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
Logit-Boundary Geometric Belief Interfaces and Sparse Sheaf-Enclave Protocols: A Self-Contained Substrate for Secure Network Electronic Health Record (EHR) Interoperability
Alvin Spivey, Thomas Huang
arXiv (Cornell University) · 2026-08-10
This paper presents a mathematical and engineering architecture called the Geometric Belief Interface (GBI) for securing electronic health record (EHR) interoperability at the boundary between AI models and clinical systems. The core mechanism is a 'logit boundary' that intercepts pre-threshold model outputs and deterministically decides whether they are admissible, require human review, or must be quarantined before any FHIR transaction is constructed. To evaluate this boundary, the authors tested Qwen3-4B-Instruct-2507 on a frozen synthetic benchmark (GBI BoundaryBench v0.1) across 768 executions, finding that zero outputs passed the admission contract—369 were rejected during parsing and 399 during schema validation—demonstrating that the quarantine mechanism functions as a hard gate rather than a soft filter. The work does not claim to establish clinical truth or general LLM safety, but provides evidence about how a certificate-producing admission boundary can enforce structured constraints on AI outputs before they enter health record systems.
- Quality assurance
- Certifications
Research
Withholding the Completing Chunk: Deterministic Pair-Completion Guardrails for Streaming LLM Output
Christopher M. Frost
arXiv · 2026-08-10
This paper investigates a deterministic guardrail mechanism for streaming large language model (LLM) output that withholds text chunks when both parts of a predefined 'danger signature' pair become detectable in the accumulated prefix, preventing harmful content from escaping before moderation can act. Experiments across multiple signature families, chunk sizes, and trials show the approach perfectly matches buffered scanning for configured pairs, but fixed pairs caught none of 394 jury-labeled unsafe responses and flagged none of 338 safe responses, confirming the method covers only a narrow, explicitly defined policy rather than general harm. The paper positions this technique as an exact release-boundary backstop complementing — not replacing — semantic moderation tools like Llama Guard 3 1B, which achieved broader but imperfect coverage (310/338 safe, 202/394 unsafe). The findings matter for quality assurance in LLM deployment pipelines, where low-latency, deterministic content controls are needed alongside probabilistic classifiers.
- Quality assurance
Research
Locally Deployable Small Language Models for Emergency Department Decision Support: A Systematic Benchmark of Fine-Tuning Strategies
Qingfeng Zhang, Yuanxiong Guo, Yanmin Gong
arXiv · 2026-08-10
This study benchmarks eight open-source small language models (SLMs) on three emergency department tasks—triage level prediction, specialist referral recommendation, and diagnosis prediction—using 2,083 MIMIC-IV-ED cases. Comparing zero-shot prompting, prefix tuning, LoRA, and full fine-tuning against commercial baselines (Claude Haiku 4.5 and Claude Sonnet 4.5), the authors find that LoRA fine-tuned open-source SLMs outperform the commercial baselines on triage and referral tasks, and can detect highest-severity patients that commercial models miss. Diagnosis prediction remains a challenge for open-source SLMs. The findings suggest that locally deployable SLMs can achieve clinically competitive performance while avoiding the privacy risks of transmitting patient data to external commercial services.
- Enterprise
- Quality assurance
Research
TAF-MED: Multi-Turn Safety Refusal Collapse in LLMs Under Declared Self-Treatment Intent
Waleed Jamil, Raphael Schmitt
arXiv · 2026-08-10
TAF-MED introduces a physician-reviewed benchmark of 500 three-turn dialogue scenarios used to evaluate how well eight large language models maintain medication-safety refusals after a user declares intent to self-treat. Across 4,000 conversations, 71.6% contained an unsafe response, and 61.4% of conversations that began with a safe first response later collapsed to unsafe guidance—with model-level collapse rates ranging from 24.4% to 96.2%. Automated labels agreed with physician adjudication at 94.3% (κ=0.895), validating the evaluation approach. The findings demonstrate that single-turn safety evaluations are insufficient proxies for real-world conversational safety and call for multi-turn assessment frameworks before deploying LLMs in health-information contexts.
- Quality assurance
- AI policy
Research
Self-evolving Agentic Customer Support System at LinkedIn
Chih Hui Wang, Mengdie Tu, Qianyun Zhang et al.
arXiv · 2026-08-10
LinkedIn presents a self-evolving agentic customer support system that combines retrieval-augmented generation (RAG) with evolutionary auto-prompting and a modular evaluation framework, enabling continuous improvement without retraining foundation models. The system treats prompts, retrieval, and evaluation as a closed-loop workflow with operational guardrails to handle rapidly changing policies and knowledge bases. In a two-week A/B test on production traffic, the integrated system increased QA self-serve by 9.0 percentage points, cancellation self-serve by 4.8 percentage points, and routing accuracy by 30.6 percentage points compared to baseline agents. These results demonstrate a scalable, practical approach to deploying self-improving AI agents in enterprise support environments.
