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
A corrective agentic hybrid RAG and an operations-grounded evaluation for a scientific facility
Rajat Sainju, Dariusz Jarosz, Hairong Shang et al.
arXiv · 2026-07-27
This paper presents APS-RAG, a deployed retrieval-augmented generation (RAG) platform built for the Advanced Photon Source scientific facility that makes decades of institutional knowledge—logbooks, technical documents, wikis, maintenance records, and live control data—accessible via natural-language queries. The system combines dense, sparse, and knowledge-graph retrieval channels with a corrective agentic loop, and is evaluated on APS-Bench, a 50-question benchmark with auditable gold answers. All RAG variants outperform a naive BM25 baseline (63.8% strict vital-nugget recall), with the full corrective Agentic GraphRAG reaching 70.3%; notably, removing the cross-encoder reranker alone drops strict recall by 32.8%. The authors release the benchmark methodology, evaluation harness, and codebase to support adoption at other large scientific facilities, demonstrating a transferable workflow for trustworthy AI assistance in facility operations.
- Enterprise
- Quality assurance
Research
Harm is not Universal: Community-Specific Toxicity Detection is Urgently Needed
Xinnuo Xu, Anja Thieme, Daniela Massiceti et al.
arXiv · 2026-07-27
This paper challenges the assumption that a single universal toxicity detector can adequately protect all users of text-to-image (T2I) generation systems. The authors provide empirical evidence that roughly 35% of images labeled safe by current detectors are considered harmful by disability communities, using a dataset of 2,400 annotated T2I-generated images covering dwarfism and blind/low vision communities. They show that both large vision-language models and general-purpose detectors perform catastrophically in zero-shot settings (F1 scores of 0.32 and 0.37, below random guessing), while prompt-based adaptation (ICL, VQA) and parameter-efficient fine-tuning offer meaningful but insufficient improvements, with the best results still far below the F1 ≈ 0.9 achieved for general-purpose toxicity detection. The findings make a strong case that community-specific toxicity detection is both necessary and feasible but requires sustained research investment to meet acceptable safety standards.
- Quality assurance
- AI policy
Research
Looping Is Not Reliability: State-Bound Evidence and Typed Revision Contracts for Agentic Code Repair
Xueping Gao, Jianwei Yang, Qiang Yang
arXiv · 2026-07-27
This paper investigates whether generate-test-revise loops in AI coding agents actually guarantee reliability, finding that repeated revision does not ensure a correct patch is retained or submitted. In a controlled study of 900 three-revision trajectories over 30 HumanEval repairs, current correctness drops from 0.820 after one revision to 0.673 after two under forced revision, even as the rate of ever-correct solutions rises to 0.847. The authors identify that stale execution traces significantly harm correct-start cases (34/135 vs. 4/135 with current traces, a 22.2-point increase), and propose a typed loop contract that binds verifier evidence to exact code states, preserves verified checkpoints, and emits auditable admission receipts. The work matters for quality assurance of agentic AI systems, showing that reliability requires formal evidence-binding mechanisms, not just iteration.
- Quality assurance
- Certifications
Research
Evaluating the Impact of Explainable AI on Trust in AI-Assisted Code Review
Zhenhan Gao, Marvin Muñoz Barón, Umm-e Habiba et al.
arXiv · 2026-07-27
This paper investigates how Explainable AI (XAI) affects developer trust in LLM-based automated code review through a within-subjects user study with 34 participants. Three conditions were compared: full explanations, review feedback only, and no explanations. Full explanations produced the highest perceived trust (M = 3.99/5), while moderate explanations achieved the highest agreement rate (89.22%), suggesting that more detailed explanations prompt developers to scrutinize AI recommendations more critically. The findings offer practical guidance for designing trustworthy AI-assisted code review systems and highlight important human factors considerations for software development teams adopting AI tools.
