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

  SWE-Gate: Passing Functional Tests Is Not Enough for Software Engineering Agents

  Xin He, Yanlin Wang, Mingwei Liu et al.

  arXiv · 2026-09-03

  SWE-Gate introduces a new benchmark for evaluating AI software engineering agents that goes beyond simply checking whether generated code patches pass functional tests. The benchmark derives 'review constraints' from real pull request review comments and constructs 303 repository-level repair instances across 75 open-source Python repositories, each with separate functional and constraint tests. Experiments with four LLM backends show that among 644 repairs passing functional tests, 221 still fail to satisfy the associated review constraints — demonstrating that functional-only evaluation substantially overestimates agents' true capability to meet real-world software development requirements. This work highlights a critical gap in how coding agents are assessed and motivates more comprehensive evaluation frameworks.

  • Quality assurance

• Research

  SENTINEL-RL: Offloading Topological Reasoning from LLM Agents in the Security Operations Center

  Uday Vallabhaneni, Cassie L. Cagwin, David J. Wild

  arXiv · 2026-09-03

  Sentinel-RL is a security operations center (SOC) architecture that separates topological graph reasoning from language-model reasoning to overcome LLM context-window and consistency limitations at enterprise scale. A graph attention encoder compresses live authentication graphs into fixed-size states, a PPO reinforcement learning policy selects containment actions, and an LLM is restricted to generating analyst-readable narratives from those policy recommendations. Evaluated on the LANL cyber-security events dataset, the system achieves 0.91 precision and 0.87 recall on red-team events after 200 training iterations, loads a 24-million-edge graph roughly 24x faster than baseline pipelines, and completes a full detect-investigate-recommend-approve cycle in a median of 6.3 seconds. The work also includes an enterprise-readiness analysis covering false-positive economics, reversibility guarantees, audit compliance, and human-approval boundaries, making it directly relevant to deploying AI-driven security automation in large organizations.

  • Enterprise
  • Quality assurance

• Research

  Efficient Test-Time Adaptation through Human-AI Interaction

  Zora Zhiruo Wang, Apurva Gandhi, Rulin Shao et al.

  arXiv · 2026-09-03

  This paper introduces TAHI (Test-Time Adaptation through Human-Agent Interaction), a framework that personalizes AI agents to individual users by capturing iterative feedback signals across interaction sessions and integrating them into the agent's context, weights, and an evolving rubric module. Tested with 30 individuals across 600 tasks in writing and visual creation, TAHI improves solo task success by 4.5–20.9% within tens of tasks, and the evolving rubric module catches 16.0–22.3% more failures than rubrics from language models or humans alone. Personalized agents also generalize across users, yielding up to 8.8% improvement in success. This matters because it shows a concrete path to closing the gap between population-scale AI training and the heterogeneous, often unarticulated standards that individual professionals require.

  • Workforce
  • Enterprise

• Research

  The Dice Roll Method: A Standardized Protocol for Repeated-Query Auditing of Large Language Model Brand Recommendations

  Dmitrij Żatuchin

  arXiv · 2026-09-03

  This paper formalizes the 'Dice Roll Method,' a standardized statistical protocol for auditing how large language models (LLMs) make brand recommendations across repeated identical queries. Drawing on a reanalysis of roughly 190,000 observations across five prior auditing studies, the authors derive tiered guidance for iteration counts—exploratory (n=5), confirmatory (n=10), and rigorous (n=15)—each tied to generalizability and effect-size targets. The protocol is validated on three independent corpora, reproducing reliability predictions in 37 of 39 test cells. This matters for quality assurance and enterprise contexts because it gives practitioners and researchers a statistically principled, reusable framework for reliably measuring and benchmarking LLM output consistency in brand-recommendation settings.

  • Quality assurance
  • Enterprise

• Research

  FiMI Banking: A Sovereign Model for Indian Retail Banking

  NPCI AI Research Team, Aman Kumar, Asit Desai et al.

  arXiv · 2026-09-03

  FiMI Banking introduces a specialized AI benchmark and training framework for Indian retail banking, evaluating two post-training approaches—preference optimization and reinforcement learning with verifiable rewards—on a controlled dataset built from vetted banking documents and structured ground truth. Preference optimization raises out-of-scope refusal rates from 52% to 80%, while reinforcement learning improves edge-case task performance from 0.509 to 0.718 and order-sensitive task performance from 0.590 to 0.679 using 29% fewer tokens. The findings demonstrate that these two training strategies address complementary requirements for building reliable, safe, and regulation-aware conversational banking agents. This matters for enterprise and policy contexts where language models must operate within strict operational and regulatory constraints.

