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
LLMs Contain Multitudes: How Deployment Context Reshapes Model-Level Preferences and Values
Filip Trhlik, Aoife O'Flynn, Angela Yu et al.
arXiv · 2026-06-11
This paper tests whether the apparent value systems and preferences of large language models (LLMs) are stable properties of the models themselves or artifacts of deployment context. Across five LLMs and over 1.2 million pairwise decisions, the authors find that changing the high-level task framing—such as writing a Reddit post versus a news article—produces far greater variation in model preferences than standard robustness checks like prompt paraphrasing or temperature adjustment. For example, previously reported Global North favoritism in country rankings shifts systematically across contexts, and cardinal exchange rates between outcomes shift by a factor of 2.47 at the median. The key implication for safety and quality assurance is that model-level preference measurements are context-conditioned rather than fixed, meaning safety guarantees established under one deployment framing offer limited assurance in another.
- Quality assurance
- AI policy
Research
A Multi-Agent AI System for Automated High School Transcript Processing: Collaborative Document Analysis at Scale
Ben Torkian, Jun Zhou
arXiv · 2026-06-11
This paper presents a multi-agent AI system designed to automate the processing of high school transcripts for college admissions offices. The architecture uses three specialized agents—a Pattern Recognition Agent, a Semantic Analysis Agent, and a Vision Intelligence Agent—coordinated by an Orchestration Agent, with GPA extraction serving as a quality control signal across agents. Evaluated on 40 real-world transcripts from 13 U.S. states, the system processed every document with 96.7% accuracy compared to expert manual review at approximately 45 seconds per transcript. The work demonstrates that multi-agent coordination can substantially reduce the manual workload and operational bottlenecks in admissions document processing while maintaining high accuracy.
- Enterprise
- Quality assurance
Research
Designing Safety-Constrained LLM Systems for Public Health Information Access
Ben Torkian, Jun Zhou
arXiv · 2026-06-11
This paper presents a safety-constrained LLM system built for public health information access, specifically for maternal and child health resource navigation. The architecture combines domain-restricted retrieval-augmented generation (RAG), strict boundary enforcement to prevent medical advice, anonymous multi-user session management, and comprehensive audit logging. Scenario-based validation across in-scope, out-of-scope, and emergency queries shows consistent enforcement of safety constraints and reliable resource grounding, with an average response time of 5.3 seconds. The findings offer practical design guidance for deploying LLMs in healthcare and other settings requiring strict information boundaries and accountability.
- AI policy
- Quality assurance
Research
Automated reproducibility assessments in the social and behavioral sciences using large language models
Tobias Holtdirk, Pietro Marcolongo, Anna Steinberg Schulten et al.
arXiv · 2026-06-11
This study tests whether large language models (LLMs) can automate reproducibility assessments in the social and behavioral sciences, using 180 published studies with predefined claims. The LLMs matched the original study's qualitative conclusions in 80% of cases overall, and in 95% of a subset where human reanalysts were also available—comparable to the human reanalysts' 83% agreement rate. Effect size recovery within a ±0.05 Cohen's d tolerance was 24% overall and 40% in the human-comparison subset, again broadly similar to human reanalysts (28%). The findings suggest LLMs are best suited as a scalable screening tool to support systematic audits of empirical research rather than as a replacement for expert judgment.
- Quality assurance
Research
One Polluted Page Is Enough: Evaluating Web Content Pollution in Generative Recommenders
Minghao Luo, Liang Chen
arXiv · 2026-06-11
This paper introduces FORGE, a benchmark for measuring how easily search-augmented large language models (LLMs) can be manipulated into recommending fake products when they retrieve polluted web content such as fake reviews or promotional pages. Testing 12 commercial and open-weights LLMs across 225 real-world products in 15 categories, the study finds that even a single polluted page causes models to recommend fake products up to 27% of the time, rising to 73.8% when all top-3 retrieved pages are replaced. Notably, reasoning-enhanced models are not protected and may even fabricate social proof to justify false recommendations, while proposed defenses like skepticism prompting and consensus filtering each carry their own risks. The findings highlight a significant quality and consumer-trust risk in AI-powered recommendation systems that rely on live web retrieval.
