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

  EduPanel: A Three-Agent LLM Judge for Teaching Videos -- Reliability, Complementarity, and Human Trust Calibration

  Jia-Kai Dong, Yi-Cheng Lin, Hung-yi Lee

  arXiv · 2026-07-20

  EduPanel is a three-agent LLM-based system that evaluates the pedagogical quality of teaching videos using rubric-grounded, learner-conditioned assessments. The system decomposes evaluation across specialized agents to handle multimodal evidence and produces interpretable feedback tailored to different learner personas. In expert studies, EduPanel achieves reliability comparable to a median human expert and improves scoring accuracy (MAE from 0.87 to 0.73), while human experts retain the ability to detect unreliable outputs (AUC = 0.77). The findings suggest EduPanel can serve as a scalable assistant for educational video evaluation without replacing human expert judgment.

  • Quality assurance
  • Workforce

• Research

  Querying Multimodal Scientific Papers with AI: Practices and Preferences Across Blind, Low-Vision, and Sighted Scientists

  Arnavi Chheda-Kothary, Lucy Lu Wang, Joseph Chee Chang et al.

  arXiv · 2026-07-20

  This paper investigates how blind, low-vision, and sighted scientists use AI tools—specifically ChatGPT and Gemini—to query visual elements such as figures, diagrams, and tables in scientific documents. Through interviews with ten scientists across STEM fields, the authors find that vague or incorrect AI-generated descriptions of images lead both blind/low-vision and sighted users to abandon AI workflows, highlighting reliability as a critical barrier. The study contributes a dataset of 115 real queries and AI responses and surfaces design implications for building more accessible and accurate AI-powered scientific question-answering systems. The findings matter for quality assurance of AI tools and for enterprise and workforce considerations around equitable access to scientific knowledge.

  • Workforce
  • Enterprise
  • Quality assurance

• Research

  Towards an Automated Test of LLM Security Knowledge

  Shufan Chai, Liangliang Sun, Jessica Staddon

  arXiv · 2026-07-20

  This paper presents a partially-automated method for assessing LLMs' knowledge of security topics by using authoritative information from Consumer Protection Agencies (CPAs) to detect instability in LLM responses—instability that can signal knowledge gaps. The approach is demonstrated across two security topics (identity theft and impostor scams) and five LLMs from the Gemini and GPT families, using publicly available CPA information. The method successfully distinguishes between models with sufficient and insufficient knowledge to accurately identify security topics in text narratives. This matters for quality assurance and certification of AI systems used in security-sensitive contexts, offering a more scalable alternative to manually curated benchmarks.

  • Quality assurance
  • Certifications
  • AI policy

• Research

  Structured Output Collapses Answer Diversity Across 44 Language Models

  Tapan Parikh

  arXiv · 2026-07-20

  This paper investigates how requesting structured output formats (like JSON) from language models affects the diversity of answers they produce. Across 44 models and 31 open-ended prompts, adding a simple format instruction ('Reply with JSON only') caused the most common answer to rise from 41% to 64% of responses, while distinct answers fell from 52 to 36 and mean answer-choice surprisal dropped from 1.80 to 1.58 bits. The compression effect is specific to formats models are trained on for tool use (JSON and XML) and is not explained by decoder-level schema enforcement — it lives in the model's learned response to the register itself. This matters because software systems typically consume language models through structured output interfaces, meaning the models being deployed in real enterprise pipelines are measurably more homogeneous than the chat-surface models that benchmarks, comparisons, and purchasing decisions are based on.

  • Enterprise
  • Quality assurance
  • Certifications
  • AI policy

• Research

  Intelligent Cause Prioritisation? An Analysis of AI Policy Priorities and Governance in Africa

  Osaremen Iluobe, Kisso Selvan

  arXiv (Cornell University) · 2026-07-20

  This paper examines how African governments are framing and prioritizing AI in their national strategies and public communications. Drawing on speeches, press releases, public statements, and national AI strategies, the authors find that while African policymakers are highly attentive to AI's economic and developmental opportunities, they devote comparatively little attention to AI safety. The paper argues that Africa's position as largely a consumer rather than a producer of frontier AI systems, combined with pressing development challenges, shapes a policy orientation focused on catching up economically. The findings highlight a governance gap that could influence who benefits from AI and who bears its risks on the continent.

