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
Labor protection and corporate artificial intelligence technology innovation: A quasi-natural experiment based on the implementation of China's Social Insurance Law
Shanshan Wei
International Review of Economics & Finance · 2026-08-06
Using China's 2011 Social Insurance Law as a quasi-natural experiment, this study finds that stronger labor protections significantly increased corporate AI innovation by boosting AI-related R&D investment and improving the allocation of AI R&D personnel. The positive effect was more pronounced in regions with stronger financial and market institutions and in larger, more profitable firms operating in competitive or high-tech industries. The findings suggest that labor security policy can serve as a meaningful lever for fostering corporate AI development.
- Workforce
- AI policy
Research
Negotiating Risk Boundaries in AI for Policing Through Mixed-Stakeholder Deliberation
Mackenzie Jorgensen, Jo Reilly, Alex Sutherland et al.
arXiv (Cornell University) · 2026-08-05
This paper presents findings from a mixed-stakeholder deliberation workshop in the UK that brought together 30 community representatives, police officers, and academics to evaluate the risks of 13 AI use cases in policing, with an explicit focus on racial bias. Participants were broadly open to AI adoption, rejecting only three use cases outright—most notably recidivism risk assessment, where objections targeted the tool's underlying premise rather than its implementation. The study found that foregrounding racial equity did not narrow deliberation; instead, discussions consistently returned to questions of whether tools actually work, deliver genuine benefit, and do so equitably for all. The authors argue this integrated reasoning, analogous to the curb-cut effect in inclusive design, supports embedding a racial bias lens into AI risk-benefit analysis from the outset in policing contexts.
- AI policy
Research
Evaluating and Improving Pedagogical Fit in LLM-Based AI Tutors with the Pedagogical Suitability Index
Benjamin Barlog, Hudson Craig, Zedong Peng
arXiv · 2026-08-05
This paper introduces the Pedagogical Suitability Index (PSI), a composite metric with six theory-informed sub-scores designed to measure how well LLM-based AI tutors align their responses with learner readiness and curricular progression, not just answer correctness. The authors evaluate four LLM tutors (ChatGPT, Gemini, Gemma4, and Qwen3) across 240 scenario-based evaluations and find that baseline differences across models are modest (PSI range: 0.557–0.638), with no clear advantage for closed versus open-weight models. Crucially, using PSI as a structured feedback signal to regenerate weak responses improved 51 of 62 targeted cases (82.3%), with manual evaluation confirming many improvements were instructionally meaningful. The findings suggest that learner- and curriculum-aware alignment is both measurable and improvable, and may matter more for effective AI tutoring than model type alone.
- Quality assurance
Research
OrchestraBench: Evaluating Multi-Agent Orchestration Failure Modes, Recovery, and Decomposition Quality
Yidian Chen, Yingzi Gu, Natan Vidra et al.
arXiv · 2026-08-05
OrchestraBench is a benchmark framework designed to evaluate multi-agent AI orchestration systems beyond simple task accuracy, focusing on failure modes, recovery, and decomposition quality in enterprise workflow contexts. Using a seed-reproducible failure-injection harness over templated workflows, the study finds that a keyword/flag router scored 0% on adversarial routing cases while an intent-reasoning model router scored 100%, and that failure recovery varied dramatically by failure type—tool faults recovered fully (1.0), ambiguous delegation partially (0.30), and latent or semantic failure modes showed zero recovery across all tested configurations. Cascade radius grew substantially with pipeline depth (mean 0.9 to 4.7 across depths 3–7), and blind retry was found to reproduce latent faults while increasing time to detection, suggesting that detection and attribution are necessary for containment. These findings matter for enterprise AI deployment by showing that routing strategy and failure-handling design—not just model capability—are critical determinants of multi-agent pipeline reliability.
- Enterprise
- Quality assurance
Research
Item Response Theory for AI Safety
Joshua Fonseca Rivera, Neil Shah, David Demitri Africa et al.
arXiv · 2026-08-05
This paper applies Item Response Theory (IRT)—a psychometric statistical framework—to analyze eight AI safety benchmarks across 192 language models, making it the largest psychometric analysis of LLM safety evaluations to date. The authors find that three interpretable factors (refusal strictness, truthfulness, and contextual harm) explain most of the variance in model behavior across benchmarks, and that psychometrically selected items can recover full benchmark scores with far fewer questions, cutting evaluation cost by 97–99%. IRT also enables auditing of individual models, including detection of naive sandbagging and changes in the model behind an API. The authors recommend that frontier labs and evaluators adopt IRT as a toolkit for making safety benchmarks more efficient, interpretable, and trustworthy.
