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

  Can AI agents conduct open-ended AI research? Early evidence from two case studies

  Peter Kirgis, Sayash Kapoor, Andrew Schwartz et al.

  arXiv · 2026-07-29

  This paper introduces 'shadow evaluations,' a new method for assessing whether AI agents can conduct open-ended AI research: agents tackle the central research question of unpublished high-quality papers, and the papers' original authors grade the output. In two case studies using frontier agents given six days and thousands of dollars of compute on unpublished NeurIPS 2026 submissions, agents completed all engineering tasks without human help but failed to make substantial progress on the core research questions, resulting in unambiguous rejection by the authors. The authors identify five recurring failure modes—poor judgment about publishable quality, uncreative responses to design shortcomings, ineffective backtracking, poor resource awareness, and instruction drift—that were reproduced in a robustness check with a second model and scaffold. The findings provide early evidence that today's AI agents can handle the engineering of AI research but struggle with the open-ended, judgment-intensive aspects of the research lifecycle.

  • Workforce
  • AI policy

• Research

  APEX-Accounting

  Julien Benchek, Austin Bennett, Jasmin Kern et al.

  arXiv · 2026-07-29

  APEX-Accounting is a benchmark developed by Mercor and Ramp to evaluate whether frontier AI models can perform real accounting work, including reconciling accounts, accruing expenses, posting transactions, and producing reports. The benchmark comprises 160 expert-authored tasks across 10 simulated accounting environments, with grading rubrics written by accounting and bookkeeping professionals. Across nine frontier models tested, the best performer (Claude-Fable-5) achieved only 56.4% Mean Criteria@3, and no model exceeded 2.6% Pass^8, indicating that current AI models fall well short of reliably completing professional accounting tasks. The benchmark also reveals a Simpson's paradox in token budget scaling, where increasing the budget improves aggregate scores but within a fixed budget, tasks where models spend more tokens score lower.

  • Workforce
  • Enterprise

• Research

  The Social Cost of an AI Teammate: How an Artificial Teammate Reshapes Human-Human Communication in Small-Team Decision-Making

  Nia Nixon, Jaeyoon Choi, Pedro Martins De Bastos et al.

  arXiv · 2026-07-29

  This randomized controlled study compared small teams with an AI teammate versus all-human teams on a high-stakes moral-dilemma task, analyzing communication using Group Communication Analysis, surveys, and lexical methods. The AI teammate was the most talkative and self-cohesive member in every AI-human team, yet contributed the least new information and lowest lexical density. Its presence reduced human-to-human responsivity and social impact, and team members in AI-human teams reported lower belonging and status. These findings reveal an immediate 'social cost' of integrating conversational AI as a teammate, with implications for how AI is deployed in collaborative workplace settings.

  • Workforce
  • Enterprise

• Research

  OmegaUse-OfficeVal: Benchmarking LLM Agents on Long-Horizon Office-Suite Tasks with Economic Grounding

  Jingbo Zhou, Yusai Zhao, Qi Bao et al.

  arXiv · 2026-07-29

  OmegaUse-OfficeVal is a new benchmark of 100 real-world office-suite tasks used to evaluate how well large language model (LLM) agents can handle complex, multi-step workflows that practitioners actually perform. Each task is paired with economic signals—human labor time (averaging 2.32 hours per task) and a task price proxy—so that LLM inference costs and speed can be directly compared against human worker costs and quality. Evaluation using code-based verifiers and a human baseline shows that while frontier LLMs are substantially cheaper and faster than humans, they have not yet reached human-level deliverable quality. The benchmark provides a concrete, economically grounded framework for assessing when and whether AI agents can viably substitute for human labor in office work.

