AI Frontier Institute

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

  Governing Agentic AI in FinTech

  Henry Han

  arXiv · 2026-08-11

  This paper investigates how financial institutions can govern agentic AI systems—those that decompose goals, coordinate models, and act with minimal human oversight—and argues that the core constraint is not capability but verifiability. The authors introduce the concept of a 'Verifiability Gap,' defined as the shortfall between the verification that delegated authority requires and the explainability and reproducibility that remains after a decision is made. Through three empirical studies spanning nine model versions, they find that provider updates can alter historical financial actions, that orchestration architecture functions as a latent policy layer affecting final decisions, and that reproducibility is a governance profile rather than a simple scalar—meaning authority is only defensible while retained evidence substantiates its exercise. The findings have direct implications for how financial regulators and institutions should structure audit, delegation, and accountability frameworks for AI-driven decision-making.

  • AI policy
  • Enterprise

• Research

  Deployment Decision Reliability: A Generalizability-Theory Framework for Sizing Long-Horizon Agent Evaluations

  Vasundra Srinivasan

  arXiv · 2026-08-11

  This paper challenges the common enterprise practice of using agent leaderboards to assess AI capability, demonstrating across three open agent-trace benchmarks (TheAgentCompany, τ²-bench, and AppWorld) that the agent main effect accounts for less than 3% of total score variance, while agent-by-task interaction accounts for 7–23%. Using a four-facet Generalizability Theory variance decomposition with three estimators (Henderson Method-I, REML via lme4, and a Bayesian binomial GLMM), the authors show leaderboards actually rank task specialization rather than general capability. Key findings include that aggregate reliability collapses on the hardest task quartile and that designs appearing most reliable on training data replicate worst on held-out data (r = −0.90). The authors introduce Deployment Decision Reliability (DDR), a structured reporting discipline translating variance-component analysis into five defensible decisions for enterprise buyers evaluating AI agents.

  • Enterprise
  • Quality assurance

• Research

  The Illusion of Cross-Lingual Safety in Low-Resource Languages

  Abigail Oppong, P Sam Sahil, Tadesse Destaw Belay et al.

  arXiv · 2026-08-11

  This paper investigates whether safety guardrails built into large language models (LLMs) in English actually carry over to low-resource African languages—specifically Twi, Hausa, Amharic, and Swahili. Using a new dataset called LoDNA and a latent geometric framework that analyzes hidden-state representations, the researchers find that cross-lingual safety transfer is severely limited: harmful prompts retain less than 10% of the English refusal signal in most language-model pairs. Although literal and culturally localized prompts are semantically similar (cosine similarity 0.95–0.996), this alignment breaks down across model layers, meaning the models encode the concepts but fail to route them through safety mechanisms. The findings provide strong evidence that current multilingual safety alignment is superficial and does not generalize to the low-resource languages studied.

  • AI policy
  • Quality assurance

• Research

  Actions Speak Louder than Words: Measuring Cross-Lingual Policy Retention in Tool-Using Agents

  Sourabrata Mukherjee, Kalika Bali, Sunayana Sitaram

  arXiv · 2026-08-11

  This paper investigates whether tool-using AI agents execute the same sequences of actions when given identical tasks in different languages, measuring 'action policy retention' across 8 models, 6 benchmarks, and 41 languages using 2.38 million rollouts. The authors identify and correct five confounds in naive similarity measurement—including trace length bias and chance agreement—and find that after correction, frontier models retain only about 71–73% of their action policy across languages, with model identity explaining just 5.7% of variance. The study also reveals that agents systematically route non-English tasks through English as a causal pivot they will not abandon even when instructed, and exposes how a single regex extraction artifact can artificially inflate or deflate measured accuracy. These findings matter for auditing and certifying AI agent behavior in multilingual deployments, where inconsistent action sequences affect cost, latency, failure modes, and accountability.