- Quality assurance
- Workforce
Research
The Visual Bottleneck: Sparse-Frame Adaptation of MLLMs for Joint Spatial-Temporal Video Grounding
Jiameng Zhang, Srikanth Madikeri
arXiv · 2026-07-27
This paper investigates how multimodal large language models (MLLMs) perform video grounding—localizing when and where events occur—when constrained to sparse inputs of 8–16 frames, as required by large-scale video moderation systems. The authors find a severe performance collapse under sparse-frame conditions (e.g., Qwen3-VL 8B drops from 56.0% to 22.3% temporal mIoU at 16 frames) and show that fine-tuning only the final three ViT layers (4% of parameters) recovers and surpasses dense zero-shot performance, reaching 68.8% temporal mIoU. A boundary-aware sampling strategy called Hybrid16 further improves temporal mIoU by 26 points over uniform sampling. The key finding is that training strategy dominates model scale: a fine-tuned 2B model consistently outperforms a zero-shot 8B model, with practical implications for scalable, cost-efficient video content moderation.
- Enterprise
- Quality assurance
Research
TRACE-CTI: Auditable Post-Extraction Governance of TTP Claims with Knowledge Graphs
Federico Valletta, Giacomo Longo, Enrico Russo et al.
arXiv · 2026-07-27
TRACE-CTI is a post-extraction governance framework for managing Cyber Threat Intelligence (CTI) claims mapped to the MITRE ATT def framework using knowledge graphs. It preserves extraction provenance, versioned trust decisions, and non-destructive revocation history, and is evaluated on two public CTI corpora comprising 65 reports and 5,303 sentences across a controlled matrix of retriever and generator configurations. The paper shows that requiring unanimous agreement across six extraction setups raises gold-aligned precision from 25.3% to 90.6% at the cost of recall dropping from 88.2% to 16.3%, and that the resulting graph can directly answer seven governance questions—about provenance, trust, versioning, and disagreement—that flat output formats cannot address without reprocessing. This matters for quality assurance in security operations, where auditable, policy-compliant validation of automated TTP mappings is essential but currently lacking.
- Quality assurance
- AI policy
Research
Hierarchical Group-Conditional Conformal Risk Control for Selective Prediction in Language Models
Murilo Salem, Luísa Böhm, Daniel Pontes et al.
arXiv · 2026-07-27
This paper introduces HG-CRC (Hierarchical Group-Conditional Conformal Risk Control), a post-hoc calibration framework that extends standard conformal risk control (CRC) to enforce simultaneous risk guarantees across all nodes of a user-defined group hierarchy in large language models performing selective prediction with abstention. The motivation is that standard CRC provides only marginal population-level guarantees, which can systematically over-expose subgroups to errors — violating the risk budget in up to 47% of trials under mild group composition shift. Evaluated on three LLMs (Qwen3-4B, Llama-3.1-8B-Instruct, Gemma-3-4B) across two benchmarks (ARC Challenge, MMLU-Pro) and eight configurations, HG-CRC achieves 0% empirical violation rate on ARC Challenge for high-accuracy models, with the trade-off of 22–37 percentage-point higher abstention rates compared to standard CRC. The framework requires only a held-out calibration set and no retraining, making it practically deployable for serving heterogeneous user populations.
- Quality assurance
Research
Early Detection of Distributed Backdoors in Multi-Agent LLM Systems: A Characterization Study
Diego Fernandez Arias, Dev Prashant Mistry, Ren Wang et al.
arXiv · 2026-07-27
This paper investigates a novel distributed backdoor attack against multi-agent LLM systems in which encrypted payload fragments are spread across multiple agents and reassembled externally after the run, evading per-step safety checks that examine each action in isolation. The authors build a working attack instance on a hierarchical multi-agent system, testing it across five language models and two task domains, and find that a prefix detector can flag 99.3% of successful attacks with a median of five steps remaining before payload execution, providing enough time to abort nearly every attack. However, the study also reveals that effective detectors rely heavily on removable surface cues—chiefly ciphertext length and entropy—and that once these cues are stripped, detection degrades significantly and transfers poorly across domains. The findings highlight fundamental challenges in securing multi-agent AI pipelines against distributed, structured attacks that evade localized inspection.