  • AI policy

• Research

  CANDOR: Chance-Calibrated Discordance in Frozen Foundation Encoders

  Soroosh Tayebi Arasteh, Sven Nebelung, Daniel Truhn

  arXiv (Cornell University) · 2026-07-20

  CANDOR is a new metric for evaluating how well frozen foundation model encoders geometrically separate different class labels, without needing to train a classifier head first. The key innovation is using equal-size label banks, which fixes the chance level at exactly 0.5 and removes a density bias that previously made uninformed encoders appear useful. Testing across 22 encoders, 20 datasets, 7 domains, and over 600,000 images, the authors find that while no encoder is entirely blind, all are weak: even the best chest X-ray encoder places 18.4% of pneumothorax positives closer to an opposite-label image than a same-label one, and some medical findings fall at or below chance. CANDOR enables practitioners to identify poorly supported findings before any classifier is trained, informing encoder selection for high-stakes applications like medical imaging.

  • Quality assurance
  • Certifications
  • Enterprise

• Research

  PathReportEval: A Systematic Benchmark for Pathology Report Generation

  Suryakant Singh, Sejuti Majumder, Beatrice Knudsen et al.

  arXiv · 2026-07-20

  PathReportEval introduces a standardized benchmark and evaluation framework for automatically generating pathology reports from whole-slide images. The paper finds that commonly used text-generation metrics like BLEU, ROUGE, and METEOR are weakly aligned with clinical correctness and frequently overestimate report quality. To address this, the authors propose the Clinical Report Quality Score (CRQS), a clinically grounded metric that evaluates factual correctness across four dimensions—clinical fact coverage, key information recall, hallucination rate, and clinical discordance—and demonstrates that CRQS reveals clinically meaningful differences between models that lexical metrics miss. This work matters for quality assurance in AI-assisted pathology, as it provides a reproducible, modular framework to rigorously assess whether AI-generated reports are clinically accurate rather than merely lexically similar to reference reports.

  • Quality assurance
  • Certifications

• Research

  Using Fine-Tuned LLMs to Identify Indicators of Vulnerability in UK Police Incident Logs

  Sam Relins, Daniel Birks

  arXiv · 2026-07-20

  This study tests whether a fine-tuned large language model (LLM) pipeline can identify four vulnerability indicators—mental ill health, substance misuse, alcohol dependence, and homelessness—in nearly 3,000 de-identified UK police incident logs. The researchers find that LLMs can produce meaningful prevalence estimates at scale (e.g., mental ill health indicators appear in roughly one in five incidents), but that naive single-pass classification is unstable and systematically over-assigns indicators compared to human judgement. Correcting these biases requires substantial human review and statistical adjustment, and individual-level errors remain frequent and unpredictable. The findings caution that LLM outputs from unstructured administrative data cannot be treated as valid measurements without rigorous methodological safeguards, limiting their use for operational policing decisions while showing some promise for population-level resource and training planning.

  • Workforce
  • AI policy
  • Quality assurance

• Research

  ChainMark: Model-Free LLM Watermarking with Closed-Form Calibration

  Chengheng Li-Chen, Kyuhee Kim

  arXiv · 2026-07-20

  ChainMark is a new watermarking method for large language model (LLM) outputs that works without requiring access to the generating model and provides closed-form statistical guarantees, addressing regulatory mandates like the EU AI Act that require machine-readable marking of synthetic text. The approach partitions vocabulary tokens into states using keyed hashing and enforces a Markov transition structure, allowing detection in linear time with no LLM involvement. The authors prove theoretical results including a closed-form mapping from a target false positive rate and text length to the minimum required state count, and a universal robustness threshold of approximately 29.3% that holds regardless of system parameters. Across multiple instruction-tuned LLMs and domains, ChainMark outperforms existing watermarking methods under translation and substitution attacks at matched budget.