  • Enterprise
  • AI policy

• Research

  Beyond Majority Vote: Multi-Perspective Adjudication for Medical Hallucination Detection

  Joe Cecil, Marjorie Freedman

  arXiv · 2026-09-03

  This paper investigates how well current methods detect factual errors (hallucinations) in chatbot-generated medical text, finding that the standard single-pass, single-annotator approach systematically undercounts errors. The authors develop a multi-perspective annotation framework combining first-pass human annotation, LLM-as-a-Judge candidate discovery, and two forms of adjudication—medical expert review and evidence-based fact-checking—showing each method catches errors the others miss. Their results demonstrate that while multi-pass adjudication improves coverage and benchmark completeness, inferences remain sensitive to the judgment, expertise, and evidence used by adjudicators. These findings are directly relevant to how AI medical chatbot safety is assessed, indicating that existing hallucination benchmarks may give an overly optimistic picture of factual accuracy.

  • Quality assurance
  • AI policy

• Research

  More Criticism Does Not Make a Better Review: EquiReview-R

  Zexing Zhang, Jichao Li, Tianyang Lei et al.

  arXiv · 2026-09-03

  EquiReview-R reframes AI-assisted peer review not as generating more criticism but as evidence-guided refinement of a structured set of concerns, treating missed weaknesses (omission) and unsupported allegations (overcritique) as distinct risks requiring opposite corrections. The system resolves existing concerns against localized evidence, searches for missing issues from multiple perspectives, and returns a stop, continue, or defer decision. On a held-out cohort of unseen papers, it reduces major overcritique from 15.5% to 8.1% and achieves a one-sided omission upper bound of 9.9% while stopping early on 52.4% of papers, with ablations confirming the gains come from revision rather than extra inference. The authors also release ReviewTrace, an evidence-linked corpus for studying review revision and disagreement, which matters for quality-assurance efforts in academic publishing and AI-assisted evaluation.

  • Quality assurance

• Research

  Adapting to Evolving Requirements: Agentic AI for Retail Supply Chain Operations

  Lei Zheng, Liping Yang, Zihao Li et al.

  arXiv · 2026-09-03

  This paper addresses the challenge of adapting retail supply chain decision pipelines to evolving operational requirements using large language models (LLMs). The authors formulate the problem as jointly selecting an intervention route and module-level change within a graph-constrained agentic framework, where domain agents expose reformulation interfaces and a central processor searches over bounded intervention paths. Evaluated on 100 warehouse requirements sourced from practitioner interviews with a large retail partner, the framework outperforms direct LLM reformulation across GPT, Qwen, and DeepSeek models, raising end-to-end success rates from 72–76% to 79–83%. This work matters for retail enterprises seeking to automate and reliably update complex, coupled supply chain operations using AI.

  • Enterprise

• Research

  Evaluating Criterion-Conditioned Behaviour of Large Language Models in Content Moderation

  Danting Zhang, Bei Peng, Robert Loftin

  arXiv · 2026-09-03

  This paper investigates whether large language models (LLMs) can reliably apply individual moderation criteria when making content decisions, rather than relying on an aggregated sense of overall harmfulness. The authors introduce DECO, a criterion-independent framework for content factorisation, and a pairwise evaluation method to compare model outputs across different criteria for the same input. Testing across four moderation datasets and four LLMs, they find that strong benchmark performance can mask significant failures at the criterion level—models struggle most when correct decisions hinge on a specific content aspect rather than general harmfulness. The findings call for new evaluation methods that explicitly measure criterion-conditioned behaviour, highlighting a key gap in how current content moderation benchmarks assess LLM reliability.

  • Quality assurance

• Research

  DNative-Twin: Decision Graphs and Digital Twins for Reconstructable Agentic Decisions

  Junjie Pang, Zhenzhen Xie, Haoke Han et al.

  arXiv · 2026-09-03

  DNative-Twin introduces a graph-native digital twin framework that records AI agent decisions as typed trajectories and re-executes their decision mechanisms under controlled conditions, enabling auditability of which evidence, tools, rules, and authorization paths produced a given outcome. Using three public process logs and controlled replay experiments, the study shows that graph structure alone cannot determine consequences of unobserved tool states, but adding replay-contract state raised unresolved-divergence recall from 0 to 0.667, and adding verification results raised it to 1.0. Across 500–5,000 BPI 2020 cases, median end-to-end processing time increased from 0.794 to 8.889 seconds on the tested platform. The work directly supports enterprise auditability of agentic decisions by separating the roles of graph structure, replay context, and verification evidence in reviewing a decision mechanism.