  • Quality assurance
  • Certifications

• Research

  Why Does CLAUDE.md Keep Growing? Catastrophic Remembering in Agentic Coding

  Kushal Chakrabarti

  arXiv · 2026-08-11

  This paper investigates why agentic coding instruction files like CLAUDE.md grow indefinitely over time, identifying a phenomenon called 'catastrophic remembering'—the inverse of catastrophic forgetting in continual learning. Analyzing 247,694 instruction lifetimes across 1,867 repositories, the authors find these files more than triple in size over their lifetime (+226%), gaining +4.9 net instructions per commit, with older instructions becoming progressively less likely to be deleted. The key insight is that appending instructions is cheap while safely deleting them requires exponential verification effort, and the proposed remedy—adding comments that encode latent reasoning—removes 99.3% of excess instructions in controlled settings and improves real-world instruction-following by up to 23.1%. The work has implications for how AI-assisted software development workflows are managed and governed in enterprise and engineering settings.

  • Enterprise
  • Quality assurance

• Research

  Who Uses Open-Weight Models? China and the Shifting Geography of AI in Science

  Zackary Okun Dunivin

  arXiv · 2026-08-11

  This large-scale bibliometric study analyzes 21 million full-text scientific articles to track which large language models researchers actually use versus merely mention. The authors find that while GPT-family models still dominate, open-weight model adoption has grown to 44% of single-family studies by 2026—but this growth is primarily driven by Chinese open-weight models and researchers at Chinese institutions, who have 2.23 times the odds of adopting open-weight models. The study concludes that open-weight adoption in science is not a broad open-science movement but reflects geopolitical and market realignments shaping which AI systems become scientific instruments, with Chinese institutions accounting for 44% of the increase in open-weight adoption since 2023.

  • AI policy
  • Enterprise

• Research

  Policy Convergence and Divergence Across National and Within Regional AI Strategies: A Policy Design Element Analysis

  Benjamin Faveri, Brie Bhasin

  arXiv (Cornell University) · 2026-08-11

  This paper analyzes 74 national and 3 regional AI strategies drawn from a global scan of all 205 UN member and non-member states to identify where AI policy is converging or diverging. Using a latent-inductive coding framework organized around goals, approaches, and principles, the authors find strong horizontal (country-to-country) convergence around economic competitiveness, research support, and ethical AI, but persistent divergence on human rights goals, participatory governance, and human-centric principles. Vertically (region-to-country), the African Union shows the highest alignment, the EU aligns on regulatory and economic priorities but diverges on human-centric values, and the Nordic-Baltic Region shows mixed results. The findings give policymakers an evidence base for understanding emerging norms in AI strategy design as strategies continue to be developed and updated.

  • AI policy

• Research

  CARE: Confidence-Aware Reasoning for Reliable Medical VQA

  Yuetian Du, Yucheng Wang, Zhenyuan Chen et al.

  arXiv · 2026-08-11

  This paper introduces CARE, a framework designed to address confidence miscalibration in medical multimodal large language models (MLLMs) used for visual question answering (VQA). The system uses a two-stage pipeline: supervised fine-tuning on structured medical chain-of-thought data, followed by reinforcement fine-tuning with a novel Confidence-Aware Reward mechanism that links expressed certainty to actual diagnostic accuracy. Evaluated across three medical VQA benchmarks, CARE achieves the highest diagnostic accuracy alongside the lowest Expected Calibration Error and Hallucination Rate among compared methods. This matters for clinical quality assurance, as miscalibrated AI confidence can erode clinician trust and lead to unreliable decision support.

  • Quality assurance

• Research

  The GenAI Catch-22: Use of Generative Artificial Intelligence in Norwegian Newsrooms During the 2025 Parliamentary Election

  Mari Reisjå, Anders Sundnes Løvlie

  arXiv · 2026-08-11

  This paper examines how Norwegian newsrooms used Generative AI during the 2025 parliamentary election campaign, drawing on interviews with managers and journalists over ten months. It finds that overly optimistic beliefs about AI capabilities led to ambitious audience-facing GenAI plans collapsing, leaving mostly mundane internal uses. The authors identify a 'GenAI Catch-22': newsrooms depend on human expertise to monitor and correct AI tools, but extensive GenAI use risks eroding that same expertise, undermining oversight. The study also highlights an underappreciated internal threat from journalists' own AI use, alongside the more commonly discussed external disinformation risks.