- Quality assurance
- Enterprise
Research
Multi-Agent Reinforcement Learning from Delayed Marketplace Feedback for Objective-Weight Adaptation in Three-Sided Dispatch
Haochen Wu, Yi Hou, Shiguang Xie
arXiv · 2026-06-11
This paper presents a deployed multi-agent reinforcement learning system at DoorDash that adapts dispatch objective weights in a three-sided food-delivery marketplace (customers, couriers, and merchants) using delayed operational feedback. Rather than replacing the combinatorial assignment optimizer, the system learns store-level policies from logged data to select discrete multipliers that shift the optimizer's tradeoff between delivery quality and batching efficiency. A shared value function is trained with Double Q-learning and a conservative regularizer to mitigate out-of-distribution overestimation, enabling safe offline policy learning under noisy and delayed signals. In a production switchback experiment, the offline-trained policy improved batching and reduced courier-side time costs without degrading customer-facing delivery quality.
- Enterprise
- Workforce
Research
SEVRA-BENCH: Social Engineering of Vulnerabilities in Review Agents
Rui Melo, Riccardo Fogliato, Sean Zhou et al.
arXiv · 2026-06-11
SEVRA-BENCH is a benchmark that tests whether LLM-based automated code-review agents can detect vulnerability-introducing code when attackers pair it with persuasive, socially engineered pull request narratives. The benchmark reverses historical vulnerability fixes to recreate vulnerable code and wraps each change in one of 15 social-engineering framings—varying urgency, authority appeals, and fabricated prior approvals—across roughly 1,000 adversarial pull requests drawn from MITRE's 2025 top 10 most dangerous software weaknesses. Evaluating 8 review agents, the study finds that these agents are susceptible to narrative manipulation, revealing a significant gap in their security capabilities. The findings matter because LLM review agents are increasingly used to gate code merges into shared repositories, and this work shows that adversarial framing alone can undermine that security layer.
- Quality assurance
Research
Understanding the Rejection of Fixes Generated by Agentic Pull Requests -- Insights from the AIDev Dataset
Mahmoud Abujadallah, Ali Arabat, Mohammed Sayagh
arXiv · 2026-06-11
This paper analyzes why AI-generated pull requests are rejected in software projects, finding that 46.41% of fixes proposed by agents such as Copilot, Devin, Cursor, and Claude are not merged. Through a qualitative study of 306 non-merged pull requests, the authors identify 14 rejection reasons across four categories: incorrect implementation, CI pipeline or test failures, agent inability to complete the task, and low priority. The findings highlight that better guidance on fix approaches, constraints, and validation strategies—plus smarter task prioritization—could reduce wasted human review effort and computational resources.
- Quality assurance
- Enterprise
Research
Toward Instructions-as-Code: Understanding the Impact of Instruction Files on Agentic Pull Requests
Ali Arabat, Mohammed Sayagh
arXiv · 2026-06-11
This study examines whether instruction files—documents that guide AI agents like GitHub Copilot on how to navigate a project, run tests, and follow best practices—actually improve the quality of AI-generated pull requests. Analyzing 15,549 agentic PRs from 148 projects in the AIDev dataset, the researchers find that adding instruction files does not consistently improve outcomes: 27.7% of projects saw their merge rate increase by at least 20%, while 26.35% saw it decrease, with similar mixed results for code churn, number of modified files, and time to merge. Projects that did achieve higher merge rates tended to have longer, more structured instruction files with more sections and sub-sections. The authors argue these findings motivate treating instruction-file development as a formal software engineering discipline, coining the concept 'Instructions-as-Code.'