- Quality assurance
- AI policy
Research
Task-Conditional Faithfulness Auditing of Multimodal LLMs for Grid Diagnosis
Tianqiao Zhao, Meng Yue, Jianhui Wang
arXiv · 2026-07-27
This paper proposes a framework for auditing whether multimodal large language models (LLMs) actually use the right evidence when diagnosing power grid faults, rather than just producing correct-sounding answers. The framework compares what a model claims to rely on, how its behavior changes when specific data modalities are removed, and what engineering knowledge says should matter for each task. When mismatches are detected, an evidence-gated correction mechanism regenerates responses under proper evidence constraints and re-audits them to confirm improved grounding without sacrificing accuracy. Case studies on IEEE 39- and 118-bus grid scenarios across three differently scaled LLMs validate the framework's ability to detect, diagnose, and correct these faithfulness failures.
- Quality assurance
Research
Bigger or Cheaper? Scale and Quantization Effects on Uncertainty Signals in Vision-Language Models Under Image Degradation
M M Asif Ferdous
arXiv · 2026-07-27
This paper investigates how model scale and 4-bit quantization affect the reliability of confidence signals in vision-language models (VLMs) when processing degraded images. Testing the Qwen2-VL family across 5,700 predictions under six photographic degradations at three severities, the researchers find that scaling from 2B to 7B parameters sharply improves the model's internal uncertainty signal (error-detection AUROC rising from 0.80 to 0.98), while verbalized confidence remains weak (AUROC 0.61–0.69). Critically, 4-bit quantization costs very little in accuracy (-1.6 points) but significantly degrades the internal confidence signal (AUROC drops from 0.95 to 0.80) and collapses the verbalized-confidence parse rate from 99% to 64%. For practitioners with a fixed memory budget, the paper recommends choosing a larger quantized model (7B-4bit) over a smaller full-precision one, as it delivers the best accuracy and uncertainty signal among configurations tested.
- Enterprise
- Quality assurance
Research
Regulating for AI Legitimacy
Gilad Abiri
arXiv · 2026-07-27
This legal article argues that AI governance frameworks focused on 'alignment'—ensuring AI systems pursue the right objectives safely—fail to address a prior question of political legitimacy: by what authority are those objectives set and enforced? The author defines legitimacy as sociological, meaning the belief among those subject to AI-governed decisions that the power is exercised rightfully, and contends that good performance alone cannot produce that belief, as demonstrated by the legitimacy crises surrounding social media and search despite their measurable benefits. The article identifies three sites where AI legitimacy falters—opacity, private exercise of public-facing authority, and administrative automation—and proposes three regulatory principles: integration (embedding AI rule-setting in already-authoritative venues), familiarity (presenting rules in locally credible forms), and contestation (guaranteeing meaningful review and real remedies). The work matters for AI policy because it reframes regulation away from purely technical compliance toward democratic authorization and public accountability.
- AI policy
Research
GAUGE: Grading Agent-Built Financial Models Without a Golden Answer
Jiacheng Lu, Sinuo Wang, Wentao Zhao et al.
arXiv · 2026-07-27
GAUGE is a new benchmark for evaluating AI agents that build financial valuation models, designed to overcome a core flaw in existing benchmarks: grading against a single expert reference answer when professionals themselves disagree substantially. The authors show that across 108 pairs of independently built analyst models covering 65 companies, median agreement scores are only 0.33 and no same-vintage pair agrees on implied price within 10%, meaning point-tolerance grading unfairly penalizes legitimate professional disagreement. GAUGE instead grades against an observed-practice envelope derived from 1,001 analyst workbooks across 196 tasks, validated with a 55-participant human study spanning students, junior analysts, and senior analysts. The best AI agent scores 53.4 on the failure-aware metric—above finance students (43.2) but below junior analysts (66.0) and all senior analysts (88.3)—and is notably stronger at mechanical model construction than at valuation judgment, where a 26-point fleet-median gap persists.