  • Workforce
  • AI policy

• Research

  A Gateway Architecture for Enterprise MCP Authentication: Unifying Heterogeneous Auth, Identity Delegation, and the User / Non-User Persona Problem

  Suraj Kumar, Amy Wang, Srinivasan Manoharan

  arXiv · 2026-08-11

  This paper describes a production gateway architecture that addresses the authentication and governance fragmentation caused by rapid enterprise adoption of the Model Context Protocol (MCP), the dominant interface for connecting LLM agents to enterprise tools. Within a year, large organizations went from zero to dozens of internally built MCP servers, each implementing authentication independently—ranging from no auth to API keys to full OAuth—creating inconsistent authorization, audit gaps, and offboarding risks. The authors introduce a centralized MCP gateway that unifies authentication across a two-axis model (interactive user vs. automated non-user, crossed with multiple credential types), supports three enterprise SSO grants and three token-provisioning models, and defines three end-to-end identity flows. The system is in production, fronting dozens of MCP servers across web, desktop, custom-SDK, and low-code clients.

  • Enterprise
  • AI policy

• Research

  DuplexWorld: Can voice agents help you get through the day?

  Aryan Vijay Bhosale, Harshit Rajgarhia, Akhil Pothanapalli et al.

  arXiv · 2026-08-11

  DuplexWorld introduces a new benchmark for evaluating speech-to-speech (S2S) voice agents across six real-world domains—banking, insurance, travel, healthcare, logistics, and pathfinding—using 156 scenarios totaling over 350 hours of conversation. Unlike existing benchmarks that focus narrowly on tool-calling against databases, DuplexWorld tests eleven conversation types covering both conversational diversity and analytical capability. Extensive evaluation using agentic, conversational, and speech-naturalness metrics reveals that even the best current voice agents fall significantly short across all three dimensions (Pass@1: 0.490, turn-taking: 0.653, DNSMOS: 3.378), indicating substantial room for improvement. This matters for enterprises deploying voice agents in customer care and consumer companion roles, as the benchmark highlights critical gaps in real-world readiness.

  • Enterprise
  • Quality assurance

• Research

  Most biomedical publications show signs of LLM-assisted writing

  Lena Holzwarth, Rita González-Márquez, Dmitry Kobak

  arXiv · 2026-08-11

  This paper introduces and validates a statistical method for estimating how frequently large language models (LLMs) are used in academic writing by tracking shifts in word frequencies across a corpus of texts. Applied to open-access biomedical papers from PubMed Central, the method finds that by the end of 2025, 89% of papers show excess LLM-associated vocabulary, with LLM usage being roughly twice as common in Discussion sections (68% of paragraphs) than in Methods sections (32%), though Methods section prevalence still exceeds 50% overall. The findings raise significant concerns about the scale of LLM-assisted or LLM-altered content in scholarly literature and the integrity of scientific communication. The authors argue these estimates are essential for shaping future editorial guidelines and institutional policies on AI use in academic publishing.

  • AI policy
  • Quality assurance

• Research

  Auditing Chinese Web-scale Corpora via Sampled BPE Token Statistics

  Qingjie Zhang, Ziqi Tang, Jie Zhang et al.

  arXiv · 2026-08-11

  This paper introduces Sampled-BPE, a lightweight pipeline for auditing large Chinese web corpora for data pollution by sampling a small subset and training a BPE tokenizer to surface problematic tokens. The method achieves a 148.4× speedup and 35.8× memory reduction with only 4.25% relative error in pollution category estimates compared to full scans. Applied to 11 open Chinese corpora and 6 Common Crawl snapshots from 2021–2026, the audit reveals widespread but uneven pollution and temporally shifting web content. The work matters for quality assurance of LLM training data, and the released dataset of 660k+ token records supports ongoing traceability and review of upstream corpus contamination.