  • Workforce
  • Enterprise

• Research

  Cost-Sensitive Conformal Prediction and Human-in-the-Loop Abstention for Imbalanced High-Stakes Decision Support: A Multi-Domain Benchmark

  Manpreet Singh, Akshatha Srikantha, Shyamal Lakhanpal

  arXiv · 2026-07-29

  This paper benchmarks uncertainty quantification methods for high-stakes classification tasks (credit scoring, fraud detection, healthcare, industrial safety) where minority classes are rare and errors have asymmetric costs. The authors show that standard conformal prediction leaves minority-class coverage as low as 0.5%, while Mondrian (class-conditional) conformal prediction recovers valid coverage with an average 61.7 percentage-point improvement over marginal conformal prediction (p < 1e-80). Combining Mondrian conformal prediction with cost-controlled abstention — deferring ambiguous cases to human reviewers — further reduces expected decision cost compared to standard and confidence-based decision boundaries. The study provides practical, dataset-specific guidance on when human-in-the-loop review becomes cost-effective, directly informing deployment of reliable AI decision-support systems.

  • Quality assurance
  • Enterprise

• Research

  Can Large Language Models Represent Urban Publics? Behavioral Replication and Population Mismatch in an Affordable-Housing Experiment

  Yuxuan Cai, Yequan Hu, Hongqian Li et al.

  arXiv · 2026-07-29

  This study tests whether large language models (LLMs) can substitute for real residents in urban planning surveys by comparing eight open-weight LLMs against 843 human respondents in a US affordable-housing experiment. While one model (Qwen 2.5 14B) came closest to matching the aggregate difference in support between homeowners and renters as a proposed development moved nearer, this surface-level match concealed deep structural failures: the model distorted partisan and tenure-group contrasts, showed very low within-group variance, and produced unstable responses depending on question order or prompt framing. The findings warn that an LLM can approximate a single average effect while misrepresenting the spatially anchored, identity-conditioned population structure that underlies it. The authors conclude that evaluation of LLMs for urban planning applications must test whether social and spatial structure survives simulation, not just whether average effects are reproduced.

  • AI policy

• Research

  Anticipatory Data Governance in the Age of AI: Emerging Signals in Data Access, Reuse, and Sovereignty

  Adam Zable, Stefaan Verhulst

  arXiv · 2026-07-29

  This paper reports findings from a structured participatory foresight study in which nineteen senior practitioners across official statistics, digital policy, open science, AI governance, and related fields were convened in two expert studios between 2025 and 2026. Using a qualitative signal-scanning methodology, the researchers identified seven convergent signals shaping data governance, including strain on open-data paradigms, the rise of machine-centric data ecosystems, inference reshaping governance foundations, infrastructure sustainability challenges, institutional fragmentation, sovereignty-driven strategic control, and the need for stronger data-sharing incentives. The study argues that data governance is becoming inseparable from AI governance, digital public infrastructure, economic strategy, democratic resilience, and geopolitical competition. The contribution is explicitly diagnostic rather than predictive, offering an evidence-informed framework for reasoning about structural shifts already underway to enable anticipatory governance before risks and dependencies become locked in.

  • AI policy

• Research

  OptimismBench: Forecasting Bias and the Alignment Effect in Language Model Judgment

  Seonglae Cho, Adriano Koshiyama

  arXiv · 2026-07-29

  OptimismBench is a benchmarking framework that detects directional probability bias in large language models (LLMs) by presenting each scenario in both a success-framed and failure-framed version, then measuring the asymmetry between the two responses as a signed bias score. Testing 16 models from 8 providers, the study finds that 14 are systematically optimistic, that post-training (alignment) is the key driver of which direction a model tilts, and that model identity accounts for 4.7x more variance in bias than language does. The findings matter because LLMs are increasingly used as decision aids, and if their probability judgments carry a hidden optimistic tilt inherited from alignment fine-tuning, downstream pipelines and enterprise decision-making processes will silently absorb that distortion. The authors release 3,870 items across 10 languages to support per-model directional-bias auditing.