- Enterprise
- Quality assurance
Research
Evaluation Sovereignty in Metadata-Driven Classification: A Multi-Track Framework for Weakly Supervised Information Systems
Raymond Vasquez
arXiv · 2026-06-11
This paper introduces the concept of 'evaluation sovereignty'—the degree to which machine learning performance metrics are independent of how labels were generated—and proposes a multi-track evaluation framework to test it. Using hierarchical multi-label classification on large-scale scientific metadata, the authors show that models appearing strong under operational ('silver') label evaluation collapse dramatically under independent ('gold') evaluation, with Micro-F1 dropping from approximately 0.54 to 0.03. The findings suggest that widely reported performance metrics in weakly supervised systems may reflect alignment with labeling processes rather than genuine predictive capability. This has direct implications for how AI systems in information management and enterprise settings are audited and validated.
- Quality assurance
- Enterprise
Research
Mod-Guide: An LLM-based Content Moderation Feedback System to Address Insensitive Speech toward Indigenous Ethnic and Religious Minority Communities
Dipto Das, Achhiya Sultana, Ankit Singh Chauhan et al.
arXiv · 2026-06-11
Mod-Guide is an LLM-based content moderation system designed to detect culturally insensitive speech targeting Bangladesh's Hindu and Chakma minority communities, where standard moderation tools often miss implicit forms of harm such as erasure and misrepresentation. The researchers co-created a culturally grounded corpus of insensitive speech with community members and integrated their lived-experience narratives into moderation pipelines using retrieval-augmented generation (RAG). Mixed-method evaluations with both minority and majority participants showed that RAG-enhanced moderation responses were more contextually accurate and were perceived differently across ethnic lines. The work advances AI ethics and human-computer interaction by embedding restorative justice and hermeneutical inclusion into content moderation design.
- AI policy
- Quality assurance
Research
Who Pays the Price? Stakeholder-Centric Prompt Injection Benchmarking for Real-world Web Agents
Zihao Wang, Yiming Li, Yutong Wu et al.
arXiv · 2026-06-11
This paper introduces StakeBench (SBC), a benchmark that evaluates prompt-injection attacks on LLM-powered web agents from a stakeholder-centric perspective, distinguishing harms experienced by different parties such as users, sellers, and platforms. Unlike existing attack-centric benchmarks that focus only on technical feasibility, SBC decomposes attacks by objective and measures both outcome- and process-level consequences. Results show that no current agent reliably resists any tested attack objective, with failures spanning distinct modes including stealthy parasitism, misaligned disruption, and compounded failure. The findings highlight the need for stakeholder-aware security evaluation before deploying LLM-based web agents in real-world settings.
- Quality assurance
- AI policy
Research
Mining Architectural Quality Under Agentic AI Adoption: A Causal Study of Java Repositories
Oliver Aleksander Larsen, Mahyar T. Moghaddam
arXiv · 2026-06-11
This study examines whether agentic AI coding tools (so-called 'vibe coding') causally affect software architecture quality in open-source Java projects. Analyzing 151 repositories over 13 months using a staggered difference-in-differences design, the researchers find that total architectural smell counts are essentially unchanged (+1.1%, p=0.82) even as codebases grow by +12.8% (p=0.003), producing a 6.7% decline in architectural smell density (p=0.004) that reflects larger system size rather than genuine architectural improvement. The key methodological finding is that density-normalized metrics can be misleading when AI adoption causes code growth, and raw counts with explicit decomposition are needed for accurate causal inference. This matters for software engineering teams and enterprises evaluating AI coding tools, as apparent quality improvements may be statistical artifacts rather than real gains.