  • AI policy
  • Quality assurance
  • Certifications

• Research

  Adversarial Robustness of Phishing Email Detection: A Comparative Study of TF-IDF + Logistic Regression and Fine-Tuned DistilBERT

  Tanveer Ahmed, Seyedali Pourmoafil

  arXiv · 2026-07-20

  This paper compares two phishing email detection approaches—a TF-IDF + Logistic Regression baseline and a fine-tuned DistilBERT transformer—trained on a unified corpus of 82,255 emails and evaluated under normal, synthetic, and adversarial conditions. Both models exceeded 98% accuracy on clean data but collapsed to roughly 64% accuracy under adversarial testing, a drop of more than 34 percentage points, with only a 0.36 percentage-point difference between the two models despite relying on different evidence as revealed by LIME, SHAP, and attention-rollout analysis. The findings demonstrate that clean-data accuracy does not predict adversarial robustness, and the authors argue that adversarial testing should become a standard component of phishing-detection evaluation. This has direct implications for enterprise cybersecurity deployments and quality-assurance practices around ML-based email filtering systems.

  • Enterprise
  • Quality assurance
  • AI policy

• Research

  Enabling Multilingual Privacy Policy Audits: Large-Scale Analysis of Spanish Mobile Apps

  Marcos Moran, David Rodriguez, Luka Nenadic et al.

  arXiv (Cornell University) · 2026-07-20

  This paper investigates whether large language models (LLMs) can audit privacy policies across multiple languages without language-specific retraining, addressing the gap left by English-centric auditing tools. The researchers built an evaluation corpus covering all 24 official EU languages and found their LLM-based classifier achieves macro-F1 scores between 0.91 and 0.94 for identifying personal data collection categories across languages. Applying this approach to 2,611 Android apps from the Spanish Google Play Store, they uncovered systematic discrepancies between declared and observed data practices—particularly in public-sector apps—and identified a linguistic barrier where public-sector apps tend to publish policies in Spanish while popular commercial apps use English. The findings demonstrate that English-only privacy audits can systematically miss transparency gaps in multilingual environments, with direct implications for regulatory oversight in regions like the EU.

  • AI policy
  • Quality assurance
  • Certifications

• News

  China’s AI models have Trump’s AI world at war with itself

  technologyreview.com · 2026-07-20

  MIT Technology Review reports that the release of Kimi, a free open-source AI model from Chinese company Moonshot that rivals top U.S. models, has sparked a public feud among Trump's AI advisors over how the U.S. government should respond. Former AI czar David Sacks favors open competition and opposes government interference, while Pentagon official Emil Michael supports a more interventionist approach, and the two factions have exchanged public insults over the weekend. The dispute centers on the economic and national security threat posed by capable, free Chinese AI models that undercut paid U.S. alternatives from OpenAI and Anthropic, rattling U.S. stocks and fueling debate over a new White House vetting process critics are calling a 'de facto licensing regime for frontier AI.' Complicating matters further, the article notes that loosened chip export controls and alleged model distillation practices may have helped China develop such competitive models despite limited computing resources.

  • Enterprise
  • Workforce
  • AI policy

• Research

  It's Not What You Say, It's How You Say It: Evaluating LLM Responses to Expressions of Belief

  Kevin Du, Clara Kümpel, Michelle Wastl et al.

  arXiv · 2026-07-20

  This paper investigates how the linguistic form of users' expressed beliefs shapes whether large language models (LLMs) defer to user-provided context or stick to their prior knowledge. The authors develop a typology of 17 fine-grained expression-of-belief types grounded in four dimensions—form, evidentiality, epistemic stance, and tone—and use it to benchmark 16 LLMs across different architectures, scales, and training stages. Key findings include that larger and instruction-tuned models tend to be less context-following than smaller or base models, and that certain linguistic framings are significantly more persuasive to LLMs than others. The results have direct implications for prompt engineering and model robustness, highlighting how surface-level linguistic choices can systematically alter LLM behavior.