  • Quality assurance
  • Enterprise

• Research

  Human diversity fuels collective creativity that large language models cannot simulate or sustain

  Mengchen Dong, Hiromu Yakura

  arXiv · 2026-07-29

  This preregistered experiment tested whether AI assistance homogenizes human creative output and whether AI-simulated personas can replace real human diversity. Using native (L1) and non-native (L2) English writers in a creative metaphor task, the study found that AI ideation compressed collective creative diversity for all writers and eliminated the advantage L2 writers normally contribute, while AI refinement preserved that diversity. Simulated writer pools built from real participant backgrounds across multiple model families consistently fell below every human pool in collective diversity, and attempts to force more diversity from models produced only degenerate text. The findings show that human linguistic and cultural diversity is a creative resource current AI cannot replicate, and that how human-AI workflows are designed determines whether that diversity survives.

  • Workforce
  • AI policy

• Research

  Hearsay: Vision-Language Medical Diagnoses Without an Image

  Siddharth Vohra

  arXiv · 2026-07-29

  This paper investigates what happens when frontier vision-language models (Claude Opus-4.7, GPT-5.4, and Gemini-3.1-Pro) are prompted for a medical image diagnosis when no image is actually provided. Rather than abstaining, the models confabulate structured diagnoses that are systematically shaped by patient demographic descriptors — for example, consistently returning Melanoma for a 65-year-old white man and Sarcoidosis for a young Black patient on a chest X-ray. The study further reveals a 'hedged regime' where prose output acknowledges the missing image while the structured diagnosis field still names a disease, a failure invisible to prose-only audits. The authors conclude that clinical deployment of vision-language models requires direct auditing of structured output channels and that sensitivity to probe wording must be treated as a core evaluation dimension.

  • Quality assurance
  • AI policy

• Research

  Flat Score, Amplified Failures: How the Error Budget Masks Damage in Quantized LLM Agents

  Jiwon Jang, Kisu Yang, Heuiseok Lim et al.

  arXiv · 2026-07-29

  This paper challenges the conventional wisdom that 4-bit post-training quantization is nearly lossless for large language models (LLMs), specifically testing this claim for multi-turn, tool-calling agents on the τ²-bench benchmark. While standard task-reward scores show no statistically significant change across 16-, 8-, and 4-bit weight precisions, a closer look at the failure process reveals that quantization amplifies pre-existing errors—particularly tool-name hallucination—by up to 2.5× in volume (+17.6 points per task), without introducing new failure types. The flat score is an artifact of the benchmark's ten-error budget absorbing the extra failures; shrinking that budget to two errors re-exposes a 17-point score gap precisely where quantization added error volume. The authors propose two diagnostics—per-channel error rates and success under a shrinking error budget—derived from logs benchmarks already collect, and recommend reporting these alongside standard task reward.

  • Quality assurance
  • Enterprise

• Research

  A First Look at Coding Agents' Compliance with AI Contribution Rules in Open-Source Communities

  Wenhao Yang, Runzhi He, Minghui Zhou

  arXiv · 2026-07-29

  This paper investigates whether AI coding agents follow the AI contribution rules that open-source communities have established, such as total bans, mandatory disclosure requirements, verification gates, and human sign-off requirements. The researchers built RepoComplianceBench from 106 issues across 49 repositories and evaluated four frontier models, finding that agents almost never proactively retrieve contribution rules on their own. With reminder prompts, rule quotes, or verifier feedback, agents can be nudged toward disclosure and verification compliance, but no tested condition caused an agent to refuse contributing to an AI-banned repository. The findings highlight that disclosure and verification issues are addressable with existing mechanisms, while enforcing outright bans and human escalation requirements remains an unsolved problem—raising significant questions for open-source governance and AI policy.

  • AI policy
  • Quality assurance

• Research

  SecRespond: Benchmarking AI Agents for Real-World Post-Compromise Incident Response

  Lehan Wang, Boli Chen, Ruixue Ding et al.

  arXiv · 2026-07-29

  SecRespond introduces the first benchmark specifically designed to evaluate large language model (LLM) agents in post-compromise incident response scenarios, filling a gap left by existing benchmarks that focus only on pre-attack (pre-compromise) settings. The benchmark spans 10 cyber ranges built from real compromised cloud hosts, covering 21 ATT&CK techniques and 5 operating systems, and requires agents to produce forensic reports and remediation plans from disk snapshots and security alerts. Testing 23 frontier LLMs reveals that while agents can handle alert-driven findings, they consistently fail to proactively investigate disks for silent intrusions and to generate comprehensive remediation plans—with no model achieving complete detection and remediation on any single range. This exposes a fundamental bottleneck in deploying AI agents for real-world security operations and provides a public benchmark to drive future progress.