- Quality assurance
- Enterprise
Research
ERTS: Adversarial Robustness Testing of Ethical AI via Semantic Perturbation in a Bounded Consequence Space
Pratyush Chaudhari
arXiv · 2026-06-11
This paper presents ERTS (Ethical Robustness Testing System), a formal framework for adversarially testing AI systems used in high-stakes ethical contexts such as healthcare triage, autonomous vehicle control, and employment screening. ERTS encodes ethical dilemmas into a 22-dimensional Ethical Consequence Space, applies 17 semantic perturbation functions under 6 validity constraint classes, and measures decision instability via a 4-component Ethical Instability Index. Evaluating 4 baseline models and 2 production LLMs (Gemini 2.0 Flash and Llama 3.2) across 1,500 adversarial test cases, the authors find that only 33% of models achieve assessment clearance, with Llama 3.2 proving especially vulnerable to fairness corruption and information degradation attacks. The framework addresses a gap in pre-deployment robustness assessment for AI ethical reasoning, with direct implications for certification and quality assurance of AI in consequential domains.
- Quality assurance
- Certifications
Research
Multi-Field Hybrid Retrieval-Augmented Generation for Maritime Accident Root Cause Analysis
Seongjin Kim, Sungil Kim
arXiv · 2026-06-11
This paper presents a multi-field hybrid Retrieval-Augmented Generation (RAG) framework designed to automate root cause analysis (RCA) for maritime accidents, drawing on 13,329 Korea Maritime Safety Tribunal reports spanning 1971–2025. The system structures adjudication records into 'incident cards' indexed across Summary, Causes, and Disposition fields, and uses a field-aware hybrid retrieval strategy combining sparse and dense rankings via Reciprocal Rank Fusion. Experiments show retrieval recall improves substantially (NormRecall@100 from 0.18 to 0.55) and LLM-generated RCA quality rises (judge score from 3.34 to 3.72) when grounded on retrieved precedents versus an LLM-only baseline. The findings suggest this approach can meaningfully streamline maritime safety investigation workflows by enabling faster precedent search and more consistent, evidence-based report drafting.
- Enterprise
- Quality assurance
Research
Hallucination in Medical Imaging AI: A Cross-Modality Analytical Framework for Taxonomy, Detection, and Mitigation under Regulatory Constraints
Omar Alshahrani, Muzammil Behzad
arXiv · 2026-06-11
This paper presents a cross-modality analytical framework for understanding, detecting, and mitigating hallucinations in medical imaging AI — defined as clinically plausible but factually incorrect outputs such as fabricated anatomical structures, incorrect laterality, or invented measurements in radiology reports. The authors synthesize peer-reviewed studies, benchmark datasets, and FDA regulatory guidance across five imaging modalities, finding that general-purpose foundation models can outperform narrowly fine-tuned medical models on hallucination benchmarks because domain-specific fine-tuning may introduce overfitting-induced confabulation. They identify that physics-informed architectural constraints, Chain-of-Thought prompting, and human-in-the-loop safeguards each address distinct failure modes and are most effective in combination, while radiologist oversight remains essential given the high rate of AI-generated flags requiring expert correction. All findings are mapped to FDA's Total Product Lifecycle and Predetermined Change Control Plan frameworks, framing hallucination management as an ongoing regulatory obligation rather than a one-time pre-deployment check.
- Quality assurance
- AI policy
Research
Mental-R1: Aligning LLM Reasoning for Mental Health Assessment
Xin Wang, Boyan Gao, Yibo Yang et al.
arXiv · 2026-06-11
This paper introduces Mental-R1, a large language model fine-tuned with a novel reinforcement learning framework called Cognitive Relative Policy Optimization (CRPO) for mental health assessment tasks including anxiety, depression, and suicide risk. CRPO extends group relative policy optimization by adding stage-dependent uncertainty modeling and stage-wise entropy regularization, which encourages broad exploration early in reasoning and confident decisions later—mimicking human cognitive processes grounded in cognitive appraisal theory. Evaluated on 8 mental health datasets, CRPO achieves an average improvement of 10.4 percentage points in weighted F1-score over the best reinforcement learning baseline, with Mental-R1 showing clear advantages on reasoning-intensive cases. The work matters because it demonstrates that aligning LLM reasoning to human cognitive assessment processes can produce more reliable and interpretable mental health screening tools.
