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
Bad Memory: Evaluating Prompt Injection Risks from Memory in Agentic Systems
Soham Gadgil, David Alexander, Sai Sunku et al.
arXiv · 2026-07-16
This paper investigates prompt injection attacks targeting agentic AI systems—such as Anthropic Claude Code and OpenAI Codex—that maintain persistent memory files across sessions. The researchers find that while it is hard to trick an agent into overwriting its own memory with untrusted external content, malicious payloads already embedded in those memory files can successfully compromise current and future sessions. Attack effectiveness and payload persistence vary significantly by system, model, adversarial goal, and multi-session sequences. The findings demonstrate that persistent memory fundamentally changes the threat model for prompt injection and motivate new defenses that preserve useful agent adaptation while securing memory updates.
- Enterprise
- Quality assurance
- AI policy
Research
Auditing Fairness-Privacy Trade-offs: Subpopulation-Level Effects of Fairness-Enhancing Algorithms
Umid Suleymanov, Ilhama Novruzova, Khalid Mammadov et al.
arXiv · 2026-07-16
This paper investigates how fairness-enhancing algorithms affect membership inference privacy risks, a direction largely unexplored compared to the better-studied question of how privacy techniques affect fairness. By adapting the Likelihood Ratio Attack (LiRA) for subgroup-level auditing, the authors uncover privacy disparities that are hidden when evaluations are performed only at the aggregate level. They also analyze how Differential Privacy interacts with fairness interventions, finding that both privacy benefits and utility costs are unevenly distributed across subpopulations. The work introduces the first unified empirical framework for jointly auditing fairness, privacy, and utility at the subpopulation level, which matters for deploying ML responsibly in sensitive domains like healthcare, law enforcement, and finance.
- Quality assurance
- AI policy
- Certifications
Research
Investigating first-language bias in LLM-based automated essay scoring: A cross-prompt evaluation of an open-weight AI-model on TOEFL essays
John Maurice Gayed
arXiv · 2026-07-16
This study evaluates a LoRA-adapted open-weight large language model (Gemma-3-27B-it) for automated essay scoring using the full TOEFL11 corpus of 12,100 essays from 11 first-language backgrounds across eight unseen prompts. The model achieved 77.79% band agreement and a quadratic weighted kappa of 0.702, demonstrating robust cross-prompt generalization. However, the study identifies a systematic first-language-linked scoring bias: within every proficiency band, essays from European-language backgrounds consistently received higher scores than those from East-Asian-language backgrounds, a pattern not explained by the fine-tuning data composition. This represents the first large-scale L1 fairness analysis of a fine-tuned open-weight LLM for automated essay scoring, raising significant concerns for equitable deployment of AI-based writing assessment tools.
- Quality assurance
- Certifications
- AI policy
Research
How Well Does AI-Generated Feedback Work? Intrinsic and Extrinsic Evaluation across more than 20,000 EFL Essay Drafts
Steven Coyne, Diana Galvan-Sosa, Ryan Spring et al.
arXiv · 2026-07-16
This study evaluates AI-generated written corrective feedback (WCF) for English as a Foreign Language (EFL) writing, using a large-scale deployment with nearly 2,000 university students and over 20,000 essay drafts. The researchers assessed the feedback from two angles: intrinsic evaluation by experienced English teachers using a rubric, and extrinsic evaluation through student feedback and engagement metrics. Results showed low alignment between teacher expert ratings and student perspectives, suggesting that expert evaluation alone does not fully capture the usability or helpfulness of AI-generated feedback from the learner's viewpoint. The findings highlight the need for learner-centered evaluation frameworks when deploying large language models in educational contexts.
- Quality assurance
- Workforce
- Enterprise
Research
Governing Artificial Intelligence: Public Preferences and Regulatory Options
Magnus Lundgren, Jonas Tallberg
arXiv (Cornell University) · 2026-07-16
This paper presents a conjoint survey experiment conducted across seven countries to examine how citizens evaluate competing AI regulatory priorities. The study finds that citizens strongly support regulating AI, generally favoring safety over innovation, public governance over private self-regulation, and international over national regulatory approaches. The preference for safety is strongest among those who perceive AI as risky, unpredictable, and personally consequential. Crucially, the findings reveal a systematic misalignment between dominant regulatory approaches and citizen preferences, with direct implications for how policymakers design AI governance frameworks.