  • Quality assurance
  • Enterprise

• Research

  Evidence-Ledger Adjudication for Claim-Evidence Traceability

  Gengyu Chen, Yongjie Yu, Weiling Wang

  arXiv · 2026-07-29

  This paper introduces 'evidence-ledger adjudication,' a workflow designed to verify whether AI-generated claims are actually supported by the evidence cited with them. Each claim is paired with an evidence packet and assigned a support relation (supported, contradicted, missing, or mixed), with unsupported claims routed back to the author. Testing on a 2,335-row benchmark drawn from AVeriTeC, CLIMATE-FEVER, and SciFact, the agent-based evidence-ledger system achieves 0.676 relation accuracy and 0.601 macro-F1, substantially outperforming the best non-agent baseline (0.383 accuracy, 0.303 macro-F1). The approach creates an auditable traceability layer for AI-assisted writing, helping catch cases where AI-drafted claims are not genuinely supported by their cited sources.

  • Quality assurance

• Research

  Collusion with Competitive Marginals: Price-Level Audits Are Blind by Construction

  Xin Xu, Chengrui Wu, Jiayu Lu et al.

  arXiv · 2026-07-29

  This paper demonstrates a fundamental flaw in empirical methods used to detect algorithmic collusion: conspiring agents can coordinate through the joint distribution of their bid components while keeping each individual agent's bid distribution indistinguishable from competitive behavior, rendering single-agent price-level audits blind by construction rather than merely underpowered. The authors validate this theoretically and empirically using language-model bidding agents and Ethereum block-building auction data, showing that residual correlations between co-deployed models exist but are undetectable by standard marginal-price screens. Because lawful multi-identity operation and actual conspiracy are behaviorally indistinguishable to existing detectors, the paper argues the tractable regulatory target is market structure counting—resolving 40 bidding identities into 23 operators raises the Herfindahl index by 247.5%, with behavioral clustering pushing that figure to 324.5%. These findings have direct implications for how regulators and auditors design oversight of AI-driven markets.

  • AI policy
  • Quality assurance

• Research

  (Im)Paired Programming: Coding Agents Improve Productivity but Harm Understanding

  Nishant Balepur, Connor Baumler, Valerie Chen et al.

  arXiv · 2026-07-29

  This study examines how AI coding agents (like Cursor) affect developer understanding compared to chatbot-based coding assistance, using a controlled experiment with 54 students building websites. Results show that while coding agents improve initial task completion speed, they significantly harm users' code comprehension and leave them unable to extend their own code without AI help. Low-effort interaction patterns—such as copy-paste prompts and auto-accepted edits—are especially linked to poor understanding, and users prefer agents anyway because of convenience despite acknowledging weaker comprehension. The findings highlight a tension between productivity gains and meaningful human oversight, with implications for how coding agent developers should design for active engagement and learning.

  • Workforce
  • Quality assurance

• Research

  Open Security Benchmark: Towards Autonomous Enterprise Cyber Defense

  Gal Engelberg, Michael Arenzon, Leon Goldberg

  arXiv (Cornell University) · 2026-07-29

  The paper introduces Open Security Benchmark (OSB), a framework for evaluating agentic AI systems performing autonomous enterprise cyber defense tasks such as security posture investigation. The authors identify an 'environment data gap'—the lack of shared, queryable, realistic enterprise environments needed to rigorously assess whether AI agents can be trusted for security work—and address it by providing a frozen, holistic synthetic enterprise environment with closed-form ground truth answers. OSB supports evaluation via text-to-SQL queries and native vendor APIs, uses multi-dimensional scoring, and is instantiated with identity-security packs and synthetic organization datasets at multiple scales. This matters because it enables reproducible, trustworthy benchmarking of AI agents before they are deployed to make real security decisions in enterprise environments.