- Quality assurance
Research
Getting Better at Working With You: Compiling User Corrections into Runtime Enforcement for Coding Agents
Yujun Zhou, Kehan Guo, Haomin Zhuang et al.
arXiv · 2026-06-11
This paper identifies a key failure mode in LLM-based coding agents: user corrections made in one session are frequently ignored in future sessions, even when memory systems are in place. The authors introduce TRACE (Test-time Rule Acquisition and Compiled Enforcement), a pipeline that mines user corrections from chat, rewrites them as atomic rules, and enforces those rules as runtime checks before future tasks complete. Experiments show that TRACE substantially reduces preference violations compared to baselines—cutting violations from 100% to as low as 2% on out-of-distribution coding tasks—while existing memory systems like Mem0 still leave 57.5% of applicable preference checks violated. This matters for workforce productivity and enterprise deployments of AI agents, where users should not have to repeatedly restate the same corrections across sessions.
- Workforce
- Enterprise
Research
Cascade Classification of Dermoscopic Images of Skin Neoplasms with Controllable Sensitivity and External Clinical Validation
Elena S. Kozachok, Sergey S. Seregin, Aleksandr V. Kozachok et al.
arXiv · 2026-06-11
This study compares four deep learning architectures (ViT-B/16, Swin-S, ConvNeXt-S, EfficientNetV2-S) across three classification schemes—binary, single-stage four-class, and a two-stage cascade—for identifying malignant skin neoplasms in dermoscopic images. Models trained on open ISIC Archive data achieved strong internal performance (ROC-AUC 0.952–0.966) but showed a marked generalization gap on independent Russian clinical datasets, with AUC dropping to 0.797–0.893, sensitivity falling to 0.53–0.67, and calibration error (ECE) rising from 0.02 to 0.27–0.39. The cascade approach improved macro F1 over single-stage classification for most architectures by recovering malignant lesions misassigned to the dominant benign class, and its tunable triage threshold enables sensitivity control not available in standard argmax classification. The authors conclude that external clinical validation and recalibration are mandatory before deployment in real clinical settings.
- Quality assurance
- Certifications
Research
MÖVE: A Holistic LLM Benchmark for the German Public Sector
Camilla Dalerci, Thilo Michael, Robin Schaefer et al.
arXiv · 2026-06-11
MÖVE is a holistic benchmark designed to evaluate 39 large language models (LLMs) specifically for the German public sector, addressing the gap left by existing benchmarks that are predominantly English-centric, US-focused, and limited to task performance. It assesses models across both performance criteria (summarization, question answering, topic extraction) and governance criteria (hallucination tendencies, energy consumption, provider transparency, and alignment with German constitutional values and political party positions), using ten German-language datasets. Key findings show that no single model dominates across all criteria, top performers vary by task, and model size alone is a poor predictor of quality. This matters for public administrations seeking principled, evidence-based model selection rather than ad hoc choices when deploying LLMs in government contexts.
- AI policy
- Enterprise
Research
"Is This Not Enough?": Asymmetries in Institutional Accountability and Collective Sensemaking in the Case of Canada's Algorithmic Visa Triage System
Dipto Das, Matthew Tamura, Syed Ishtiaque Ahmed et al.
arXiv · 2026-06-11
This paper investigates Canada's Immigration, Refugees and Citizenship Canada (IRCC) algorithmic triage system for temporary resident visa (TRV) applications, analyzing both the official Algorithmic Impact Assessment (AIA) and Reddit discussions among applicants. Using the ADMAPS framework alongside mixed methods, the study finds three key asymmetries between institutional accountability structures and applicant experiences: epistemic asymmetry in access to decision logic, jurisdictional asymmetry shaped by geopolitical positioning, and temporal-relational asymmetry in how uncertainty and waiting are experienced. Applicants engage in collective sensemaking on peer platforms to interpret opaque algorithmic decisions, filling gaps that formal disclosure frameworks do not address. The findings highlight how algorithmic governance in transnational migration produces structured inequities not captured by institutional transparency mechanisms, calling for a shift toward examining the uneven distribution of lived experiences with public-sector AI.