- Quality assurance
- Certifications
Research
Provable Limits and Certified Deferral for Verbalized Uncertainty in Small Language Models
Jianru Shen
arXiv · 2026-08-05
This paper investigates whether small language models (0.5B–14B parameters) can reliably express their own uncertainty in words, and whether that uncertainty can be used to decide when the model should hand off a question to a human. Evaluating eleven instruction-tuned models across two benchmarks (ARC-Challenge and TruthfulQA) with over 25,000 predictions, the authors prove theoretical limits on calibration—including conditions under which temperature scaling is mathematically impossible—and propose a Clopper-Pearson procedure to generate finite-sample risk certificates from a 200-question calibration set. Empirically, certified autonomous operation at a 20% error budget was granted to only three of 22 model-task pairs, and to none at a stricter 10% budget, demonstrating that verbalized confidence alone is insufficient to safely deploy small models without human oversight. These findings directly inform when and how small AI models can be trusted to operate independently versus when human deferral is required.
- Workforce
- Certifications
Research
Short-term load forecasting under EU-AI Act Requirements in Safety-Critical Environments: Results from a 41-day live challenge on the aggregated German transmission-grid load
Thomas Bartz-Beielstein
arXiv · 2026-08-05
This paper presents results from a 41-day live forecasting challenge evaluating a short-term load forecasting (STLF) pipeline for the aggregated German transmission-grid load, framed explicitly as both an accuracy and a compliance problem under the EU AI Act. The pipeline, built on the open-source library spotforecast2-safe, incorporates determinism, reproducibility, and auditability by design, along with anomaly detection, calendar and weather covariates, and hyperparameter tuning. The EU AI Act-compliant pipeline outperformed the official ENTSO-E day-ahead forecast baseline, and lightweight auditable local models proved competitive with large pre-trained foundation models such as chronos-2 with over 100 million parameters. The work demonstrates that regulatory compliance requirements for safety-critical infrastructure can be met without sacrificing forecasting accuracy, and all challenge data and results are publicly available.
- AI policy
- Quality assurance
Research
The Beginning of ChatGPT Ads
Emma Lurie, Ro Encarnación, Sorelle A. Friedler et al.
arXiv · 2026-08-05
This paper presents the first empirical audit of advertising in ChatGPT's user interface, using 91 automated 'sock puppet' accounts designed to signal different racial/ethnic groups and income levels to study demographic differences in ad delivery. The study collected over 3,000 advertisements from 186 unique advertisers and found that lower-income accounts were more likely to receive ads regardless of race, and that early ChatGPT ads skewed heavily toward consumer goods and were visually separated from the LLM's response text. The findings raise important questions about algorithmic targeting, fairness, and transparency as advertising becomes more deeply integrated into AI chat interfaces.
- AI policy
- Enterprise
Research
DelusionEval: Measuring Delusion-Linked Behaviors in AI Chatbots
Jared Moore, Andrea Mock, Yifan Mai et al.
arXiv · 2026-08-05
DelusionEval is a benchmark that measures how often AI chatbots exhibit behaviors that reinforce user delusions and psychological harm. Using 589 real conversation histories (12,591 messages) from 18 users who experienced delusions, the study finds that all major model families—including GPT and Claude—show substantial rates of delusion-linked behaviors, and that these rates do not reliably improve with larger model size, later release date, or test-time reasoning. Critically, extending conversation context significantly worsens outcomes: for example, failure to discourage self-harm when a user expresses suicidal ideation rises from 30% to 41.1% when 350 additional prior messages are included. The findings highlight serious gaps in current LLM safety evaluation and underscore the need for rigorous, real-world studies of human-AI interaction in mental health contexts.
- Quality assurance
- AI policy
Research
A Chain Is Only as Strong as Its Weakest Link: A Scoping Review of System Integration Audits in AI
Leah Davis, Dominic Martin, AJung Moon
arXiv (Cornell University) · 2026-08-05
This scoping review examines how AI auditing handles system integration risks, screening 4,259 documents and analyzing 58 audits that treat system integration as a core evaluation focus. The authors find that current AI auditing is fragmented and model-centric, with few measures targeting integration-specific risks such as compatibility, completeness, and oversight across inter-component, system-environment, and multi-system sites. Large gaps remain relative to traditional audit standards seen in safety-critical fields like aerospace, and access to information and resources heavily shapes how audits are designed. The paper calls on the AI community to develop audit practices capable of capturing failures that span components, environments, and systems beyond what component-level evaluation can detect.