- Enterprise
- Quality assurance
Research
Closed-Loop Validation-Repair for Healthcare Interoperability: A Multi-Model Study of Schema Compliance in Clinical LLMs
Jianru Shen
arXiv · 2026-07-27
This study tests whether a closed-loop validation-repair framework can bring open-source clinical language models (Qwen2.5 7B, Llama 3.1 8B, and Gemma2 9B) into compliance with healthcare interoperability standards such as ICD-10, CPT, and HL7 FHIR. Across 960 model-scenario pairs covering ten medical specialties, baseline schema compliance ranged from 85.9 to 91.6 percent, with most failures being representation-level format violations rather than clinical reasoning errors. Applying the validation-repair loop raised overall compliance to 99.0 percent, with statistically significant improvements of 7.8 to 12.5 percentage points confirmed by McNemar tests (p < 0.001). The findings suggest that closed-loop validation-repair is a practical system-level safeguard for integrating AI into electronic health record pipelines.
- Quality assurance
- Enterprise
Research
Retrieval-Augmented Large Language Models as Components of Cognitive Computing architecture for Regulatory Knowledge Management
Dariusz Nowak-Nova
arXiv · 2026-07-27
This paper investigates whether combining large language models (LLMs) with Retrieval-Augmented Generation (RAG) can transform them into reliable components of cognitive computing infrastructure for regulatory knowledge management. The study proposes and validates a locally deployed, on-premises architecture—using Ollama and LM Studio with Polish-language models Bielik and PLLuM on consumer-class hardware—that pairs LLMs with external knowledge repositories for controlled, traceable information retrieval. Results show that RAG integration significantly improves factual consistency, domain specificity, and normative precision while enabling auditability and dynamic updating of regulatory content without retraining. The authors conclude that such systems should be viewed as semantic processing modules capable of supporting regulatory compliance and organizational decision-making in legally volatile environments.
- AI policy
- Enterprise
Research
DeepFaith: Evidence-Grounded LLMs for Faithful Incident Reporting in Multi-Stage APT Defense
Trung V. Phan, Tri Gia Nguyen, Thomas Bauschert
arXiv · 2026-07-27
DeepFaith is a framework that uses large language models (LLMs) to generate faithful, evidence-grounded incident reports for multi-stage Advanced Persistent Threat (APT) cyberattacks. By integrating unified evidence representation, evidence-grounded prompting, faithfulness-aware generation, and post-generation verification, the system converts outputs from autonomous defense modules into natural-language reports explicitly tied to underlying system evidence. Tested in a realistic enterprise testbed, DeepFaith improved faithfulness scores from 0.68 to 0.92, reduced unsupported claims from 0.32 to 0.08, and increased temporal consistency from 0.6 to 0.88, outperforming both template-based and existing LLM-based solutions. These results demonstrate that evidence-grounded LLM generation can produce reliable and actionable reports for security operations centers.
- Enterprise
- Quality assurance
Research
Beyond Aggregate Risk: Role-Stratified Conformal Risk Control for LLM Tool Calls
Md Ashikur Rahman, Md Arifur Rahman, Niamul Hassan Samin et al.
arXiv (Cornell University) · 2026-07-27
This paper addresses the risk that AI language-model agents can be manipulated through structured tool calls, where dangerous inputs in high-stakes argument fields (e.g., recipients, credentials) may be masked by benign inputs elsewhere when risk is evaluated only in aggregate. The authors introduce role-stratified conformal risk control, a calibration method that assigns separate risk thresholds and budgets to different semantic argument roles within tool calls, providing finite-sample statistical guarantees per role rather than per whole action. Tested across AgentDojo and InjecAgent benchmarks with six language models, the method achieves more consistent role-specific budget compliance than aggregate approaches under a variety of challenging conditions including model transfer, adaptive attacks, and distribution shift. The results argue that certification of AI agent actions should be performed at the level of individual semantic roles, not entire tool calls.