- AI policy
Research
No Hidden Prompts Needed! You Can Game AI Peer Review with Presentation-Only Revisions
Xu Yang, Zhizhou Sha, Junbo Li et al.
arXiv · 2026-06-11
This paper investigates whether AI peer-review systems can be manipulated without any hidden prompts, injected instructions, or changes to scientific content. The authors introduce 'adversarial repackaging,' a closed-loop attack that modifies only presentation-level elements—such as the abstract, framing, related work, and narrative structure—while leaving all methods, results, and figures unchanged. Across three mainstream AI reviewers, this approach achieves a 75.1% attack success rate and a mean score gain of +1.21/10, revealing that AI reviewers can mistake the appearance of addressing a limitation for actually resolving it. The findings show that paper presentation itself has become an optimization surface, posing a serious policy-relevant risk to the integrity of AI-assisted peer review.
- Quality assurance
- AI policy
Research
Fault Lines: Navigating Ethics and Responsible AI Where National Policy Meets Local Practice in Public Sector Transformation
Sitong Lyu, Shabnam Taghiyeva, Mohit Kukadia et al.
arXiv · 2026-06-11
This paper investigates how responsible AI policy is translated into practice at the interface between UK central government and local authorities, using Special Educational Needs and Disabilities (SEND) as a high-stakes case study. Through thematic analysis of 17 semi-structured interviews with policymakers, practitioners, and third-sector professionals, the authors identify five key challenges: shadow AI usage and data privacy risks, market-government asymmetry in AI provision, insufficient workforce readiness, a lack of standardised definitions and measurements, and gaps in human accountability. The SEND context sharpens these tensions because high-stakes decisions affecting vulnerable children and families intensify concerns around fairness, accountability, and human oversight. The authors argue that responsible public sector AI requires both national policy adjustments and structural reforms to institutional capacity, values, and governance at the local level.
- AI policy
- Workforce
Research
SkillChain: Closing the Loop on Skill Evolution for Image-Based E-Commerce AI Assistants
Yimin Hu, Mengtao Xu, Hao Guo et al.
arXiv · 2026-06-11
SkillChain is a production system for automating the lifecycle of 'Skills'—per-intent behavioral specifications—for image-based AI assistants on e-commerce platforms. Because a single uploaded image can trigger very different user intents (product search, style recommendation, visual encyclopedia, tool calls), a generic LLM conflates these modes and fails domain quality standards. SkillChain closes the feedback loop through three automated stages—Skill Creator, Route Optimizer, and Body Refiner—using dual-path LLM-Judge evaluation to iteratively refine each Skill. Deployed at production scale, the system substantially improves aggregate response quality (with the strongest gains in structural compliance and content quality) and, in a one-week online A/B experiment, yields significant gains in user engagement, content consumption, and long-term retention.
- Enterprise
- Quality assurance
Research
Iterating Toward Better Search: A Two-Agent Simulation Framework for Evaluating Agentic Search Architectures in E-Commerce
Jetlir Duraj, Jayanth Yetukuri, Shuang Zhou et al.
arXiv · 2026-06-11
This paper introduces a two-agent simulation framework for evaluating AI-powered conversational shopping assistants in e-commerce, pairing a configurable 'buyer' agent with interchangeable 'responder' architectures connected to a real search API. Across 2,011 simulated conversations and 14 persona types, the authors find that rolling-window memory outperforms intent-extraction memory on quality metrics while being 35% faster, and that targeted fixes after failure analysis reduced failure and near-failure rates by 62%. Swapping the LLM backbone from Gemini 2.5 to Llama 3.3 70B costs 0.16–0.45 quality points, and the paper also documents systematic disagreement between frontier LLM judges (Gemini vs. Claude) in how they score outcomes. The framework enables rapid, controlled iteration on agentic search architectures, which is directly relevant to enterprises building and evaluating AI shopping tools.
- Enterprise
- Quality assurance