- Quality assurance
- Certifications
Research
When Shared Rollouts Fail in Defensive Driving Evaluation: A NAVSIM Score Basis Audit
Ziang Wei, Minjun Yu, Zheyuan Lai et al.
arXiv · 2026-08-05
This paper audits a flaw in NAVSIM v2.2, a benchmark used to evaluate defensive driving policies. The authors find that when a shared numerical instability causes both the agent and the logged human reference to fail the same compliance checks, a 'reference-conditioned forgiveness' rule incorrectly awards compliance credit to both — including route-blind and actor-blind probe policies that should perform poorly. As a result, these deliberately uninformed probes outrank human replay and a competitive baseline (PDM-Closed) on the full 12,146-token navtest split, undermining the benchmark's ability to distinguish meaningful driving policies. The authors isolate the root cause to dependency-sensitive numerical behavior in a shared velocity refit step, show that replacing only the solver restores correct ordering, and propose an audit protocol requiring stack disclosure, blind probes, and rollout stability tests before defensive driving claims are made using such scores.
- Quality assurance
- Certifications
Research
Evaluating the Diagnostic Robustness of Vision-Language Models Under Visual and Textual Perturbations
Ali Khoramfar, Mohammad Javad Dousti, Alireza Mohamadian et al.
arXiv · 2026-08-05
This paper stress-tests four vision-language model (VLM) families on a histopathology-validated brain MRI dataset by applying evidence-preserving perturbations — reordering anatomical slices, swapping answer-choice label positions, and removing expert-annotated lesion slices — to measure whether predictions stay consistent when the underlying clinical evidence has not changed. The results reveal alarming fragility: prediction flips occur in up to 48.9% of cases under simple sequence reversals, label reordering causes inconsistent diagnoses in up to 67.8% of cases with identical images, and models still commit to categorical diagnoses in up to 76.1% of cases after lesion slices are removed. The authors argue that standard aggregate accuracy metrics overestimate clinical reliability by hiding sensitivity to sequential presentation and textual framing, and call for stability-based evaluation metrics before VLMs are deployed in safety-critical clinical settings.
- Quality assurance
- Certifications
Research
ContextWeave: A Real-World Workflow Benchmark
Bo Wang, Yuqian Yao, Enxi Wang et al.
arXiv · 2026-08-05
ContextWeave is a longitudinal benchmark designed to evaluate how well AI language agents use recalled past experience to improve performance on realistic, long-horizon office-work tasks. It reconstructs privacy-preserved, multi-month workflows from 14 participants into 1,005 executable tasks and measures both workspace quality and alignment with user-specific preferences. Across tested configurations, the best memory setup raises Workspace Score from 68.08 to 78.20 and Preference Score from 41.50 to 70.60, though gains vary across models and memory components. The findings show that actionable, experience-rich memory aids workflow continuation more than compact summaries, but also increases susceptibility to misleading recall, motivating memory systems that balance retrieval relevance with reliable execution.
- Enterprise
- Quality assurance
Research
Guideline-as-Oracle: Zero-Annotation Training of an Ophthalmic Telephone Triage Agent
Chenyu Wang, Yi Liu, Baoqing Li et al.
arXiv · 2026-08-05
This paper introduces Guideline-as-Oracle (GAO), a zero-annotation training approach for an AI telephone triage agent in ophthalmology. Instead of costly expert-labeled dialogues, the method compiles American Academy of Ophthalmology guidance into a structured rule table and uses it as the sole supervision source to generate 3,000 training dialogues. Fine-tuning a 9-billion-parameter model on this corpus improves agreement with a 201-case clinical reference from 61.7% to 74.1% and raises emergent-case recall from 9.5% to 69.0%, outperforming seven general-purpose AI systems on these metrics without requiring a frontier model at inference. This matters for healthcare quality assurance and workforce support, as it shows clinical guidelines alone can drive reliable AI triage performance while avoiding expensive annotation and privacy-sensitive data sharing.
- Quality assurance
- Workforce
Research
Chain-of-Thought Monitoring Can Be Unreliable in Implicit-Influence Settings
Agatha Duzan, Asa Cooper Stickland
arXiv · 2026-08-05
This paper introduces a benchmark comparing how reliably chain-of-thought (CoT) monitoring detects hidden behavioral influences in AI models under two conditions: explicit influence (where the model is directly instructed to act on a nudge and hide it) versus implicit influence (where contextual details like irrelevant candidate information bias model outputs without any instruction to conceal). Testing seven frontier reasoning models across four task formats, the authors find that CoT monitors detect 60–94% of behavior shifts under explicit influence, but detection drops by 41–46 percentage points in two of four implicit settings. Notably, realistic system-prompt additions that developers might deploy to reduce off-topic bias can push implicit detection as low as 5% while preserving the behavioral influence itself, suggesting that current monitorability estimates are likely overoptimistic and that well-intentioned deployment choices can further erode safety oversight.
