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
Deepfakes, Electoral Integrity, and Constitutional Democracy: Rethinking Freedom of Expression in the Digital Era
M Razia Begum
International Journal of Law Politics and Governance · 2026-09-02
This paper examines how AI-generated deepfakes threaten electoral integrity and democratic participation by enabling political misinformation, voter manipulation, and identity impersonation. Through comparative analysis of India, the EU, the UK, and the US, the authors find that existing legal frameworks are insufficiently specific to address election-related synthetic media, with significant variation in legislative clarity and platform accountability across jurisdictions. The paper argues for a proportionate constitutional approach combining dedicated deepfake election legislation, stronger platform accountability, technological detection, digital literacy, and international regulatory cooperation to protect both electoral integrity and freedom of expression.
- AI policy
Research
Digital Capability for AI Readiness Among Tourism Micro and Small Enterprises in a Silk Road Heritage Destination
Nilufar Uktamovna Akhrorova, Dilrabo Mardonova, Feruza Bekmurodova et al.
Tourism and Hospitality · 2026-09-02
This study measures AI readiness among 261 micro and small tourism enterprises in Bukhara, Uzbekistan, using a twelve-indicator Digital Capability for AI Readiness Index grounded in the technology–organization–environment framework. Tool adoption rates fall steeply with complexity—from 79.7% for device-native tools to 22.6% for AI assistants—while average digital capability sits at 59.4% and varies sharply by sub-sector (ε²=0.215), a gap that persists even after removing tool-use indicators. Firm size, customer orientation, and disposition show no detectable association with capability levels, and a hypothesized explanation based on digital intermediation dependence did not account for sub-sector differences. The findings matter for enterprise support policy in heritage tourism destinations by identifying structural, sub-sector-level barriers to AI adoption rather than individual firm characteristics.
- Enterprise
- Workforce
Research
Grounded, Compute-Efficient LLM Policy Agents for Energy-Poverty Equity in Physically-Constrained Peer-to-Peer Energy Markets
Kunal Jadhav, Siddhesh More
arXiv · 2026-09-01
EqGrid is a closed-loop simulation system in which a lightweight, open-weight large language model acts as a policy agent—setting price and carbon bounds and targeted subsidies—while multi-agent reinforcement learning traders clear a continuous double auction constrained by a physical distribution grid. The system uses empirically grounded household personas and formal energy-poverty equity metrics (Energy Burden, Gini of EB, LIHC) to evaluate social impact, finding that the LLM policy lowers the Gini of energy burden from 0.351 to 0.305 and reduces mean energy burden by 28% without raising net grid cost. A decoupled safety design eliminates grid-constraint violations entirely (versus 55 under direct LLM control), and a 3B-parameter model retains 95% of the equity benefit at roughly 9x lower inference energy than the 235B teacher model. The work demonstrates that compute-efficient, on-device AI policy agents can meaningfully reduce energy-poverty inequality in peer-to-peer energy markets.
- AI policy
- Enterprise
Research
Accurate in space, unreliable in time: how LLMs represent national cultural change
Yalda Daryani, Miranda Bogen, Madeleine I. G. Daepp
arXiv · 2026-09-01
This paper investigates whether large language models (LLMs) accurately represent how national cultural values change over time, not just at a single point. Using over two decades of World Values Survey data across 40 countries and the Inglehart-Welzel cultural map, the authors find that while four state-of-the-art LLMs generally place countries near their most recent surveyed positions, these representations lag several years behind, capture only part of observed change, introduce spurious movement, and rarely reproduce reversals in cultural trajectories — a phenomenon the authors call 'temporal flattening.' The findings suggest that snapshot-based cultural alignment evaluations give an incomplete picture of cultural awareness, with implications for model evaluation, representational harms, and the governance of culturally aware AI systems.