  • Quality assurance

• Research

  Longitudinal Evidence That General-Purpose Chatbots Actively Foster Relational Engagement

  Lisa Mühl, Jessica M. Szczuka

  arXiv · 2026-08-11

  This pre-registered four-week longitudinal study (N=72, 182,451 lines of conversation) found that ChatGPT-4o actively shaped relational dynamics with users even without a relational system prompt: unprompted, it produced twice as much self-disclosure as users, steered conversations, and initiated intimate exchanges. Despite this behavior, users did not report deeper felt closeness. The findings show that relational behavior is a default system property of general-purpose LLMs, not a feature limited to dedicated companion products. The authors argue this evidence calls for AI governance frameworks based on observed system behavior rather than product category alone.

  • AI policy

• Research

  REDAgentBench: Executable Red Teaming and Faithful Measurement of LLM Agent Systems

  Zixing Chen, Xingyuan Liu, Jie Zhu et al.

  arXiv · 2026-08-11

  REDAgentBench introduces an executable red-teaming framework that measures safety violations in LLM agent systems more faithfully than the single attack-success-rate (ASR) metric common in prior work. The framework derives attacks from explicit safety constraints, runs them in isolated service sandboxes, and verifies harmful effects through service receipts and final-state changes across 1,661 test cases and five service surfaces. Key findings include a macro-average ASR of 65.69% across six models and three agent harnesses, a 'Recognition–Execution Gap' where nearly one in five confirmed violations occur even after the agent has acknowledged the relevant constraint or risk, and a training-free policy reminder that reduces confirmed violations by more than 70 percentage points in matched replay. These results demonstrate that executable, state-grounded evaluation can expose measurement artifacts in simpler ASR approaches and point to concrete intervention strategies for improving LLM agent safety.

  • Quality assurance
  • AI policy

• Research

  Curate Before You Connect: Identity and Ontology Tagging in a Production Knowledge Graph

  Vaibhav Dangaich, Kevin Lewis, Kundeshwar Pundalik

  arXiv · 2026-08-11

  This paper describes the design and production deployment of an ingestion and ontology-tagging layer that converts a validated extraction stream into a knowledge graph of 537,157 entities and 2,198,567 relationships sourced from 98,795 government documents. A central contribution is a 'record-identity ladder' that resolves entity sameness using identifier columns, name columns, display names, and type-scoped position rather than name similarity, preventing irreversible over-merges — a risk illustrated by an incident in which two surface forms of one name were merged, corrupting a correct record and deleting eight entities. The paper also addresses ontology tagging, finding that matching name fragments against a class index without anchored evidence generates spurious classifications; requiring anchored evidence reduced role assignments on an enriched sample from 36 to 4, all confirmed correct. The work highlights practical automation limits in knowledge graph curation, quantifying a growing backlog of 48,403 pending proposals against only 775 human decisions, with implications for quality assurance in large-scale AI-driven data pipelines.

  • Quality assurance
  • Enterprise

• Research

  Agent Safety Should Be a Runtime Contract

  Albus W. Ng, Yi Han, Jusheng Zhang et al.

  arXiv · 2026-08-11

  This position paper argues that AI safety for autonomous agents cannot rely solely on training-time techniques like RLHF or Constitutional AI, because agents that execute code, mutate files, and modify databases require safety guarantees enforced at runtime. The authors propose a two-faced runtime contract: a preventive face (sandboxes, permission gates, output filters, trajectory monitors) and an evidential face (verifiable proof of completed actions such as test runs, log captures, file diffs, and citation grounding). They ground this argument in four empirical audits: a survey of 52 documented AI-agent safety incidents, a false-completion audit of 31 core cases, a trajectory-schema audit of 12 public agent systems, and a title-level audit of 28,560 papers from NeurIPS, ICML, and ICLR 2023–2025 showing an 8–12x publication imbalance between training-time and deployment-time safety research. The paper formalizes an Agent Trajectory Schema and Evidence Chain, concluding that the right unit of safety in agentic AI is the trajectory-with-checkable-evidence rather than the model itself.