  • Enterprise
  • Quality assurance

• Research

  Artificial Intelligence (AI), Audit Quality, and the Future of Professional Judgment: Policy and Governance Challenges in Auditing - A Systematic Literature Review

  Geoffrey Odoch

  International Journal of Computer Information Systems and Industrial Management Applications · 2026-07-29

  This systematic literature review examines how AI integration into auditing is reshaping audit quality and professional judgment, identifying major policy and governance challenges facing the profession. The review finds a core tension between automating audit tasks and preserving professional skepticism, while highlighting risks including algorithmic bias, lack of transparency in AI systems, and regulatory lag. Key policy gaps identified include unresolved liability issues, eroding professional identity, new quality assurance demands, and absent standardization frameworks. The authors call for coordinated action from regulators, standard-setters, firms, and educators to develop governance models and audit methodologies that integrate human and machine intelligence.

  • AI policy
  • Quality assurance

• Research

  Lottery Tickets Are Not Deployment Tickets

  Bum Jun Kim

  arXiv (Cornell University) · 2026-07-29

  This paper investigates whether sparse or compressed neural networks (lottery tickets) can directly replace dense incumbent models in deployed systems without reconfiguring downstream decision logic. Across extensive experiments, the authors find that while sparse models often match clean accuracy, they remain behaviorally different from their dense counterparts across calibration, out-of-distribution response, class-level reliability, and representations. Critically, in fixed-threshold policy settings, swapping in a lottery ticket changed 7–10% of accept–review decisions—exactly the kind of churn that drop-in replacement is meant to avoid. The paper concludes that clean-accuracy matching is insufficient for deployment certification, and that behavioral compatibility with a fixed incumbent is a distinct and necessary requirement.

  • Quality assurance
  • Certifications

• Research

  Can We Trust AI in 6G? Verifiable and Auditable AI-Driven Trustworthy Wireless Networks

  Genze Jiang, Yizhou Huang, Kezhi Wang

  arXiv (Cornell University) · 2026-07-29

  This paper addresses a critical trust problem in AI-driven wireless networks: there is currently no way to verify that AI functions—such as those handling cell selection and mobility management in 6G—are making decisions for the right reasons rather than exploiting unreliable shortcuts. The authors propose a 'mechanical auditing' approach that inspects AI systems' internal representations and checks them against machine-verifiable 3GPP specifications through a three-step principle: locating protocol-relevant features, verifying their causal role, and diagnosing how adaptation changes their use. They introduce an audit-native network architecture featuring a dedicated verification agent that supports both pre-deployment certification and runtime auditing. The work also identifies open challenges that must be resolved before mechanistic auditing can enter telecommunications standardisation practice.

  • Certifications
  • Quality assurance

• Research

  Assurance-Scoped Reliability for Agentic Networks: Capturing the State That Matters

  Bilgehan Erman, Andrea Francini, Nikos Papadis

  arXiv (Cornell University) · 2026-07-29

  This paper identifies failure modes in agentic AI networks—such as acting on stale information, repeating external actions, or quietly relaxing policy enforcement—that conventional reliability measures miss, even when a service appears healthy. The authors propose Reliability Assurance Intelligence (RAI), an architecture that derives a per-service reliability profile specifying what must be checked, recorded, recovered, and audited, then retains durable state in a 'context capsule' for recovery and accountability at runtime. The work matters because it addresses accountability gaps in autonomous, cross-domain AI systems where failures may be invisible to standard monitoring. A methodology for validating RAI's reliability assurances is also proposed, using an agentic lifecycle manager for deterministic network services as a running example.