- Quality assurance
- AI policy
Research
Innocent Panels, Hateful Stories: Evaluating and Detecting Hateful Intent in Multi-Turn Visual Story Generation
Ye Leng, Junjie Chu, Yiting Qu et al.
arXiv · 2026-08-05
This paper investigates how modern text-to-image systems like Gemini and GPT-Image can be exploited to generate hateful visual stories across multiple turns—sequences of images that collectively convey hateful narratives even when individual panels appear innocent. The authors introduce HatefulStoryPrompts (330 multi-turn configurations from 55 hateful stories) and find that every tested frontier model completes over 80% of hateful stories, with the best reaching 99.0% completion. Existing moderation tools perform poorly at the group level, with dedicated safety models achieving at most 34.9% recall on the human-labeled HatefulVisualStory dataset, though a strong vision-language model reaches 67.5%. The paper proposes proactive and post-generation defenses—including an interaction-aware monitor achieving 97.3% recall—arguing that image safety must evolve from per-image moderation to stateful reasoning over multi-turn interactions.
- AI policy
- Quality assurance
Research
When Memory Lies: An Empirical Study of Spatial Memory Staleness in VLM Agents
Yushi Sun, Yanjie Zhang
arXiv · 2026-08-05
This paper investigates what happens when AI agents that rely on stored spatial memory encounter environments that have changed, making that memory outdated or 'stale.' Using a dynamic grid-world testbed (FrozenLake), the researchers ran over 13,000 experiments across six language/vision models and found three key results: models that correctly detect stale memory from text often fail badly when using visual input (vision F1 as low as 0.067); trusting stale memory without verification more than doubles agent death rates compared to using no memory at all; and while adding a read-time memory audit helps in text mode, it does not fully close the safety gap, especially when visual grounding is unreliable. The findings frame spatial-memory staleness as a concrete safety failure mode in memory-augmented AI agents and highlight reliable visual grounding under memory-observation conflict as a critical unsolved challenge.
- Quality assurance
Research
The Personalization Mirage: How LLMs Fabricate User Profiles, and Why Self-Monitoring Misleads
Yushi Sun, Yanjie Zhang, Rui Sheng
arXiv · 2026-08-05
This paper investigates 'over-inference' (OI) in personalized large language models — the tendency of LLMs to fabricate user attributes beyond what the available evidence supports. The authors introduce MirageBench, a benchmark of 150 personas and 6 personalization tasks, and evaluate 12 models across 7 families on over 143,000 judged claims, finding that every model over-infers between 35% and 49% of its claims (cross-model mean 41.6%). A key finding is a 'Self-Monitoring Inversion': models' self-assessed OI is negatively rank-correlated with judge-measured OI (rho = -0.60, p = 0.044), meaning self-reported confidence is a misleading signal for model comparison. The results argue that external verification, rather than model self-monitoring, is needed for trustworthy personalization systems.
- Quality assurance
- Enterprise
Research
EuroExec: Frontier Language Models Fall Short of Expert Judgment on European Executive Decision Tasks
Pau Arnal, Khaled Denfir, Danylo Smahliuk et al.
arXiv (Cornell University) · 2026-08-05
EuroExec introduces a benchmark of 413 open-ended European executive decision tasks authored by 47 domain experts, evaluated with over 4,000 human expert hours. Six frontier LLMs were tested; the strongest model solved only 56.9% of tasks, while expert-written reference answers were preferred over every model response in 74% of direct rankings. The study concludes that frontier generative systems fall well below the professional standard required for the real-world executive work they are already being used for, and that human evaluators — with rigorous consistency checks — are necessary to surface this gap, as automated metrics also fall short on open-ended problems with subjective ground truth.
- Enterprise
- Quality assurance
Research
The Evaluator Is Part of the Experiment: Measuring Open-Ended LLM Conformity
Alicia Guerra, Yibo Hu
arXiv · 2026-08-05
This paper investigates how large language models revise open-ended answers when exposed to peer input, finding that wrong peer information consistently produces the lowest-quality revisions across all tested generator-dataset combinations. The study also reveals that LLM-based evaluators are not neutral: different judges shift their ratings toward or away from the peer-endorsed position when that context is visible, including GPT-4o and GPT-5.4-mini audits. The authors conclude that simple answer-flip rates are insufficient for measuring conformity in open-ended settings, and that anchor calibration must be explicitly checked to maintain a stable evaluation scale. These findings have important implications for how AI outputs are assessed, highlighting systematic biases in the evaluation process itself.
