical sales professionals in China finds that AI adoption is positively associated with perceived organizational performance, both directly and through two parallel pathways: operational efficiency and employee motivation. Using PLS-SEM analysis, the results show that AI adoption significantly predicts operational efficiency and employee motivation, which in turn predict performance, confirming partial mediation through both pathways. The findings suggest that AI generates performance value in compliance-sensitive sales environments when integrated into workflows alongside motivational support for employees.
- Workforce
- Enterprise
- Quality assurance
Research
Taxing Artificial Intelligence
Juliette Faivre, Sarah H. Cen
arXiv (Cornell University) · 2026-07-02
This paper examines whether taxation can serve as an effective policy tool to address harms associated with AI development and deployment, including environmental pressures, labor and creative displacement, and systemic risks. The authors survey a range of potential tax instruments—corporate income taxes, consumption taxes on AI services, and excise taxes tied to specific AI activities—and evaluate their feasibility, measurement challenges, incidence, leakage effects, and innovation costs. The paper argues that AI taxation should go beyond simple Pigouvian correction to also redistribute unevenly borne costs and gains and fund regulatory capacity, but cautions that policy design must be carefully matched to specific harms and objectives.
- AI policy
- Workforce
- Enterprise
Research
<b>BRIDGING THE DIGITAL TRANSFORMATION GAP IN NIGERIA'S CONSTRUCTION INDUSTRY: CHALLENGES FROM AI ADOPTION</b>
Henry Asuquo Okpo, Numbere Jamaibi Thanks
VUNOKLANG MULTIDISCIPLINARY JOURNAL OF SCIENCE AND TECHNOLOGY EDUCATION · 2026-07-02
This systematic review of 20 peer-reviewed studies examines why Nigeria's construction industry has failed to adopt AI and machine learning tools despite their global impact. Using the Technology-Organisation-Environment framework and a PRISMA-compliant methodology, the authors identify five barrier clusters: AI-specific data and infrastructure constraints, human capital deficits, financial limitations, regulatory and governance voids, and organisational inertia. The paper proposes a multi-level intervention framework and offers evidence-based recommendations for policy reform and industry strategy in Nigeria's built environment, with broader relevance for AI adoption in emerging economies.
- Workforce
- Enterprise
- AI policy
Research
The Impact of Artificial Intelligence Application on Corporate ESG Performance: Evidence from Chinese A-Share Listed Firms
Haixia Feng, Renbo Shi, Qingjin Wang
Systems · 2026-07-02
Using panel data from Chinese A-share listed firms spanning 2009 to 2025, this study finds that AI adoption significantly improves corporate ESG performance, with firms showing higher AI usage achieving better environmental, social, and governance outcomes. The positive effect is strongest among firms with lower technological intensity, non-heavily polluting industries, and regions with stricter environmental regulation. Mediation analysis identifies human capital upgrading and green technological innovation as key mechanisms through which AI drives ESG gains. These findings have direct implications for corporate sustainability strategy and the role of AI-enabled workforce and innovation capabilities in meeting ESG goals.
- Enterprise
- Workforce
- AI policy
Research
A question of style? Regulating artificial intelligence in the European Union and the USA
Alison Harcourt, Claudio M. Radaelli, Philipp Trein
Big Data & Society · 2026-07-02
This comparative policy study examines how the EU and USA regulate artificial intelligence, finding that the EU's 2024 AI Act offers a comprehensive but potentially rigid framework while the US takes a piecemeal, sector-specific approach that is more enforceable in practice. The authors argue that the EU's top-down architecture makes adaptive regulation difficult and raises concerns about enforceability and human rights protection, whereas the US state-by-state and sector-by-sector model allows for greater learning, imitation, and diffusion through 'laboratory federalism.' The paper uses the concepts of policy style and regulatory style to explain these divergent governance approaches and their trade-offs for AI oversight.
- AI policy
Research
Assured Automated Peer Review: Proof-Carrying Papers and the Trust Architecture AI Review Requires
Robert Tilley Jr.