- Quality assurance
- Certifications
Research
The Tokenizer Tax: Quantifying and Explaining the Cross-Lingual Cost of Subword Tokenization for Indian Languages
Priyansh Srivastava
arXiv · 2026-07-27
This paper measures how much more expensive it is—in token usage—to process Indian-language text compared to English text when using popular subword tokenizers. Using the FLORES-200 parallel corpus across 14 languages and 6 tokenizers, the authors find that GPT-3.5/GPT-4's tokenizer imposes an average 8x token penalty on Indian languages (up to 13x for Malayalam), shrinking the effective context window to as little as 12% of what English users receive for equivalent content. The root cause is identified as failed byte-pair merges, which strongly correlate with the penalty (Pearson r = 0.89), but multilingual tokenizers can reduce the disparity by 73%, showing it is a design choice rather than an inherent limitation of Indic scripts. This matters for enterprise and policy discussions because it means Indian-language users face higher costs, reduced context capacity, and degraded content preservation compared to English users of the same LLM systems.
- Enterprise
- AI policy
Research
Epistemic Norms for AI Safety and Alignment Research
Keivan Navaie
arXiv (Cornell University) · 2026-07-27
This paper argues that AI safety and alignment research operates under fundamentally different standards than mainstream AI research, requiring epistemic norms focused on bounding worst-case outcomes and demonstrating the absence of hazardous behaviors rather than optimizing average-case performance. The authors identify five key gaps in current alignment research practices—including a near-absence of institutionalized independent verification—through a preregistered bibliometric synthesis. To address these gaps, they propose ECAISA (Epistemic Code for AI Safety and Alignment), an eight-principle framework with scoring rubrics, disclosure ladders, and anti-gaming mechanisms aimed at improving how safety-relevant research claims are documented, checked, and relied upon. The framework targets auditability rather than certification, making it directly relevant to governance and quality-assurance processes for AI safety research.
- AI policy
- Quality assurance
Research
Success Is Not Self-Explanatory: Auditing Success Provenance in Agent Evaluation
Jingkun Luo, Da-Tian Peng
arXiv · 2026-07-27
This paper identifies a gap in how AI agents are evaluated: a correct answer does not reveal whether the agent succeeded through intended reasoning or by exploiting information acquired during the evaluation process itself. The authors introduce AcquaBench, an auditing framework that uses three matched conditions—CLEAN (benchmark-authorized info), GOLD (correct target available), and SHAM (incorrect but structurally matched value)—to test whether agent success genuinely depends on correct-target availability versus mere exposure to a source. Key findings show that in one dataset (D0), GOLD exceeds SHAM by 19.1 to 25.9 percentage points, confirming success tracks the correct value, while in another (D2) behavioral dependence persists even when a localization marker fails, and an apparent 5.0-point model score gap collapses to -0.6 points under GOLD conditions. The paper argues that agent benchmarks should report not just scores but whether the evaluated information state actually supported the observed success, calling this missing property 'success provenance.'
- Quality assurance
- Certifications
Research
Beyond Local Inspection: Global, Guideline-Grounded Evaluation of Post-hoc XAI Methods for ECG Classification
Nils Gumpfer, Michael Guckert, Samuel Sossalla et al.
arXiv · 2026-07-27
This paper evaluates whether post-hoc explainable AI (XAI) methods reliably identify clinically relevant patterns in ECG classification, rather than merely following signal amplitude. The authors introduce a global, guideline-grounded framework that aggregates explanations across heartbeats and benchmarks them against clinically defined regions of interest derived from ECG guidelines, testing 13 gradient-based XAI methods on four binary classifiers trained on the PTB-XL dataset. Results reveal that methods transferred from computer vision frequently track signal amplitude instead of diagnostic relevance—mean Spearman correlations up to 0.69—causing them to miss critical low-amplitude regions such as the ST segment, where LRP-ε assigns only 4.6% of relevance compared to 63.8% for LRP-SIGN, with 9 of 13 methods falling below chance for at least one condition. These findings highlight that sample-level heatmaps can mask systematic explanation failures, underscoring the need for domain-grounded, global evaluation before deploying XAI tools in clinical settings.