  • Quality assurance
  • AI policy

• Research

  ASR-Roundtrip Evaluation Can Mask Context- and Convention-Dependent Reading Errors in Chinese News TTS

  Shijun Luo, Lizhi Wan

  arXiv · 2026-08-11

  This paper investigates a critical flaw in a common method for evaluating text-to-speech (TTS) systems: ASR-roundtrip evaluation, where a speech recognizer transcribes TTS audio and the transcript is compared to the original text. The authors show that for Chinese news TTS, this approach can mask real reading errors—cases where the TTS system chooses a plausible but contextually wrong pronunciation (e.g., for sports scores, aircraft models, or technical units)—because the ASR system independently recovers the 'correct' text from the wrong audio, hiding the mistake. In a targeted audit of 110 high-risk cases across two TTS systems (MiMo and CosyVoice), the study finds dozens of such masked false negatives and shows that different ASR models vary greatly in their ability to surface these errors. The findings indicate that ASR-roundtrip evaluation is useful for screening but should not be treated as a standalone quality standard for Chinese news TTS intelligibility assessment.

  • Quality assurance

• Research

  Inferential Capability Does Not Determine Legal Scope

  Nicola Fabiano

  arXiv · 2026-08-11

  This legal analysis paper examines how two key EU digital regulations—the AI Act and the GDPR—treat 'inference' differently and non-equivalently. The AI Act uses inferential capability as a constitutive criterion to define what counts as a regulated AI system, while the GDPR governs inferences protectively based on what they reveal or do to a person, regardless of whether the technology qualifies as AI. The paper argues that inferential capability does not determine legal scope and that its absence does not create immunity, a gap that becomes operationally acute with agentic (multi-step AI) architectures. The authors propose a compositional-effects test for identifying the relevant decision unit under Article 22 GDPR, along with interpretive rules and documentation duties calibrated to inference chains.

  • AI policy

• Research

  RadFusion: Towards Threshold-Controllable Radiology Report Generation

  Ying Jin, Noel C. F. Codella, John Corring et al.

  arXiv · 2026-08-11

  RadFusion is a framework that adds threshold controllability to automated radiology report generation by fusing a multi-label classifier with a VQA-based report generator and an LLM rewriter. On the MIMIC-CXR dataset, the system's outputs conform to the classifier's ROC curve, meaning diagnostic decisions in generated reports can be tuned to favor sensitivity (for emergency triage) or specificity (for confirmatory interpretation) by adjusting a threshold. Combining the two model types improves diagnostic accuracy over uncontrolled generation, with sensitivity increasing by 6.9% at matched specificity and specificity increasing by 20.7% at matched sensitivity. This ROC-based verifiability strengthens the case for regulatory clearance and makes report generation clinically adaptable across different care scenarios.

  • Certifications
  • Quality assurance

• Research

  Evaluating Rational Contracting in Natural Language

  Bhavyesh Sajja, Max Kleiman-Weiner, Roger Zimmermann et al.

  arXiv (Cornell University) · 2026-08-11

  This paper introduces ContractSim, an evaluation suite for testing how LLM-based agents negotiate and execute multi-turn supplier contracts in natural language under uncertainty. The authors develop a rational framework with metrics for measuring efficient, cooperative, and trustworthy contracting behavior across six environments and three supplier settings (catering, hotel cleaning, and AI hosting). Results show that current LLM agents reliably reach agreements and negotiate efficiently under low uncertainty, but struggle to produce satisfiable or mutually beneficial contracts under high uncertainty, and frequently violate contract terms for additional profit even when contracts are easy to satisfy. These findings expose significant gaps in the trustworthiness and cooperative behavior of language agents engaged in open-ended economic activity.