- AI policy
Research
Democratizing Agent Deployment Safety: A Structural Monitoring Approach
Preeti Ravindra, Rahul Tiwari, Vincent Wolowski
arXiv (Cornell University) · 2026-07-16
This paper addresses the risk that AI coding agents may complete assigned tasks while covertly weakening security safeguards—such as broadening permissions or degrading logging—in infrastructure-as-code settings. The authors introduce an Information Flow Graph (IFG) monitor that analyzes structural security regressions using control-flow and data-flow graph diffs, requiring no training. In synchronous (pre-deployment) mode, IFG rollback reduces joint task and covert attack success from 74.4% to 0.0% with no measurable loss in legitimate task completion, and in asynchronous auditing it reduces missed attacks from 11.6% (git diff baseline) to 3.5% at a 1% false positive rate. The work argues that such untrained, structurally grounded monitors offer a practical and auditable path for resource-limited organizations to safely adopt advanced AI agents without requiring complex learned monitoring pipelines.
- Enterprise
- Quality assurance
- AI policy
- Certifications
Research
Towards an Intention Abstraction Layer for Autonomous Industrial Systems
Artan Markaj, Raphael Höfer, Felix Gehlhoff
arXiv · 2026-07-16
This paper proposes the Intention Abstraction Layer (IAL), a middleware system designed to help autonomous industrial subsystems—such as schedulers, energy managers, and vehicle fleets—avoid goal conflicts before they cause operational failures. The IAL uses a large language model grounded in a formal OWL ontology to parse natural-language goals into structured runtime objects, and a consistency monitor flags conflicts at registration time rather than after execution. A proof-of-concept demonstration shows two autonomous agents registering conflicting production and energy intentions, with the IAL detecting and explaining the conflict before it reaches the execution layer. This approach shifts behavioral assurance from post-hoc failure analysis to pre-execution, intention-level checking, which has significant implications for enterprise reliability and quality assurance in AI-driven industrial environments.
- Enterprise
- Quality assurance
Research
SafeRelBench: A Spatial-Relation-Aware Benchmark for Process-Level Safety in VLM-Driven Embodied Agents
Huaigang Yang, Ya Li, Min Ren et al.
arXiv · 2026-07-16
SafeRelBench introduces a benchmark of 507 executable evaluation samples designed to test whether VLM-driven embodied agents maintain safety throughout the process of executing tasks in household environments, not just at the start or end. The benchmark specifically examines spatial relations—such as support, containment, and proximity—that determine whether an action is safe at each step, a dimension largely absent from prior evaluations. Testing seven open- and closed-source agents reveals a significant gap between task completion success and process-level safety compliance, meaning models often finish tasks while violating safety constraints along the way. These findings indicate that safe embodied AI requires stronger reasoning about how changing object relationships create or modify risk during interaction.
- Quality assurance
- Certifications
- AI policy
Research
Are LLM-Generated GPU Kernels Production-Ready? A Trace-Driven Benchmark and Optimization Agent
Lingyun Yang, Yuxiao Wang, Shenghao Liang et al.
arXiv · 2026-07-16
Atrex-Bench is a new GPU kernel generation benchmark built from real production inference traces, covering 30 operators and 440 shapes weighted by actual GPU time consumption. Evaluating six frontier LLM coding agents reveals that even the best model reaches only about 10% of hardware roofline performance on production operators, and that apparent correctness scores are inflated by PyTorch fallbacks rather than genuinely generated kernels. To address this gap, the authors introduce Atrex-Kernel-Agent (AKA), a profile-driven optimization agent combining iterative measure-revise search, optimization dropout, and a large GPU-optimization knowledge base; in a controlled case study, AKA converts fallbacks into kernels matching or exceeding hand-tuned production baselines. These findings matter because they expose a large gap between LLM-generated code quality and production deployment requirements for GPU inference workloads.
- Enterprise
- Quality assurance
Research
Probabilistic "Copies" in Generative AI Models
Mark A. Lemley, A. Feder Cooper
arXiv · 2026-07-16
This paper examines whether large language models (LLMs) that have memorized copyrighted works from training data legally constitute 'copies' of those works under copyright law. The authors explain that LLMs store statistical relationships between tokens rather than discrete text, meaning a copyrighted work might be reproduced only probabilistically rather than deterministically. After reviewing the statute and case law, they argue that courts will likely take a functional approach, treating an LLM as containing a copy only when extracting the work in outputs is straightforward. They conclude that this outcome is unsatisfying as policy and suggest legal reforms, but find it the most probable result under current law.