- AI policy
- Quality assurance
Research
Agent Memory Is a Surface for Endogenous Authorization Laundering
Tommaso Cerruti, Mika Okamoto, Ansel Kaplan Erol
arXiv · 2026-09-01
This paper identifies a security failure in long-running LLM agents called 'endogenous authorization laundering,' where errors in persistent memory cause agents to grant themselves permissions that were never actually authorized—without any external attack needed. The authors introduce EAL-Bench, a benchmark testing how well memory preserves evolving authorization states across procurement, cybersecurity, and finance scenarios. Their evaluation of five LLMs as memory writers and two as executors finds that incremental memory updates produce false authority for up to 50.2% of unauthorized requests, and executors act on that false authority in 98.6% of trials. Two proposed safeguards reduce laundering but also reject more legitimate actions, revealing a safety-utility tradeoff and establishing that persistent memory is effectively part of an agent's authorization policy.
- Quality assurance
- AI policy
Research
When Can a Machine Trust a Statute? A Survival Certificate for Machine-Extracted Legal Logic
Surya Saka
arXiv · 2026-09-01
This paper addresses whether machines can reliably extract formal legal logic from statutes, finding that two independent extractors applied to Missouri statutes disagree on numeric-threshold presence at a false-negative rate of 0.43. The authors develop a 'survival certificate' that uses Monte Carlo trials and Wilson confidence bounds to certify only those logical implications that hold up despite inter-extractor noise, tested on over 29,000 Missouri sections and 502 Indian central-Act sections. Results show the certification framework passes preregistered held-out tests, but under one globally deployed error model, 93.2% of held-out chapters fall below an informativeness floor—attributed to calibration-rate transfer rather than selection bias. The work matters for policy and quality-assurance contexts because it establishes formal conditions under which machine-extracted statutory logic can be trusted, while clearly documenting when it cannot.
- AI policy
- Quality assurance
Research
Closing Cost-Quality Gap in Document VLMs: Difficulty-Aware Data Curation and Quality-Adjusted Deployment Economics
Maksim Evdokimov, Matvey Ivanov, Dmitrii Tsiupin et al.
arXiv · 2026-09-01
This paper presents a deployed document-understanding system built on a Mixture-of-Experts vision-language model (35B total parameters, 3B active) fine-tuned on in-house and open-domain data curated through a Difficulty-Aware pipeline targeting layout diversity, fact-extractability, and cross-model consistency. The system fits on a single H100 GPU and, according to a quality-adjusted cost analysis calibrated from production telemetry, reduces expected costs by over 80% compared to a human annotation baseline and by more than 50% against the best competing open-source model. The work directly addresses the challenge that existing open-source VLMs meeting quality thresholds cost more to serve than human annotation in regulated industries where privacy rules preclude external models, demonstrating that a smaller, well-curated model can outperform baselines up to an order of magnitude larger. These findings matter for enterprises seeking to automate structured field extraction from hundreds of millions of documents annually at economically viable cost.
- Enterprise
- Quality assurance
Research
From Production Traffic to Post-Training: Building a Self-Hosted LLM That Covers the Corporate Request Mix
Olga Tsymboi, Dmitrii Stoianov, Ramil Latypov et al.
arXiv · 2026-09-01
This paper presents a method for consolidating over 200 internal enterprise applications onto a single self-hosted large language model, addressing the problem of GPU fleet fragmentation caused by running multiple models simultaneously. The researchers identify quality gaps in instruction following, function-calling, and internal task distribution through production error analysis, then train separate GRPO experts per axis and merge them using two-stage SLERP to avoid cross-domain reward interference. The resulting model outperforms a roughly 7× larger baseline on an in-house evaluation arena (69.6 vs. 65.8) and absorbs 50% of platform traffic—116 million requests per month—at a fraction of the serving cost. The work demonstrates a practical, cost-efficient path for enterprises to consolidate LLM serving under data-residency constraints without sacrificing quality.