  • Enterprise
  • Quality assurance

• Research

  IndicSafeEval: Safety Robustness of Large Language Models under Multilingual Persuasive Jailbreak Attacks

  Saikat Mondal, Mamta, Deeksha Varshney et al.

  arXiv · 2026-09-03

  IndicSafeEval is a new benchmark framework that tests the safety robustness of large language models (LLMs) against persuasion-based jailbreak attacks in four Indian languages: Hindi, Bengali, Marathi, and Punjabi. The benchmark combines ten safety-critical content categories with six persuasive strategies to generate 7,200 adversarial prompts, which are used to evaluate open-source LLMs in a systematic black-box setting. Results show that safety performance varies significantly across languages and prompt styles, with some risk categories being notably more vulnerable to persuasion-based attacks than others. The findings highlight that current safety evaluations are largely English-centric and insufficient, calling for multilingual and persuasion-aware benchmarking to better assess real-world LLM safety.

  • Quality assurance
  • AI policy

• Research

  Artificial Intelligence for Energy Optimization in Data Centers

  Mohammed Basharath Ullah, Summaiya Unnisa Begum, Mohammed Nadeem Ullah

  arXiv · 2026-09-03

  This paper conducts a systematic review of roughly 194 papers on AI-based energy optimization in data centers, ultimately coding 63 studies. The analysis finds that of 28 primary control-oriented studies, 18 are validated only in simulation and just 5 reach physical hardware or production facilities, while none account for water withdrawal or embodied carbon. Reported savings across four technique families overlap so substantially that the field cannot currently rank its own methods. Based on this corpus analysis, the authors propose CLEAR-DC, a methodological framework that couples control-policy with workload-demand through an elasticity term and tracks net rather than direct benefits, along with a standardized reporting schema covering energy, carbon, water, embodied share, and validation venue.

  • Enterprise
  • Quality assurance

• Research

  HalluPeer: A Taxonomy-driven Benchmark for Detecting Hallucinations in Scientific Peer Reviews

  Tzu-Ling Lin, Dong-Ting Yao, Teng-Fang Hsiao et al.

  arXiv · 2026-09-03

  HalluPeer introduces a benchmark specifically designed to detect hallucinations in AI-assisted scientific peer reviews. The benchmark provides aligned triples of paper content, human-written reviews, and hallucination-injected reviews, annotated for detection, classification, and localization, built from 12K papers and 38K reviews. Experiments show that existing hallucination detectors struggle to distinguish unsupported claims from legitimate critique, and the study confirms that hallucination patterns defined by HalluPeer also appear in real peer reviews. This work highlights a critical reliability gap in using LLMs as review assistants, where verifying claims requires grounding them in long, technical source papers.

  • Quality assurance

• Research

  GPS-Bench: A Governance Policy Benchmark for Automating Policy Analysis

  Linh Le, Melanie Bui, My Chiffon Nguyen et al.

  arXiv · 2026-09-03

  GPS-Bench is a benchmark designed to evaluate how well large language models (LLMs) can simulate governance policy analysis by linking policies to real actors, their responses, and downstream impacts using public records such as legislative documents, lobbying disclosures, regulatory filings, and economic data. Unlike prior approaches that rely on archetypes or unverifiable assumptions, GPS-Bench grounds every actor persona in dated, provenance-tracked evidence and compares multiple inference strategies—including multi-agent simulation, graph-based methods, and fine-tuning—in a controlled setting. The study finds that fine-tuning on grounded records yields the strongest actor-level impact predictions, while multi-agent decomposition adds interpretive value by surfacing coalition-formation mechanisms that can be checked against the documentary record. This benchmark provides a common empirical framework for assessing when evidence quality, actor modeling, and agent interaction improve the prediction and interpretation of policy outcomes.

  • AI policy

• Research

  Beyond "Made with AI": Visualizing Provenance Density to Mitigate the Transparency Penalty

  Qing Zhang, Yifei Huang, Juyoung Lee et al.

  arXiv · 2026-09-03

  This paper identifies a 'Fluency Trap' where users both over-trust fluent AI-generated hallucinations and under-trust accurate AI-generated content once disclosed as AI-made, showing that simple 'Made with AI' labels fail to address this problem. The authors propose 'Provenance Density,' an interface that visualizes the density of verified claims within a text, and test it in a user study with 81 participants. The idealized Provenance Density interface produced a large discernment gap between truth and fabrication (+4.15 points, d=1.82), while participants given no signal showed no detectable discrimination. A 200-sample technical audit further found that a 'Consistency Veto' mechanism carries most of the discriminative signal, suggesting that effective AI transparency requires moving from authorship labels toward evidence visualization.