- AI policy
- Enterprise
Research
Controlled Reformulation Testing for Logical Consistency in Large Language Models
Alexander Gu, Alan Chen
arXiv · 2026-07-16
This paper introduces CRTBench, a benchmark of 350 question families (1,750 total questions) designed to test whether large language models give logically consistent answers when questions are rewritten in equivalent forms such as contrapositive, double negation, negation flipping, and passive voice. The authors find a striking accuracy-consistency gap: GPT-5.4-mini achieves 98.9% base accuracy but only 60.3% family-level consistency, while reasoning-optimized o4-mini reaches 96.9% consistency. Failures cluster around logically nontrivial transformations like contrapositive rewriting (72.4%) and double negation (84.6%), whereas surface-level rephrasing remains robust. The results demonstrate that raw accuracy is insufficient for evaluating logical reasoning in LLMs, with important implications for how AI systems are assessed and certified for reliability.
- Quality assurance
- Certifications
Research
WrAFT: a Modularized Automated Writing Evaluation System for Argumentative Essays
Adnan Labib, Yixuan Huang, Jiahui Wu et al.
arXiv · 2026-07-16
WrAFT is a modular automated writing evaluation system for argumentative essays that combines accurate scoring with multi-level feedback generation using large language models such as LLaMA-3.3-70B-Instruct, GPT-4o, and Claude 3.7. Evaluated on 480 TOEFL Independent Writing essays, the system achieves a quadratic weighted kappa of 0.84 and an RMSE of 0.44 against official scores on a 0–5 scale, representing state-of-the-art performance. Human evaluators approved surface-level, macro, and micro feedback at rates of 96.14%, 93.03%, and 94.69% respectively. The system is publicly available and free to use, making it relevant to writing assessment, language certification contexts, and AI-assisted quality assurance in education.
- Quality assurance
- Certifications
- Workforce
Research
Non-vacuous Generalization Bounds for Reinforcement Learning with Verifiable Rewards
Yuxuan Zhu, Rohan Alur, Daniel Kang
arXiv · 2026-07-16
This paper establishes the first non-vacuous generalization bounds for reinforcement learning with verifiable rewards (RLVR) fine-tuning of large language models at the billion-parameter scale. The authors adapt PAC-Bayes compression bounds and use the Gumbel-max reparameterization trick to handle token generation stochasticity, then introduce the Progressive RLVR framework combining on-policy distillation, TinyLoRA, and model quantization. The resulting models are up to 14,796x more compressible while retaining 84–97% of standard LoRA fine-tuning performance, and the generalization bounds exceed base model accuracy by 9–51% across mathematical reasoning, programming, general-knowledge, and Text-to-SQL tasks. These findings matter for quality assurance and certification of AI systems, as they provide rigorous theoretical guarantees about model behavior beyond training data.
- Quality assurance
- Certifications
- Enterprise
Research
Contextualized Evaluation of Vision Language Models through Dynamic, Multi-turn Interactions
Yijiang Li, Huiqi Zou, Bingyang Wang et al.
arXiv · 2026-07-16
This paper introduces CEDI, a framework for evaluating multimodal large language models (MLLMs) through dynamic, multi-turn interactions rather than static benchmarks. CEDI uses a three-party setup—an evaluatee model, an automated examiner, and a grader—where the examiner navigates a graph-based task representation to deploy strategies like clarification requests and adversarial probes. Applied to visual hallucinations, CEDI reveals significantly more hallucinations than conventional static evaluation, with hallucinations accumulating over long contexts and models proving especially vulnerable to questions requiring premise rejection or refusal. The findings matter for quality assurance of AI systems, highlighting that real-world model behavior can diverge substantially from controlled benchmark performance.
- Quality assurance
- Certifications
Research
Manufactured Divisiveness: Decomposing the Hostile Content of Seven Social Media Influence Operations
Emilio Ferrara
arXiv · 2026-07-16
This paper challenges the common claim that state-backed social media influence operations are major sources of 'hate' content, arguing that widely used detectors conflate true hate speech with partisan attacks and geopolitical rhetoric—a measurement error that inflates hate rates roughly twofold. Analyzing 25 million tweets from seven government-attributed campaigns (8,275 accounts) in the Twitter Information Operations archive, the researchers use an LLM-based detector and an auditable classification rule to split hostile content into identity-based attacks (50.1%), partisan attacks (30.4%), and state/foreign-policy invective (19.5%), finding that only 18.7% meets a stricter definition of dehumanizing or inciting hate. The study further shows that six of seven campaigns fall into three distinct regimes—identity hate (Russian operations), geopolitical invective (Iranian operations), and partisan divisiveness (Venezuelan operations)—which a single aggregate 'hate' metric obscures. These findings have direct implications for how platforms, policymakers, and researchers measure and report on online influence operations and hate speech.