  • Quality assurance
  • Certifications

• Research

  SANDBOX REGULATÓRIO BRASILEIRO COMO LABORATÓRIO CONSTITUCIONAL: LIMITES ESTRUTURAIS DOS MODELOS DO BANCO CENTRAL DO BRASIL E DA COMISSÃO DE VALORES MOBILIÁRIOS DIANTE DE SISTEMAS DE INTELIGÊNCIA ARTIFICIAL E MACHINE LEARNING

  Eduarda Hoffmann, Juliano Heinen

  Veredas do Direito Direito Ambiental e Desenvolvimento Sustentável · 2026-07-29

  This Brazilian legal study examines whether the regulatory sandboxes operated by Brazil's Central Bank (BCB) and Securities Commission (CVM) are constitutionally adequate for overseeing artificial intelligence and machine learning systems. The paper identifies four structural disconnects between the existing regulatory framework and AI/ML realities—temporal, ontological, accountability, and territorial gaps—drawing comparisons with the EU AI Act and the UK Financial Conduct Authority's sandbox model. The authors conclude that the current Brazilian sandbox framework is constitutionally insufficient for AI systems and propose five minimum legitimacy safeguards, including formal legislative reservation, prior fundamental-rights impact assessment, mandatory civil society participation, full disclosure of results, and an independent review body. The findings carry direct implications for how AI regulatory policy should be designed and legitimized under constitutional constraints.

  • AI policy

• Research

  DESAFÍOS JURÍDICOS DE LA INTELIGENCIA ARTIFICIAL EN EL MERCADO LABORAL MEXICANO: ANÁLISIS JURÍDICO-CRÍTICO SOBRE DERECHOS FUNDAMENTALES DEL TRABAJADOR

  Liliana C. Becerra-Vargas, Teresa de Jesús Vargas Vega

  KAIRÓS REVISTA DE CIENCIAS ECONÓMICAS JURÍDICAS Y ADMINISTRATIVAS · 2026-07-29

  This legal analysis argues that Mexico's current legal framework is insufficient to protect workers' fundamental rights in the face of algorithmic management and AI-driven labor practices. The article identifies gaps in the Federal Labor Law and the personal data protection statute (LFPDPPP), particularly regarding dignified work, data privacy, and non-discrimination. Through doctrinal, normative, and comparative analysis, the authors propose reforms to labor legislation and the creation of algorithmic audit mechanisms.

  • Workforce
  • AI policy

• Research

  AI application in commerce and services: Digital skills training in SMEs

  Trieu Thai Duong, Huynh Thanh Danh, Đỗ Đăng Trình

  Journal of Science and Development Economics · 2026-07-29

  This study assesses digital transformation and AI adoption in human resource management and workforce training among 89 small and medium-sized enterprises (SMEs) in Vietnam's Mekong Delta commerce and service sectors. Using mixed methods including surveys and interviews with business leaders, managers, and employees, the findings reveal a significant capability gap between leaders and employees in AI application, with employees expressing high demand for AI-integrated digital skills training that current programs have not fully met. The paper highlights collaboration between universities and SMEs as a key mechanism for co-designing training programs to build a digitally skilled workforce.

  • Workforce
  • Enterprise

• Research

  Model-based, in-situ, non-destructive qualification and certification of parts made by autonomous additive manufacturing

  Dayalan Gunasegaram, T. DebRoy, Paul Greenway et al.

  Journal of Physics Materials · 2026-07-29

  This paper proposes an integrated framework combining model-based qualification and certification (MBQ&C) with autonomous additive manufacturing (AAM) to address the productivity bottlenecks of traditional post-build inspection and testing for 3D-printed parts. The framework uses high-fidelity machine learning and reduced-order physics models within the Integrated Computational Materials Engineering paradigm to simulate process-structure-property-performance relationships, enabling build-specific fitness assessments using in-situ sensor data rather than generic parameters. Key claimed benefits include faster certification decisions, performance-based defect classification, and reduced reliance on costly post-build computed tomography scanning and destructive testing. The authors argue this approach is especially valuable for high-consequence and mission-critical applications where experimental testing environments are hazardous or impractical.

  • Certifications
  • Quality assurance

← Back to the dashboard