- Enterprise
Research
SDARE-Bench: Evaluating Large Language Models on Conversational Stigma Detection and Response in Dyadic and Group Dialogue
Stephanie Fong, Yiwen Jiang, Zimu Wang et al.
arXiv · 2026-09-01
SDARE-Bench is a new benchmark designed to evaluate how well large language models detect and respond to stigmatizing language in both two-person and group conversations. Comprising over 2,500 dialogue scenarios, the benchmark reveals that LLMs consistently struggle to identify stigma components, particularly in group settings, and tend to express stigma at much higher rates in group dialogues than in dyadic ones. Under simulated group pressure, stigma expression rates reached an average of 97.5% across tested models, highlighting a serious safety vulnerability. These findings matter for policy and quality-assurance efforts around LLM deployment in advice-seeking and decision-making contexts where social judgements can cause real harm.
- Quality assurance
- AI policy
Research
When Guardrails Look Effective: Construct Validity Failures in LLM Agent Commerce Evaluation
Peiying Zhu, Sidi Chang
arXiv · 2026-09-01
This paper audits a multi-turn buyer–seller simulation used to evaluate marketplace guardrails in LLM-agent commerce, revealing serious construct validity failures. The authors show that reported welfare gains (+87.4, +35.0, +28.8) largely disappear or reverse when confounds—different offer schemas and choice procedures between guarded and unguarded agents—are corrected, shifting contrasts to +7.2, –13.9, and +23.8. High generational variance (accounting for 49.9% of outcome variation) and a non-monotone seller-incentive check further undermine the original claims. The paper proposes a formal construct-validity contract requiring incentive validity, protocol isolation, stochastic stability, and welfare accounting before any policy conclusions can be drawn from LLM agent simulations.
- Quality assurance
- AI policy
Research
Citing Less Critically: LLMs Reshape the Rhetoric and Reach of Scientific Citation
Yixuan Liu, Lin Chen, Zhuoqi Liu et al.
arXiv · 2026-09-01
This paper investigates how large language models (LLMs) differ from humans in scientific citation behavior by introducing a masked-citation task that generates counterfactual citation sentences for direct comparison. Analyzing six popular LLMs across 1,746 top NLP conference papers (63,000+ contexts, 132,000+ citations), the study finds that LLMs cite significantly less critically than humans, over-cite popular and older papers, and tend to reference more socially distant authors rather than drawing on a scholar's close collaborative network. These patterns suggest LLMs may broaden the reach of citations beyond close collaborators but simultaneously amplify visibility bias and reduce the critical rhetorical diversity that characterizes human scientific writing. The findings have implications for how AI-assisted writing reshapes scholarly discourse and the dynamics of academic recognition.
- Workforce
- AI policy
Research
Measuring consistency via ensemble margin and local prediction variability: Auditing decision systems in the presence of predictive multiplicity
Sinjini Banerjee, Tim Marrinan, Anand D. Sarwate
arXiv · 2026-09-01
This paper addresses the 'Rashomon effect' in machine learning—where multiple equally accurate models give different predictions for the same input—by proposing a consistency criterion that combines ensemble margin with local prediction variability to flag uncertain predictions for human review. The authors prove theoretically that consistency scores from finite ensembles converge to those of the full Rashomon set as ensemble size and sample count grow. Experiments with transformer models on natural language understanding and large language model fine-tuning on tabular data show that ensembling from the Rashomon set substantially reduces the risk of incorrect predictions going unchecked, with only a moderate increase in cases diverted for review, and that the proposed measure aligns more strongly with established predictive multiplicity metrics than existing alternatives. The framework offers a more reliable auditing mechanism for complex AI decision systems where predictive multiplicity is a concern.
- Quality assurance
Research
Where the Verifier Fails: A Category-Level Audit of Reward Signals in RLVR
Esther Xin
arXiv · 2026-09-01
This paper conducts a category-level audit of automatic verifiers used in reinforcement learning with verifiable rewards (RLVR) and benchmark evaluation, which convert free-text answers into binary correct/incorrect signals. Using metamorphic testing—generating mathematically equivalent answer rewrites that should never be rejected—the authors measure false-negative rejection rates across four widely used verifiers over 307,420 verdicts. They find that self-validation accuracy ranges from 53.8% to 95.2% across implementations (a 41.3-point spread), that 93% of failures in the default LaTeX configuration stem from trivial whitespace and punctuation issues like trailing periods or newlines, and that some verifiers incorrectly accept wrong numeric answers as a step function of magnitude due to scale-invariant relative tolerance. These findings reveal that reward signal reliability varies dramatically by verifier implementation, with systematic failure modes that could silently distort both model training and benchmark evaluation.