Zenodo (CERN European Organization for Nuclear Research) · 2026-07-02
This paper proposes a trust architecture for automated AI-assisted peer review systems, addressing assurance problems such as adversarial submissions, undisclosed tool use, and incomplete accountability. The architecture centers on five controls: verification contracts, layered provenance controls, measurement of reviewer dependence, anytime-valid sequential monitoring, and signed decision receipts. The work is motivated by real deployments like Google's Paper Assistant Tool and the AAAI-26 AI Review Pilot, and aims to provide structured accountability without overclaiming unrestricted verification. The contribution is an architectural and evaluation framework rather than an empirical validation of a deployed system.
- Quality assurance
- Certifications
- AI policy
Research
Entrepreneurial orientation, dynamic capabilities, and startup performance: The amplifying role of AI
Nguyen Kien Quoc, Nguyen Ngoc Long
Problems and Perspectives in Management · 2026-07-02
This study of 315 Vietnamese startup founders and executives finds that entrepreneurial orientation and dynamic capabilities both drive startup performance, largely through the mediating role of innovation capability. AI adoption intensity significantly amplifies all three relationships—between entrepreneurial orientation, innovation capability, and dynamic capabilities and performance—with the strongest amplifying effect observed for dynamic capabilities. The model explains 66.5% of variance in startup performance, highlighting AI as a critical enabler of capability-conversion in resource-constrained, emerging-economy contexts. The findings offer practical guidance for startup managers on building competitive advantage through combined capability development and AI adoption.
- Enterprise
- Workforce
Research
The Eticas AI Risk Taxonomy: Open Infrastructure for Operationalizing AI Audits
Gemma Galdon Clavell, Pablo Accuosto, Usman Gohar
arXiv (Cornell University) · 2026-07-02
This paper presents the Eticas AI Risk Taxonomy v2.0.0, an open framework designed to bridge the gap between identifying AI risks and actually operationalizing them into structured audits with measurable, graded findings. The authors demonstrate the system end-to-end using PII leakage as a test case on GPT-4-0314, showing disclosure rates of 0%, 51%, and 84% under increasing adversarial conditioning, ultimately yielding a severity grade of E with a SYSTEMIC pattern. The taxonomy organizes 76 active subcategories across 10 categories and 20 sub-groups, with mappings to 18 external frameworks, and its upper layers are published under CC BY 4.0 as open semantic infrastructure. This work matters because it provides shared, standardized infrastructure for AI auditing that goes beyond risk catalogs to enable defensible, reproducible audit findings in high-stakes deployment contexts.
- Quality assurance
- Certifications
- AI policy
- Enterprise
Research
Artificial Intelligence and Inequality: Policy Paths in a Polarized Future
Hassan Daliri
Perspectives on Global Development and Technology · 2026-07-02
Using an agent-based simulation model grounded in theories of skill-biased technological change and labor market polarization, this study examines how AI and automation affect income inequality under different policy regimes. Results show that rapid automation without skills investment worsens inequality and displaces workers, while combining targeted skill subsidies with progressive taxation significantly reduces inequality and shifts employment toward AI-complementary sectors. The findings demonstrate that AI's inequality effects are policy-contingent rather than technologically predetermined, highlighting the importance of redistribution and adaptive human capital strategies for inclusive growth.
- Workforce
- AI policy
Research
The Agentic Garden of Forking Paths
Jiacheng Miao, Jonathan K Pritchard, James Zou
arXiv · 2026-07-01
This paper investigates how AI agents, when assigned different analytical personas, reproduce the ideological variation seen among human researchers analyzing the same dataset. Across four high-stakes domains, AI agents with different personas produced divergent and often opposing conclusions from identical data, replicating 72% of the human ideological gap observed in a study where 42 human teams analyzed the same immigration dataset. Critically, 86% of the AI-generated analyses passed independent AI review and 78% passed majority human expert review, suggesting the problem is not flawed analysis but selective exploration of a large space of defensible analytical choices. To address this, the authors introduce the 'm-value' (multiverse value) and 'Agentic Bootstrap,' a method that uses AI agents to sample plausible analysis paths and assess how extreme any single reported finding is relative to the full distribution of defensible analyses — offering a new criterion for evaluating scientific credibility.