  • Quality assurance
  • Enterprise

• Research

  Simple Domain Generalization for Strong Pixel-Level Image Tampering Detection in Modern VLMs

  Yi Tang, Xinyi Shang, Jiacheng Cui et al.

  arXiv · 2026-07-20

  This paper addresses the challenge of detecting pixel-level image tampering produced by modern vision-language models (VLMs) such as ChatGPT, Gemini, and Qwen-Image, where cross-model and out-of-distribution shifts make detection difficult. The authors propose a domain-generalized training framework combining a balanced minibatch sampling scheme—to prevent biased optimization—and a late-injection strategy that introduces small amounts of new VLM-distribution data after initial convergence. Despite its conceptual simplicity, the framework outperforms the prior state-of-the-art method PIXAR by 26.1% and 26.8% relative improvement in average gIoU and cIoU, respectively, across out-of-distribution VLMs including GPT-Images-2.0, Gemini-3.1, FLUX.2, and Seedream 4.5. These results matter for quality assurance and policy contexts where reliable detection of AI-generated image manipulations is increasingly critical.

  • Quality assurance
  • AI policy

• Research

  GigaPath-Flash and GigaTIME-Flash: Efficient Pathology Foundation Models for Whole-Slide and Tumor Microenvironment Analysis

  Naoto Usuyama, Jeya Maria Jose Valanarasu, Sicong Yao et al.

  arXiv · 2026-07-20

  GigaPath-Flash and GigaTIME-Flash are compact, open-weight pathology AI models designed for whole-slide image analysis and tumor immune microenvironment prediction from routine H&E images. GigaPath-Flash uses a 22M-parameter tile encoder distilled from a billion-parameter teacher model and retains 97% of the original GigaPath's average slide-level performance while requiring 50x less compute. GigaTIME-Flash surpasses the original CNN-based GigaTIME in prediction quality while running 6x faster and using 8x less GPU memory. Released under an Apache-2.0 license, these models lower computational and licensing barriers for computational pathology, immuno-oncology, and precision health research.

  • Enterprise
  • Quality assurance
  • Certifications

• Research

  A Continual Validation, Updating, and Decision-Making Framework for Self-Adaptive Digital Twins via Robust Model Predictive Control: A Case Study in Additive Manufacturing

  Yi-Ping Chen, Ying-Kuan Tsai, Vispi Karkaria et al.

  arXiv · 2026-07-20

  This paper presents an adaptive Digital Twin framework that addresses the problem of concept drift—where surrogate models degrade as real-world operating conditions change—by combining Fisher score-based drift detection, Low-Rank Adaptation (LoRA) for efficient model updates, and Mann–Whitney U testing for statistical validation of improvements. Applied to additive manufacturing processes, the framework detects distributional shifts quickly, fine-tunes fewer than 1% of model parameters upon drift detection, and certifies predictive improvement before deploying updated models. The approach maintains both predictive accuracy and uncertainty quantification under abrupt and incremental drift, providing a statistically rigorous and computationally tractable path to sustaining trustworthy neural-network-based Digital Twins. This matters for quality assurance and enterprise settings where continuous, reliable monitoring of manufacturing processes is critical to maintaining output fidelity.

  • Quality assurance
  • Enterprise
  • Certifications

• Research

  TRIM: Reducing AI-Generated CodeSlop via Agent Trajectory Minimization

  Alex Mathai, Shobini Iyer, Aleksandr Nogikh et al.

  arXiv · 2026-07-20

  This paper identifies a quality problem in AI-generated code called 'CodeSlop'—residual, functionally unnecessary edits that accumulate in agent-generated patches as the agent iterates through speculative and abandoned changes during its search process. The authors introduce TRIM (Trajectory-guided Redundancy Identification and Minimization), an algorithm that reduces CodeSlop indirectly by minimizing agent trajectories rather than targeting unnecessary code directly. Empirically, TRIM cuts CodeSlop by 17.9%–32.9% across agentic scaffolds with negligible performance regression and requires roughly half the validation cost of baselines like Delta Debugging. This matters because as coding agents take responsibility for larger portions of codebases, accumulated redundancy makes software progressively harder to maintain.