- Quality assurance
- Certifications
Research
The Half-Lives of Generative-AI Evidence: A 40-Record Audit, a Claim-Currency Framework, and a Reflexive Case of Frontier-Model-Assisted Research
Carlo Iacono
arXiv · 2026-07-27
This paper audits 40 empirical records on generative-AI evaluations published between July 2025 and July 2026, finding that the newest named model cited was a median 281 days old at publication, with journal articles citing models a median of 395 days old versus 56 days for preprints. The study distinguishes 'model age' from 'claim currency,' noting that 35 of 40 records included a superseded model family and only 7 supplied a precise dated model identifier, and proposes six reporting practices to make AI-assisted research more inspectable. It also serves as a reflexive case study of its own two-day AI-assisted production process using GPT-5.6 Sol Pro, demonstrating how rapid frontier-model-assisted research can be documented transparently. The findings matter for quality assurance in AI research, highlighting how publication pipelines risk rendering empirical evaluations outdated before they reach readers.
- Quality assurance
- AI policy
Research
Randomness in large language models: What researchers need to know (and report)
Guillaume Coqueret, Joan Llull, Florian Oswald et al.
arXiv (Cornell University) · 2026-07-27
This paper investigates how outputs from large language models (LLMs) vary across repeated requests even when prompts and settings are held constant, due to sources such as deliberate sampling, silent model updates, numerical rounding, and expert routing. Using sentiment classifications of corporate filings as an illustrative case, the authors show that this variability has meaningful downstream consequences for regression results. They argue that LLM outputs should be treated as draws from a distribution rather than fixed measurements, and propose a reporting standard for researchers, data editors, and replication packages to address reproducibility challenges. The findings are directly relevant to how LLM-generated data is validated and reported in research workflows.
- Quality assurance
- AI policy
Research
Generative Artificial Intelligence in Scientific Research: Individual Benefits, Collective Risks, and a Framework for Responsible Research with AI
Fulvio Castellacci, Tommaso Ciarli, Yuan Gao et al.
arXiv (Cornell University) · 2026-07-27
This paper examines the tension between individual productivity gains from generative AI in scientific research and broader systemic risks, drawing on an academic roundtable and a growing empirical literature. While AI-assisted research shows measurable gains in publication volume and citation share, evidence on novelty, disruption, and breakthrough output is ambiguous or negative. The authors identify three mechanisms driving divergence between private and social returns—information asymmetry, negative externalities on a shared knowledge base, and depletion of research capacity—and propose a 'Responsible Research with AI' (RRAI) framework built on four principles: disclosure, differentiation, narrative, and proportionality. RRAI is designed to integrate with existing governance structures such as the EU AI Act, UNESCO, and the OECD to preserve AI's productivity benefits while managing systemic risks.
- AI policy
- Quality assurance
Research
EXE-Bench: Ranking the Tradeoffs of AI-based Windows Malware Detectors for Real-World Usability
Andrea Ponte, Daniel Gibert, Matouš Kozák et al.
arXiv (Cornell University) · 2026-07-27
EXE-Bench introduces a comprehensive benchmark for evaluating AI-based Windows malware detectors across four dimensions: predictive performance, temporal robustness, adversarial robustness, and computational overhead. The benchmark aggregates these into a single score to enable fair model comparison, revealing that post-deployment-only evaluations give an incomplete picture. A key finding is that feature-engineered models outperform deep learning approaches in withstanding both the passage of time and adversarial attacks, whereas deep networks tend to degrade significantly after initial deployment.
- Quality assurance
- Certifications
Research
Recursive Governance: A Graph-Theoretic Framework for Risk Propagation and Drift Detection in Agentic AI Systems
Sriram Nagaraj, Advaith Nila Narayanan
arXiv (Cornell University) · 2026-07-27
This paper proposes a graph-theoretic governance framework for managing risk in autonomous agentic AI systems used by financial institutions. It introduces four key contributions: a calibrated Degree of Autonomy materiality score, a Directed Acyclic Graph-based risk propagation algorithm that penalizes downstream agents when upstream validations fail, a trajectory monitoring protocol using cosine drift across Chain-of-Thought embeddings to detect reasoning drift, and practical solutions for complications like LLM version changes and latent feedback loops. The framework is designed to replace static model inventory practices with a dynamic 'Inventory-as-Code' governance loop suited to continuously evolving agentic systems. This work directly addresses emerging gaps in model risk management as financial institutions adopt increasingly autonomous AI.
- AI policy
- Enterprise