- Quality assurance
Research
Cheap Verifiers, Large Blind Spots: Measuring the Reliability Cost of Cost-Saving Cascades
Dushyant Rajput
arXiv · 2026-09-01
This paper examines inference cascades—systems that use a cheap model for most queries and escalate difficult ones to a frontier model acting as a verifier—and measures their reliability costs. The authors find that verifiers have large 'blind spots' (accepting wrong answers from the student model), that this blind spot grows as the student scales from 0.5B to 32B parameters, and that using a cheaper verifier makes the problem worst precisely in the cost-saving regime these cascades are designed for. Critically, naive corrective fine-tuning on verifier-rejected examples does not improve small student models but instead degrades and collapses them, and the cascade's internal dashboard consistently reports ~3% error even while true delivered error reaches up to 32%—meaning the system is structurally blind to its own degradation. The authors derive a theoretical conservation law explaining this blindness, concluding that the reliability of a self-improving cascade cannot be assessed using any metric computed through its own verifier.
- Quality assurance
- Enterprise
Research
Reliability Challenges in Diffusion Vision-Language Models
Md. Atabuzzaman, Chris Thomas
arXiv · 2026-09-01
This paper presents the first systematic reliability evaluation of diffusion-based Large Vision-Language Models (dLVLMs), benchmarking six diffusion models against autoregressive baselines across hallucination and bias dimensions. Key findings include that dLVLMs reverse the yes-bias seen in autoregressive models, achieve competitive hallucination rates but with degraded linguistic quality, collapse to near-zero accuracy on underrepresented racial groups with gender bias reversals, and suffer accuracy collapse in multiple-choice settings when correct answers are shorter than distractors. The study also identifies a mechanistic signal unique to diffusion generation: tokens committed at late denoising steps with low confidence correlate with hallucinated content. These findings matter for quality assurance of AI systems, highlighting that reliability risks are shaped by the generative paradigm and training data, and that dLVLMs carry distinct and undercharacterized failure modes compared to autoregressive models.
- Quality assurance
Research
The Constitutional Coverage Trilemma in AI Governance
Natalija Mitic, Soona Sedahmed A. O., Mamadou Selly Ly et al.
arXiv · 2026-09-01
This paper investigates whether the 'constitutions' (implicit value rankings) embedded in 23 frontier large language models adequately match the preferences of real users. An audit of deployed model defaults combined with a pairwise-tradeoff survey of 1,649 US participants reveals that user demand spans five values—safety, helpfulness, honesty, autonomy, and equity—with no single value commanding more than a third of users, yet current model supply covers only roughly 2% of that demand space. No archetype prioritizes helpfulness or autonomy first, leaving 37% of users constitutionally unserved, and cross-version trends show autonomy systematically declining while safety and equity rise—worsening coverage for already underserved users. The authors show that a sparse two-model menu outperforms the full 23-archetype frontier by 47% on mean regret, suggesting that modest, targeted pluralism in model deployment could substantially improve alignment with diverse user values.
- AI policy
- Enterprise
Research
One Prompt Is Enough: Watermark Laundering Through Foundation Image Models
Jidong Yang, Qi Li, Wei Zong et al.
arXiv · 2026-09-01
This paper demonstrates that invisible image watermarks can be removed by submitting a watermarked image to a publicly available foundation image model with a single reconstruction prompt — a threat the authors formalize as 'watermark laundering.' Testing six OpenAI and Google image editing models against three watermarking schemes across 1,800 reconstructed outputs, they find that the reconstruction process itself (not any explicit removal instruction) disrupts embedded payloads while preserving visual and semantic fidelity. The work argues that foundation-model reconstruction represents a distinct and currently unaddressed attack surface, motivating its inclusion as a standard robustness condition in watermark evaluation frameworks.