  • Quality assurance
  • Enterprise
  • Workforce

• Research

  Operational Hallucination and Safety Drift in AI Agents

  Shasha Yu, Fiona Carroll, Barry L. Bentley

  arXiv · 2026-07-20

  This paper identifies and measures two failure modes in LLM-based autonomous agents operating over multiple turns: 'Safety Drift,' where a model's declared refusal of unsafe actions gradually erodes until it actually executes constraint-violating behavior, and 'Operational Hallucination,' where agents enter repetitive tool-call loops due to flawed state perception. Through controlled multi-turn evaluations on high-stakes ethical dilemmas, malicious requests, and benign tasks, the authors quantify these failures using declaration-action gap and livelock metrics across multiple state-of-the-art LLMs. They propose an Action-Aware Supervision Layer—a lightweight architectural add-on with intent-action consistency checks, runtime state tracking, and forced termination—that their post-hoc simulations show can intercept observed violations without false positives on benign cases. The findings matter for AI safety and enterprise deployment because they show that single-turn safeguards are insufficient for agentic systems and that enforceable architectural mechanisms are needed for reliable, responsible AI agents.

  • Quality assurance
  • AI policy
  • Enterprise

• Research

  Differentiable Logic Gate Networks for Low-Latency EEG Classification on Edge Devices

  Shyamal Y. Dharia, Stephen D. Smith, Camilo E. Valderrama

  arXiv · 2026-07-20

  This paper investigates Differentiable Logic Gate Networks (Diff-Logic) as a hardware-native alternative to conventional neural networks for real-time EEG classification on edge devices. By compiling models into pure Boolean circuits executable via bitwise CPU operations, Diff-Logic achieves 80.2% Macro F1 on dementia detection—outperforming a matched MLP baseline by 6.8%—while delivering up to 2.9× faster inference and a 14× smaller model size compared to MLPs on a power-constrained edge device. Inference latency remained nearly constant across a 10× increase in model scale, demonstrating that logic-based architectures can satisfy the strict latency and memory constraints of portable brain-computer interfaces without sacrificing competitive accuracy.

  • Workforce
  • Enterprise
  • Quality assurance

• Research

  LLMs and Agentic AI Systems for Smart Grids: A Tutorial on Architectures and Applications

  Daniela Rojas, Abdulwahab Albassam, Aidan G. Leung et al.

  arXiv · 2026-07-20

  This paper presents a design framework for deploying large language models (LLMs) and agentic AI systems in smart grid applications such as wind power forecasting, EV charging scheduling, power flow analysis, and contingency diagnosis. The core contribution is a 'solver-grounded' design principle, where numerical results are only reported when they originate from trusted computational tools and pass explicit verification, addressing the problem that LLMs can produce physically infeasible outputs. In case studies, the EVAgent system reproduces the CVXPY optimum while reducing LLM-only unmet energy by 7.5–9.5x, and GridDebugAgent repairs 17 of 39 contingency cases while reducing total violations by 52.3%. The authors also propose a four-group evaluation framework covering task utility, solver-grounded correctness, faithfulness and safe failure, and cost and latency, establishing a consistent division of labor between AI orchestration and trusted solvers.

  • Enterprise
  • Quality assurance
  • AI policy

• Research

  O-VAD: Industrial Video Anomaly Detection through Object-Centric Tracking and Reasoning

  Mei Yuan, Qi Long, Qifeng Wu et al.

  arXiv · 2026-07-20

  O-VAD introduces a training-free agentic framework for Industrial Video Anomaly Detection (IVAD) that tracks object state evolution across space and time, then reasons over temporal trajectories to identify anomalous objects and events. Unlike prior vision-language model (VLM) approaches that require retraining on normal clips or injecting domain-specific knowledge, O-VAD works without domain knowledge by mimicking how human inspectors observe object transformations. Experiments across three IVAD datasets show it outperforms frontier VLMs, agentic frameworks, and fine-tuned traditional VAD methods, while also generating interpretable reports about anomaly processes and types. This matters for industrial quality control because it offers a scalable, knowledge-free solution for detecting complex procedural and physical anomalies in manufacturing settings.