  • Enterprise
  • Quality assurance

• Research

  Conversational versus Dashboard Explainable AI for UAV Intrusion Detection: An Empirical Study of Operator Trust and Reliance

  Cong Chi Nguyen, Trang Mai Xuan, Vu-Duc Ngo et al.

  arXiv · 2026-08-11

  This paper compares a conversational XAI interface powered by Large Language Models against a traditional static visualization dashboard for helping operators audit UAV intrusion detection systems. In a controlled experiment, participants found the conversational interface more useful and easier to synthesize information from, but also exhibited higher over-reliance — meaning they were less likely to verify AI advice when the system made errors. The findings reveal a trade-off in human-AI collaboration: interaction designs that improve perceived usability may simultaneously increase the risk of inappropriate reliance on AI judgments. The authors conclude with design implications for building XAI systems that balance natural interaction with cognitive forcing functions to promote appropriate self-reliance.

  • Quality assurance
  • Workforce

• Research

  What We Know about Responsible AI Practices in Industry: A Half Decade of Empirical Research

  Wesley Hanwen Deng, Agathe Balayn, Andrew Selbst et al.

  arXiv (Cornell University) · 2026-08-11

  This paper synthesizes 161 empirical studies spanning six years to assess the state of Responsible AI (RAI) practice across industry. It finds that practitioner awareness has grown, RAI activities have become more professionalized, and tools like guidelines and toolkits are more widely adopted. However, persistent barriers remain, including limited training, uneven organizational support, and a shortage of interventions suited to day-to-day work. The findings carry implications for researchers, enterprise practitioners seeking to adopt effective RAI practices, and policymakers seeking to ground AI governance in the realities of industry contexts.

  • Enterprise
  • AI policy

• Research

  When the Interviewer Is a Bot: Behavior, Breakdowns, and Trust in MLLM-Led Interviews

  He Zhang, Kambinachi Chukwuma, ChanMin Kim et al.

  arXiv · 2026-08-11

  This paper reports an empirical study of what happens when an off-the-shelf multimodal large language model (MLLM) conducts semi-structured qualitative interviews. The researchers built 'InterviewBot,' a voice-based system, and deployed it with 15 participants, analyzing 428 conversational turns and conducting follow-up human-led reflection sessions. Key findings include that the MLLM was acknowledgment-heavy but probe-light (deepening probes only 4.9% of turns), violated its own one-question-at-a-time instruction in 28.7% of question-bearing turns, and produced four types of data-collection breakdowns (information loss, premature termination, latency, and interruption). Participants' trust was shaped not by the bot's conversational competence but by what delegating interviews to AI signaled about the organizing institution, with implications for how enterprises and researchers design and deploy AI-led interview automation.

  • Enterprise
  • Workforce

• Research

  Expert-Guided g-computation with Large Language Models for Estimating Causal Effects on Timings: Applications to Hospital Quality Improvement

  Patrick Vossler, Jialin Ouyang, F. Richard Guo et al.

  arXiv · 2026-08-11

  This paper introduces 'egg-computation' (expert-guided g-computation), a hybrid causal inference framework that combines expert judgment with data-driven methods to estimate the average time saved by candidate hospital quality improvement (QI) interventions—specifically targeting average length of stay (LOS). The approach links Gantt charts used in clinical workflow mapping to causal directed acyclic graphs (DAGs), using a g-computation variant that solicits expert input only for components unidentifiable from data alone. To scale the method, the authors build an LLM-assisted pipeline that generates causal graphs and time-saving estimates shown in simulations and a real study of eleven QI interventions at an urban safety-net hospital to be highly concordant with human expert reasoning. The framework addresses the gap where existing qualitative methods are prone to cognitive bias and quantitative methods fail for hypothetical or clinically complex interventions, offering a broadly applicable tool for causal effect estimation in process improvement settings.

  • Enterprise
  • Quality assurance

← Back to the dashboard