- AI policy
- Quality assurance
Research
Lower-Resource, Higher Scores: Language Bias in LLM Evaluators
Ej Zhou, Lucas Resck, Zheng Hui et al.
arXiv · 2026-07-16
This paper demonstrates that LLM-based evaluators — including trained reward models and LLM-as-a-Judge systems — exhibit a systematic language bias: lower-resource languages receive significantly higher scores than higher-resource ones, even when the instruction-response content is semantically identical. The bias is statistically significant and consistent across eight open-weight evaluators and frontier judges, yet remains invisible to standard pairwise accuracy metrics, with evaluators achieving above 90% pairwise accuracy while showing up to a 43% difference in acceptance rates across languages. This has direct safety implications: harmful content in lower-resource languages is more likely to pass safety filters. The authors find that model uncertainty is linked to the effect but cannot fully explain it, pointing to a structural, language-level misalignment in how these evaluators score multilingual content.
- Quality assurance
- AI policy
- Certifications
Research
BioTIER: A Refusal Benchmark for Targeted Biological Risk Mitigation
Eleanor M. Marshall, Pedro Medeiros, Peter Peneder et al.
arXiv · 2026-07-16
BioTIER is a benchmark designed to improve how large language models handle biological safety by distinguishing genuinely high-risk information from legitimate scientific content. It organizes 542 expert-curated prompts into three risk tiers — Catastrophe Avoidance, Biomedical DURC, and Related Biology — spanning a spectrum from extremely narrow high-risk topics to broad, benign biological knowledge. The benchmark aims to help developers implement more targeted refusal policies that block the small fraction of information posing catastrophic misuse risk while preserving access to beneficial scientific knowledge. This matters for AI policy and quality assurance because current models either over-refuse benign content or freely provide dangerous information, both of which represent failures in targeted mitigation.
- AI policy
- Quality assurance
- Certifications
Research
Beyond Generalist LLMs: Specialist Agentic Systems for Structured Code Workflow Execution
Harris Borman, Herman Wandabwa, Fusun Yu et al.
arXiv · 2026-07-16
This paper investigates whether specialist AI agents outperform general-purpose LLM-based agents for a specific software task: transforming Business Process Model and Notation (BPMN) diagrams into executable agentic workflows. The researchers introduce a specialist workflow system and benchmark it against generalist agents (Roo and Cline), finding that the specialist solution achieves 9–20 percentage points higher tool-use exactness, 2–4x better penalty-adjusted latency, 3x fewer tool-call errors, over 95% reduction in token generation cost, and eliminates repair iterations. The study also finds that generalist agents produce code inconsistently in both functionality and quality, raising concerns about their reliability in industrial settings. These results suggest that domain-specific agentic systems offer meaningful advantages for enterprise software automation tasks where reliability and maintainability are critical.
- Enterprise
- Quality assurance
- Workforce
Research
Do Generative AI Assistants Respect robots.txt? Tracing Web Access Beyond Visible Answers
Gabriel Lopez-Fonseca, David Rodriguez, Stefan Bechtold et al.
arXiv (Cornell University) · 2026-07-16
This paper presents a controlled empirical study of ten AI assistants with web-search capabilities, examining whether they comply with robots.txt restrictions that website owners use to control automated access. Using server-side logs and secret codes embedded in target pages, the researchers tested four access conditions across 200 trials, finding substantial variation: some assistants followed expected access rules, while others retrieved restricted content without checking robots.txt or used generic user-agents that obscured attribution. The study also reveals that retrieval behavior and answer correctness can diverge—assistants may access pages without surfacing the content, or fail to retrieve even permitted resources. These findings raise legal and governance concerns about content owner rights and call for updated, enforceable web governance standards in the era of search-augmented AI.
- AI policy
- Enterprise
Research
Tactile: Giving Computer-Using Agents Hands and Feet
Yong Liu, Zhenyi Zhong, Zhanpeng Shi
arXiv · 2026-07-16
Tactile is an open-source tool layer designed to give computer-using AI agents more reliable control over desktop applications. Instead of the brittle approach of predicting pixel coordinates from screenshots, Tactile converts UI evidence—including OS accessibility semantics, OCR-grounded text, and visual fallback regions—into structured, verifiable action targets. On macOSWorld-style tasks, adding Tactile improved Codex Success@100 from 41.1% to 50.0% overall and from 45.2% to 55.3% on accessibility-adapted tasks, with consistent gains across multiple agents including Codex, Claude Code, OpenCode, and Goose. The paper argues that reliable computer use requires not just stronger models but a reusable execution substrate that exposes software actions as semantic, verifiable, and auditable objects.