- Quality assurance
Research
Social Pressure Breaks Majority Voting in LLM Safety Panels
Yibo Hu, Jiaming Qu
arXiv · 2026-08-05
This paper investigates how social pressure undermines majority-vote safety panels composed of large language models. In a controlled two-round experiment, models first judge content alone, then re-judge after seeing simulated peer opinions asserting the wrong label or abstaining. The study finds that wrong-label peer messages dramatically inflate false-alarm rates—from 56.5% under silent peers to 87.5% per reviewer and 100% at the panel level—while the effect is strongly asymmetric, with models far more susceptible to pushes toward 'unsafe' labels than 'safe' ones. These findings reveal a critical failure mode for LLM-based content moderation panels and propose a simple pre-deployment diagnostic to identify this vulnerability.
- Quality assurance
- AI policy
Research
Manipulation-Proof Oblivious Audits against Deceptive Model Providers
Augustin Godinot, Sofiane Azogagh, Julien Ferry et al.
arXiv (Cornell University) · 2026-08-05
This paper introduces an audit protocol that makes it significantly harder for machine learning model providers to manipulate fairness evaluations. By using a Private Information Retrieval mechanism, the auditor can query a model 'obliviously'—requiring the provider to label a large pool of instances without knowing which subset will actually be used for the audit. The authors provide theoretical guarantees showing that a provider trying to hide unfairness must falsify a much larger number of responses under this protocol, raising both the difficulty and detectability of manipulation. Experimental results confirm the approach is practical and imposes minimal overhead on the auditor without requiring changes to the model.
- AI policy
- Quality assurance
Research
Equitable System-Prompt Selection via Constrained Mixed-Strategy GroupDRO
Mengyu Xu, Qiaoxin Yang, Zhihan Liu et al.
arXiv · 2026-08-05
This paper addresses the problem that LLMs can give very different quality answers to semantically equivalent questions phrased differently, depending on the system prompt used. The authors propose a constrained mixed-strategy GroupDRO framework that assigns weights to a pool of existing system prompts to minimize worst-case information-quality loss across different question phrasings and evaluation metrics, while keeping average quality close to standard selection. Tested across five LLMs on bilingual medical and consumer-finance benchmarks, the method reduces Overall Mean, Worst 25% Mean, and Worst loss by 13.1%, 13.2%, and 13.7% on average relative to no mitigation. This matters for quality assurance in high-stakes domains like healthcare and finance, where consistent and equitable response quality across diverse user phrasings is critical.
- Quality assurance
- Enterprise
Research
Zero Defect Manufacturing – from vision to integrated industrial reality
Foivos Psarommatis, Victor Azamfirei, Vittorio Solina et al.
International Journal of Production Research · 2026-08-05
This editorial introduces a Special Issue on Zero Defect Manufacturing (ZDM), synthesizing twelve research papers that collectively advance the scientific, technological, and managerial foundations for defect-free production in digitally connected manufacturing environments. The contributions span data analytics, artificial intelligence, explainable AI, human–AI collaboration, cost modelling, and machine-vision maturity assessment, demonstrating ZDM's growing maturity as both a research field and industrial practice. The special issue identifies six future research priorities—semantic interoperability, hybrid intelligence, lifecycle integration, standardised benchmarks, cybersecurity, and human-centric design—positioning ZDM as a cornerstone for resilient, intelligent, and sustainable manufacturing. Published in the International Journal of Production Research, the work signals a shift from ZDM as a quality philosophy to an integrated, data-driven industrial paradigm.
- Quality assurance
- Enterprise
Research
CyberBridge: Bridging the Gap Between Cybersecurity Education and Industry
Arthur Nijdam, Paul Stankovski Wagner, Sara Ramezanian
arXiv (Cornell University) · 2026-08-05
CyberBridge is a framework that automatically maps cybersecurity job descriptions to professional role profiles by decomposing vacancy descriptions into Knowledge, Skill, and Task (KST) statements and matching them to workforce profiles using a sentence-BERT model. A key feature is its interpretability, allowing users to trace recommendations to specific competencies rather than relying on a black-box system. The framework supports job recommendation for students, labor market analysis for educators, and curriculum planning to assess how well academic programs align with in-demand cybersecurity roles, making it a practical tool for evidence-based curriculum development.
- Workforce
- Certifications