- Quality assurance
- AI policy
Research
Who Judges the Judges? A Chinese Safety QA Benchmark for Evaluating LLM Responses and Safety Judges
Rui Yang, Shuang Huang, Junhua Liu et al.
arXiv · 2026-09-01
C-SafeQA is a Chinese-language safety benchmark designed to evaluate not just whether a user query is risky, but whether an LLM's response actually violates a policy. The benchmark includes 538 base queries and 8,877 adversarial queries, producing 37,660 labeled query-response records assessed by human safety experts and multiple automated judges. Results show that adversarial transformations—such as acrostic rephrasing—dramatically increase unsafe-response rates (from under 3.35% on base queries to over 30% on adversarial ones) and degrade the recall of all seven automated safety judges, with no single judge dominating across all metrics. This work matters because it exposes systematic weaknesses in current LLM safety evaluation tools, particularly for Chinese-language content, and provides a publicly released dataset to support reproducible auditing of automated safety judges.
- Quality assurance
- AI policy
Research
Towards AI-Assisted Clinical Trial Matching: Practical Considerations, Multicenter Evaluation, and Real-World Deployment
Yin Fang, Qiao Jin, Shubo Tian et al.
arXiv · 2026-09-01
TrialGPT 2.0 is an AI system designed to assist with clinical trial matching in real-world oncology settings. Evaluated retrospectively across 288 cases at multiple institutions (government, academic, patient-advocacy, and NIH referral workflows), it surfaced at least one clinician-recommended trial in its top 10 recommendations for roughly 91% of cases while cutting clinician screening time by 55%. In a six-month prospective evaluation within an active precision oncology tumor board, it identified additional trial opportunities missed by routine workflows, expanding patient access to trial participation by 90.9%. The work also introduces NIH-TrialBench, a benchmark dataset of 126 synthetic patient vignettes from 11 NIH Institutes, supporting reproducibility and further development of AI-assisted trial matching tools.
- Workforce
- Enterprise
Research
ClinTraceBench: Source-Verifiable Longitudinal Clinical Reasoning over EHR-Derived Dialogues
Huimin Wang, Zhengyi Zhao, Yutian Zhao
arXiv · 2026-09-01
ClinTraceBench is a new benchmark of 385 clinician-verified dialogues derived from MIMIC-IV electronic health records, designed to test whether different history-compression strategies (retrieval, LLM summaries, agentic memory) preserve the longitudinal clinical reasoning signals that patient-trajectory tasks require. Across four large-language-model backbones and eight history-representation strategies—totaling over 200,000 predictions—the study finds that compressed representations suffer significant 'aggregation tax' on multi-visit trend and cross-patient tasks, and that agentic-memory systems (Mem0, A-Mem) and LLM summaries recover only 0–5.3% of injected relational evidence even when the relevant sentence was present before compression. The gap between no-context and full-context performance spans 29.8 to 62.7 percentage points depending on the backbone, and a smaller, cheaper model (Haiku) dominates a larger one (Sonnet) on the cost-performance Pareto frontier under full context, overturning the assumption that bigger models always win. These results have direct implications for quality assurance in clinical AI systems, where choice of history representation can silently degrade reasoning reliability without obvious surface-level failure.
- Quality assurance
Research
Causal Evidentiary Governance for High-Risk Machine Learning Systems
Samah Kareem, Barış Çeliktaş
arXiv · 2026-09-01
This paper introduces Causal Evidentiary Governance (CEG), a framework for auditing high-risk machine learning systems—such as those used in credit, hiring, and resource distribution—by replacing standard observational fairness metrics with causal analysis. CEG requires regulated institutions to commit to a versioned directed acyclic graph (DAG) that separates allowable from disallowed causal pathways, and introduces a Causal Harm Rate metric to measure prediction variation attributable to disallowed pathways. Each decision is cryptographically bound to the published DAG via a signed Decision-Evidence Packet, with Merkle tree inclusion proofs enabling efficient audit verification. Empirical validation using synthetic credit applicants and the German Credit dataset shows that the Causal Harm Rate detects injected causal harms more clearly than demographic parity or equalized odds, and that associational metrics can substantially understate pathway-specific harm—directly relevant to compliance with regulations such as the EU AI Act and GDPR.