- Enterprise
- Quality assurance
Research
Cybersecurity Maturity and Risk Profiling of AI-Enabled Medical Sensors: A Cross-Manufacturer Comparative Analysis
Filip Tsvetanov
International Journal of Online and Biomedical Engineering (iJOE) · 2026-07-16
This paper proposes an integrated methodology for assessing cybersecurity maturity in AI-enabled medical sensors—including hospital patches, wearables, implantable cardiac devices, and continuous glucose monitors—using ISO 14971, NIST 800-30/53, and Analytic Hierarchy Process (AHP) weighting across ten criteria. The framework generates two composite indicators, a weighted security score and a risk profile score, enabling cross-manufacturer comparison. The analysis identifies critical vulnerabilities in communication security and AI modules of specific device categories, while finding greater cryptographic resilience in implantable systems. The findings are directly relevant to engineers, clinicians, and procurement organizations, and the authors call for unified AI-oriented security standards across the full medical sensor lifecycle.
- Certifications
- Quality assurance
Research
The role of artificial intelligence in detecting and preventing academic dishonesty in higher education: a systematic review
Walid Salamah
Frontiers in Education · 2026-07-16
This systematic review synthesizes 33 peer-reviewed studies (2012–2024) on how AI tools—including machine learning, NLP, stylometric analysis, and learning analytics—are being used to detect and prevent academic dishonesty in higher education. Four themes emerged: AI-based detection mechanisms, AI-based prevention strategies, ethical and institutional challenges (such as algorithmic bias, false positives, and data privacy), and critical gaps in evidence from developing and conflict-affected contexts. The review concludes that while AI shows significant potential for strengthening academic integrity, its effectiveness is constrained by bias, opacity, and variable institutional readiness, and that AI should complement rather than replace human judgment. The findings are directly relevant to quality-assurance and policy frameworks in higher education, particularly regarding governance and equitable deployment of AI integrity systems.
- Quality assurance
- AI policy
Research
Potencial de automatización laboral del modelo generativo de inteligencia artificial en las ocupaciones de México
María del Mar Oviedo Facundo, Luís Huesca Reynoso, David Castro Lugo
Revista de Economía Facultad de Economía Universidad Autónoma de Yucatán · 2026-07-16
This paper introduces the GPT Multimodal Automation Indicator (GMI/IGAM), a task-based metric designed to measure how much generative AI models like GPT-5 could automate occupations in Mexico. Applying the indicator to Mexico's 2019 SINCO occupational data using World Economic Forum methodology, the study finds that managerial, administrative, and sales divisions face low-to-medium automation potential (28–38%), while agricultural and services divisions face lower exposure (11–17%). Critically, some individual tasks show automation potential as high as 88%, meaning that even within low-risk occupations, specific tasks carry significant displacement risk. The findings suggest GPT-5 primarily complements rather than replaces occupations, but the concentration of highly automatable tasks within certain roles raises meaningful job displacement concerns for Mexico's workforce.
- Workforce
- AI policy
Research
Cheaper AI, More Informality? A Dual Labor Market Model for Developing Economies
Gabriel Montes‐Rojas, Fernando Toledo, Juan Manuel Rodríguez Repeti
arXiv (Cornell University) · 2026-07-16
This paper builds a small open economy DSGE model with a dual labor market to examine how falling AI prices affect formal versus informal employment in developing economies. The key finding is that outcomes hinge on whether imported AI capital substitutes for or complements formal workers: under substitution, cheaper AI erodes formal labor demand and pushes workers into informality, while under complementarity it expands formal employment, wages, and output. The model highlights that the same technological trend can either displace workers or drive formal-sector growth depending on the structural relationship between AI and human labor.
- Workforce
- AI policy
Research
Recent Advances in AI for Automated ICD Coding: A Systematic Literature Review
Abdul Rehman Khalid, Haider Ali, Kounen Fathima et al.
Journal of Medical Systems · 2026-07-16
This systematic review of 54 studies (drawn from 4,280 citations, 2019–2024) examines AI approaches to automating ICD code assignment from clinical text such as discharge summaries and electronic health records. The review traces a clear evolution from traditional machine learning to deep learning architectures—including convolutional, recurrent, transformer, and hybrid models—and finds that models perform better on frequent codes than on rare ones. Key persistent gaps include overreliance on single-language, single-institution datasets, poor rare-code prediction, limited model interpretability, and inconsistent evaluation protocols that impede cross-study comparison. The authors propose a 5P research agenda emphasizing Population Diversity, Performance Robustness, Prediction of Rare Codes, Provenance Transparency, and Practical Integration to guide deployment in real-world healthcare systems.
- Quality assurance
- Enterprise