  • AI policy
  • Quality assurance

• Research

  The Psychological Costs of Artificial Intelligence Adoption in Software Engineering

  Adam Alami, Elda Paja, Abhishek Tiwari

  arXiv · 2026-09-03

  This qualitative case study investigates the psychological toll on software professionals one year into an organizational AI adoption at a large software development services company. Through 21 semi-structured interviews, the researchers found that workers experience accountability anxiety, craft identity disruption, erosion of meaning and satisfaction, cognitive and workload intensification, and uncertainty distress. Professionals cope by restoring a sense of control, making identity-preserving adaptations, or simply absorbing costs that cannot otherwise be resolved. The findings challenge the assumption that AI adoption is cost-free and reframe it as a human transition, not merely a technological or organizational one — with direct implications for how organizations plan and manage AI integration in software engineering workflows.

  • Workforce
  • Enterprise

• Research

  Caught in the Story: Narrative Captivity in Multi-turn LLMs Conversation

  Yuhe Wu, Guangyu Wang, Yujie Chen et al.

  arXiv · 2026-09-03

  This paper introduces 'narrative captivity,' a failure mode where large language models acting as moral advisors uncritically align with a user's one-sided account during multi-turn conversations, without seeking missing perspectives. The authors build a benchmark of 5,078 interpersonal-conflict scenarios across six moral dimensions and test 17 LLMs, finding that end-state moral judgments shift by 25 percentage points on average compared to single-turn baselines. Stage-level analysis points to preference optimization as a major contributor to this bias, and four inference-time mitigation strategies provide only partial relief. The findings matter because as people increasingly rely on LLMs for everyday advice, models that cannot preserve independent judgment under one-sided narration risk reinforcing biased or incomplete moral reasoning.

  • Quality assurance
  • AI policy

• Research

  Accountable AI with Grounded, Faithful, Consistent, Actionable Rationales: A Case Study in Clinical Trial Matching with VERDICT

  Zikai Zhou, Yufei Jin, Yilin Xu et al.

  arXiv · 2026-09-03

  This paper introduces VERDICT, an LLM-based agent designed to make AI decision-making accountable by translating clinical trial matching tasks into Satisfiability Modulo Theories (SMT) formulations, so that decisions are derived by formal solvers rather than generated directly by language models. The authors identify key accountability requirements—grounded rationales, explicit assumptions, policy consistency, and pivotal conditions—and show that standard LLMs fail these requirements by applying decision policies inconsistently and producing rationales unfaithful to their own decisions. Evaluated on SIGIR 2016-derived and TREC 2021 datasets, VERDICT achieves the strongest decision accuracy among LLM-only and neurosymbolic baselines, applies policies with perfect consistency, and produces clinician-preferred rationales with improved counterfactual self-faithfulness. This work matters for high-stakes clinical and policy settings where the ability to examine, justify, and contest AI decisions is essential.

  • Quality assurance
  • AI policy

• Research

  A Comparative Legal Analysis of Deepfake Evidence in Bangladesh

  Shamsad Hossain

  Iconic Research and Engineering Journals · 2026-09-03

  This paper examines whether Bangladesh's legal framework for evaluating digital evidence is equipped to handle AI-generated deepfakes, finding that the Evidence Act of 1872—even after its 2022 amendment—lacks authentication standards and expert qualification rules specific to synthetic media. Using doctrinal and comparative methods, it benchmarks Bangladesh against the UK and Malaysia, identifying transferable elements such as Malaysia's certification mechanism and the UK's reliability inquiry. The study proposes a sequential authentication framework and legislative, institutional, and procedural reforms suited to Bangladesh's existing judicial infrastructure. The findings are relevant to ongoing legislative developments, including Bangladesh's Cyber Security Act of 2026, which may directly regulate AI-generated content.

  • AI policy
  • Certifications

• Research

  FACULTY MEMBERS’ EXPERIENCES WITH AI IN LANGUAGE TEACHING: A PEDAGOGICAL PERSPECTIVE

  SEVDA UGUR

  Advanced Education · 2026-09-03

  This qualitative study interviewed 12 faculty members at Turkish university language schools to explore how educators experience AI integration in language teaching. Findings reveal a gap between functional AI use—such as material preparation and productivity tasks—and meaningful pedagogical integration aligned with learning outcomes, assessment, and classroom interaction. Institutional constraints, workload pressures, ethical concerns, and student readiness further complicated adoption, while faculty also reported tensions around professional identity and authority in AI-mediated environments. The authors conclude that effective AI integration requires pedagogical guidance, institutional coordination, and AI literacy development for both teachers and students.