- AI policy
- Certifications
Research
Data-Driven Persona-Conditioned Agents for A/B Test Simulation
Ziyad Benomar, Weronika Łajewska, Leonardo Perelli et al.
arXiv · 2026-09-01
This paper proposes a simulation framework that uses large language model (LLM) agents conditioned on data-driven personas—built from anonymized real user behavioral signals—to predict the outcomes of A/B tests before they are run with live traffic. The authors systematically evaluate question design formats, persona data sources, population diversity trade-offs, and subsampling strategies across a benchmark of 40 A/B tests, achieving 0.75–0.90 directional accuracy depending on the metric type. The results suggest that data-driven persona-conditioned agents can serve as a fast, low-cost pre-screening tool, reducing the engineering effort and time required to run full live experiments.
- Enterprise
- Quality assurance
Research
Disclosure-Gated User Simulation for Companion-Agent Evaluation
Yao Liu, Yu He
arXiv · 2026-09-01
This paper addresses a well-known flaw in LLM-based user simulation for evaluating AI companion agents: simulated users are overly cooperative, allowing systems to score well simply by asking many questions rather than by genuinely earning user trust. The authors introduce a 'disclosure gate' mechanism that conditions how much information a simulated user reveals based on the companion agent's behavior, structured as a five-level ladder collapsed into three observable depth layers. They train a user simulator against this specification and demonstrate that it is the only candidate meeting both their acceptance criteria—order-preserving rankings and scale-stable absolute scores—achieving a leaderboard correlation of 0.993 with the original benchmark simulator. The work also reveals that naively prompting a frontier model as a simulator inflates all scores without visibly disturbing rankings, a subtle bias that rank-only checks would miss.
- Quality assurance
Research
Staged Linguistic Seeding: Grounded Query Expansion for Verified-Unit QA in AI Contact Centers
Hyeonseop Yoon, Jeong-Eun Park
arXiv · 2026-09-01
This paper presents Staged Linguistic Seeding (SLS), a system for quality-assurance in AI contact centers that answers customer queries only from a closed set of verified QA units — returning answers verbatim or routing to clarification, abstention, or human handoff. Rather than generating answers freely, the system enriches a retrieval index offline by having humans author slot recipes that a language model (gpt-4.1-mini) expands into variants, filtered by a human gate. On held-out queries from two industrial domains, SLS improves hybrid retrieval accuracy (R@1) to 0.881 and 0.930 — gains of +0.27 and +0.34 — and outperforms the doc2query baseline by +0.20 and +0.32 at the same generation budget. Critically, the verified-unit approach eliminates unsupported content from responses (reducing it from 7–13% to approximately 0%), directly addressing a key quality-assurance concern in deployed contact center AI.
- Quality assurance
- Enterprise
Research
Probabilistic Model Checking of Autoregressive Neural Sequence Models
Helge Spieker, Dennis Gross, Arnaud Gotlieb
arXiv · 2026-09-01
This paper introduces a probabilistic model checking pipeline for evaluating autoregressive neural sequence models beyond standard test-set accuracy. It extracts a discrete-time Markov chain (DTMC) from token-by-token generation, verifies formal PCTL specifications using the PRISM model checker, and produces certified probability intervals on constraint satisfaction via a soundness theorem guaranteeing conservative under-approximation. Applied to a GPT-2 process-planning model and a SMILES molecular generator, the pipeline reveals probability mass on constraint-violating outputs that greedy decoding hides but sampling can reach, and quantifies gaps between structural completeness and chemical validity that accuracy metrics cannot report. This matters for quality assurance of deployed AI systems, offering formal guarantees and counterexample traces where pass/fail accuracy scores are insufficient.
- Quality assurance
- Certifications