- Quality assurance
- AI policy
Research
Grounded Optimization: A Layered Engineering Framework for Reducing LLM Hallucination in Automated Personal Document Rewriting
Shashank Indukuri, Adarsh Agrawal
arXiv · 2026-07-01
This paper introduces Grounded Optimization, a five-layer engineering framework designed to reduce hallucinations when large language models rewrite resumes for applicant tracking systems. The framework combines temporal context validation, contamination detection, structural enforcement, prompt-level grounding, and an evaluator agent. Ablation experiments across three LLMs, four temperature settings, and six layer configurations on 25 synthetic resumes spanning 14 industries show that undefended baselines produce 2.48–5.36 detected hallucinations per resume, while the full framework reduces the overall detected hallucination rate to 0.04–0.24 and cuts temporal hallucinations by 50–95%. The work matters because it demonstrates that deterministic, layered defenses are necessary complements to prompt-level grounding, especially for weaker models or higher temperature settings, improving the reliability of AI-assisted document rewriting in high-stakes job application contexts.
- Quality assurance
- Workforce
Research
On the Utility and Factual Reliability of Pruned Mixture-of-Experts Models in the Biomedical Domain
Atsuki Yamaguchi, Szymon Palucha, Léo Bijar et al.
arXiv · 2026-07-01
This paper examines how structured expert pruning of Mixture-of-Experts (MoE) language models affects both utility and factual reliability in the biomedical domain. Testing four MoE models across six pruning methods and multiple pruning ratios, the authors find that moderate pruning can preserve in-domain biomedical utility without an immediate drop in reliability, but extreme pruning raises hallucination risks. Performance degrades rapidly when pruned models are applied outside their target domain. The study concludes that evaluating pruned models on utility benchmarks alone is insufficient for safe deployment in high-stakes settings like biomedicine, where factual reliability must also be assessed.
- Quality assurance
Research
When Should Service Agents Reconsider? Difficulty-Routed Control in Customer-Service Operations
Qian Chen, Chengyuan Liu, Xin Yu
arXiv · 2026-07-01
This paper addresses how autonomous AI customer-service agents—which now execute backend operations like refunds, cancellations, and order modifications—can balance speed for routine requests with safeguards for complex, error-prone ones. The authors propose a 'difficulty-routed' architecture that uses a lightweight router to send routine sessions down a fast baseline path while escalating operationally complex sessions to a more deliberative workflow featuring conflict-aware communication and write-triggered reconsideration. Evaluated on human-verified retail and airline tasks from the τ²-bench benchmark, the method improves reliability on conflicted service requests without broadly expanding interactions across all sessions. The findings suggest targeted control—concentrating deliberation before consequential backend writes rather than applying safeguards uniformly—is an effective design principle for enterprise AI service agents.
- Enterprise
- Quality assurance
Research
Adoption and Impact of Command-Line AI Coding Agents: A Study of Microsoft's Early 2026 Rollout of Claude Code and GitHub Copilot CLI
Emerson Murphy-Hill, Jenna Butler, Alexandra Savelieva
arXiv · 2026-07-01
This study examines Microsoft's early-2026 rollout of Claude Code and GitHub Copilot CLI across tens of thousands of engineers, finding that adoption spread primarily through social networks rather than demographic factors, and that retention was more closely tied to engineers' existing coding activity levels. Engineers who adopted these command-line AI coding agents merged roughly 24% more pull requests than they otherwise would have, a lift that persisted across the four-month observation window. The findings indicate that CLI coding agents produce measurable productivity gains rather than mere novelty effects, and that organizations should prioritize visible peer use as a central strategy when rolling out agentic tools at scale.
- Workforce
- Enterprise