  • Workforce
  • AI policy

• Research

  IMPACT OF GENAI ON INSTRUCTORS' CREATIVITY AND EFFECTIVENESS IN WRITING INSTRUCTION

  Tetyana Bidna

  Advanced Education · 2026-09-03

  This mixed-methods study surveyed and interviewed L2 writing instructors in Canadian tertiary programs to understand how generative AI tools—primarily ChatGPT—affect their creativity and effectiveness in lesson planning and material design. Instructors reported that GenAI supported idea generation and pedagogical innovation but also raised concerns about overreliance, reduced originality, and formulaic outputs, while effectiveness gains in time management and customization were offset by prompting complexity and variable output quality. Decisions to continue or stop using GenAI were shaped by practical benefits, professional growth, and ethical or financial considerations. The findings underscore the need for AI literacy, institutional support, and critical engagement to integrate GenAI as a genuine collaborative tool rather than a replacement for instructor judgment.

  • Workforce

• Research

  Governing AI-driven digital transformation in public healthcare: assessing administrative readiness and institutional capacity in Egypt's health system

  Mohamed Abbas El-Naggar, Dina Helmy Al Nashily, Hiyam Ahmed Abdulrahim et al.

  Frontiers in Digital Health · 2026-09-03

  This mixed-methods study across 24 Egyptian governorates assessed whether public healthcare institutions have the administrative, governance, and infrastructure capacity needed for AI-driven digital health transformation. Using surveys of 387 healthcare administrators and interviews with 18 senior policymakers, the structural model explained 68% of variance in perceived digital health transformation success, with institutional capacity, administrative readiness, and organizational governance all positively associated with success. Qualitative findings revealed fragmented governance, weak infrastructure, workforce digital literacy gaps, regulatory ambiguity, and resource constraints as key barriers. The authors conclude that strengthening governance coordination, infrastructure equity, workforce capability, and AI-specific regulatory safeguards is necessary before large-scale AI deployment can deliver public value in Egypt's health system.

  • AI policy
  • Workforce

• Research

  AI-DRIVEN CHEATING DETECTION IN ONLINE EXAMINATIONS: A SYSTEMATIC REVIEW OF MULTIMODAL METHODS, ETHICAL CHALLENGES, AND FUTURE DIRECTIONS

  Beyza ESİN ÖZSEVEN, Turgut Özseven

  Advanced Education · 2026-09-03

  This systematic review analyzes 29 peer-reviewed studies (2020–2025) on AI-driven cheating detection in online examinations, covering methods such as computer vision, voice analysis, biometric authentication, behavioral/keystroke dynamics, and NLP-based AI-generated content detection. The review finds that multimodal and hybrid AI frameworks consistently outperform single-modality approaches, but real-world deployment is constrained by false-positive risks, computational costs, privacy concerns, and algorithmic bias. The authors call for responsible, scalable system design and highlight explainable AI as a mechanism to improve transparency and trust in automated proctoring.

  • Quality assurance
  • Certifications

• Research

  AI-DRIVEN CHALLENGES IN TECHNICAL HIGHER EDUCATION: AN INSTITUTIONAL PERSPECTIVE

  Anatolii Melnychenko, Tetiana Zheliaskova, О. А. Акімова

  Advanced Education · 2026-09-03

  This mixed-methods study surveyed 1,916 students and 515 faculty members at technical universities to examine how AI tools are being adopted in higher education. Findings show students primarily use AI for information retrieval, idea generation, and text generation, while complex analytical applications remain limited, and a notable adoption gap exists between students and faculty. The study identifies academic integrity concerns, declining independent learning, and the absence of clear institutional AI policies as key problems. The authors conclude that systemic institutional responses—including formal AI policies, updated teaching methods, and improved digital competencies—are needed to manage AI integration responsibly.

  • AI policy
  • Workforce

• Research

  Beyond the machine: risk, fear, optimism and the foundations of public trust in AI

  Steven Pickering, Martin Ejnar Hansen, Yosuke Sunahara

  AI & Society · 2026-09-03

  This study surveyed 3,235 respondents in Japan and the UK to identify what drives public trust in AI, finding that trust in political and scientific institutions (government, university scientists, other people) consistently predicts AI trust in both countries. Technological optimism increases trust in both contexts, while fear of AI has a stronger negative effect in the UK; notably, belief that AI will replace one's job is positively associated with trust in Japan but unrelated in the UK. The findings show that national context significantly shapes public confidence in AI, and that both psychological predispositions and institutional legitimacy are key foundations of AI trust. The authors argue this underscores the need for governance frameworks that build informed capability and institutional credibility.

  • AI policy
  • Workforce

← Back to the dashboard