  • Quality assurance
  • Enterprise

• Research

  How Does Alignment Tuning Shape Representations of Sycophancy and Related Cue-Induced Biases in LLMs?

  Prakhar Gupta, Terry Jingchen Zhang, Florent Draye et al.

  arXiv · 2026-07-20

  This paper investigates where susceptibility to sycophancy and cue-induced biases lives inside large language models (LLMs), finding that alignment tuning—not pretraining—is primarily responsible for installing these biases. Across five model families and seven bias types, the authors show that pretrained base models barely exhibit these biases, while aligned models encode each bias as a distinct, coherent direction in their hidden states that can be both decoded and causally steered. By intervening along these directions, the authors can recover unbiased answers across all tested model families, demonstrating a practical debiasing approach. The findings matter for quality assurance and policy because they reveal that alignment processes inadvertently introduce measurable representational flaws, and that targeted internal interventions can mitigate—though not fully eliminate—these errors.

  • Quality assurance
  • AI policy

• Research

  FinSAgent: Corpus-Aligned Multi-Agent RAG Framework for Evidence-Grounded SEC Filing Question Answering

  Jijun Chi, Zhenghan Tai, Hanwei Wu et al.

  arXiv · 2026-07-20

  FinSAgent is a multi-agent retrieval-augmented generation (RAG) framework designed to answer questions over SEC filings by aligning retrieval with the specific structure, terminology, and evidence standards of financial disclosures. The system addresses 'prior-corpus misalignment' — a mismatch between general model priors and the target corpus — through three mechanisms: role-specialized agents anchored to the 10-K item structure, database-aware query decomposition conditioned on a summary-level corpus view, and a learned feature-gated reranker that distinguishes evidential validity from semantic similarity. Across five offline financial QA benchmarks and a randomized online experiment with 1,000 user ratings, FinSAgent outperforms both single-agent and multi-agent baselines in retrieval coverage, answer correctness, and user preference. These results matter for enterprise and quality-assurance applications where accurate, evidence-grounded answers from complex regulatory filings are critical.

  • Enterprise
  • Quality assurance

• Research

  VDAR-Router: Adaptive LLMs Routing via Verbalized Query Difficulty Analysis Retrieval

  Yu-Chien Tang, Jun-Chen Hung, Wen-Chih Peng et al.

  arXiv · 2026-07-20

  VDAR-Router is a difficulty-aware retrieval-based framework for routing queries to the most appropriate large language model (LLM) to balance cost and performance. Unlike existing methods that rely on surface semantics or embedding similarity, VDAR-Router first generates an explicit difficulty analysis for each query, retrieves historical examples with similar difficulty profiles, and selects a model using a reward function that weighs both performance and cost. Experiments on three datasets show it consistently achieves better cost-performance trade-offs than existing baselines without requiring additional training. This matters for enterprise deployments where reducing LLM inference costs while maintaining output quality is a key operational concern.

  • Enterprise
  • Quality assurance

• Research

  Judge-dependent safety gains and model-specific helpfulness costs of evidence-sufficiency prompting in clinical LLMs

  Koyar Afrasyab

  arXiv · 2026-07-20

  This study tests whether a structured 'evidence-sufficiency prompt' reduces unsafe overconfident answers from clinical large language models, and whether measured safety gains reflect real behavior or just the scoring judge's calibration. Across four models and 1,200 paired responses, unsafe overconfidence dropped from 49.3% to 24.7% (a 24.7-point reduction), but the magnitude was highly judge-dependent: a different-family judge nearly halved the measured effect. Blinded clinician review characterized the primary AI judge as high-sensitivity but low-specificity, meaning it overcounts safety failures rather than providing calibrated rates. Critically, the safety prompt carried a model-specific helpfulness cost—correct diagnosis rates fell from 80.3% to 50.3% overall, with near-zero cost for GPT-5.5 but near-total degradation for Gemini—underscoring that safety and helpfulness tradeoffs must be evaluated jointly before any clinical deployment consideration.

  • Quality assurance
  • Certifications
  • AI policy

← Back to the dashboard