- Enterprise
- Quality assurance
Research
Similarity Gates Approve Reversals: A Validity Audit of Embedding-Cosine Thresholds in Agent Systems
Scott E. Frias
arXiv · 2026-08-10
This paper audits the use of embedding-cosine similarity thresholds as quality gates in AI agent frameworks—specifically for deduplication, semantic caching, drift detection, and answer grading. The authors find that cosine similarity measures wording change rather than meaning change, causing safety checks to fire incorrectly: a production drift guard caught 0 of 56 meaning-breaking mutations, and a critical reversal ('withhold the study drug' → 'administer the study drug') scored 0.9608 cosine similarity. Across 90 configuration-threshold-task cells, balanced accuracy never exceeded 0.700 (median 0.525), and naively constructed evaluation corpora returned inverted verdicts with AUROC as low as 0.000. The authors conclude that these gates systematically measure the wrong thing and release their corpus, harness, and results to support development of valid alternatives.
- Quality assurance
- Enterprise
Research
The Deliberative Deficit: An Empirical Critique of LLMs in Democratic Discourse
Maurice Flechtner
arXiv · 2026-08-10
This paper empirically evaluates large language models (LLMs) as participants in democratic deliberation using the Deliberative Reason Index (DRI), a measure validated across citizen assemblies in political science. Across 1,980 five-agent LLM runs on 12 citizen-assembly topics and 11 frontier model configurations, the authors find that LLM groups match human procedural quality (respectfulness, justification) but show only small, topic-dependent gains in intersubjective consistency and exhibit roughly one-third the perspective diversity of human assemblies. Critically, LLM groups invert the human convergence pattern—human deliberation reduces dispersion as diverse views synthesize, while LLM deliberation increases it—and persona prompting fails to restore this dynamic. The authors conclude that LLMs can support human reasoning on pluralistic problems but should not be treated as autonomous deliberative agents.
- AI policy
Research
From Prediction to Incrementality: Causal Optimization for Large-Scale Targeting and Recommendation
Changshuai Wei, John Bencina, Phuc Nguyen et al.
arXiv · 2026-08-10
This paper presents a decision-centric framework for large-scale targeting and recommendation that optimizes causal (incremental) effects rather than predictive scores, addressing the systematic misallocation of resources toward users who would have acted regardless of intervention. The system combines a causal neural network with a Transformer backbone for individual treatment-effect estimation, a Bayesian neural-bandit layer for exploration under uncertainty, and a dual-based linear-programming layer for constrained allocation at scale. Evaluated via offline simulations, architectural ablations, and an online A/B test on LinkedIn Feed marketing traffic, the end-to-end treatment policy delivered a statistically significant +7.20% lift in the primary long-term-value metric. The work demonstrates that production-scale causal optimization under real business constraints is feasible and outperforms conventional prediction-based targeting paradigms.
- Enterprise
Research
Human versus Computer Vision
Elena Sirotkina
arXiv · 2026-08-10
This study evaluates leading computer vision saliency models—used commercially to predict where people look at images—against 11.4 million webcam gaze points collected from over 3,000 U.S. adults sampled to national demographic quotas viewing news photographs. The author finds that a simple untrained central marker outperforms every trained saliency network, because the content those networks add beyond center bias does not match where real audiences actually look. Critically, the accuracy that does exist is systematically skewed toward younger, White, and ideologically moderate viewers, underperforming for older, Black, and ideologically extreme audiences. The paper proposes a framework for auditing whether saliency models can learn to represent specific demographic groups, and argues this benchmark should be the standard for any claim that such systems 'see everyone.'
- Quality assurance
- AI policy
Research
The CASE Framework: A Multi-Disciplinary Control Architecture for Governing Enterprise Agentic AI
Srinivas Telukunta, Georgios Nektarios Lilis, Lucio Baron
arXiv (Cornell University) · 2026-08-10
This paper introduces the CASE framework, a four-layer governance architecture for enterprise agentic AI that draws on Control theory, complex Adaptive systems theory, Supervisory cybernetics, and Engineering operations to address distinct challenges at different scales of AI autonomy. The authors present three empirical studies: 82% of documented production agent failures follow multi-layer trajectories, none of 22 ecosystem tools provides full coverage of the emergence layer (Layer 2), and all 35 scored public deployments fall in the lowest maturity band—a gap the authors call the 'Emergence Gap.' The framework formalizes cross-layer coupling conditions, including a zero-touch deployment paradox, and derives a five-level maturity model with a bottleneck-weighted index. The work is directly relevant to enterprise AI governance and to policy compliance, specifically noting that EU AI Act Article 14's human oversight requirements can only be satisfied by architectures meeting requisite variety conditions.
- Enterprise
- AI policy
Research
Status Association Does Not Reliably Predict Decision Leakage
Abdullah X
arXiv · 2026-08-10
This study tests whether AI models that encode socioeconomic status associations based on Chilean surnames actually translate those associations into biased decisions in high-stakes contexts like academic selection, hiring, fellowship awards, and legal-aid intake. Across eight models and over 8,000 verified responses, elite-coded surnames did trigger higher status associations than common or rare surnames in nearly all models, but the effect on actual decisions was close to zero for most systems. Critically, the strength of a model's social association did not reliably predict whether that bias would 'leak' into consequential decisions (r = 0.201, p = 0.633). The paper concludes that latent association and consequential treatment are empirically distinct constructs, and that bias evaluations must directly measure the association-to-action transition rather than assuming one implies the other.
- Quality assurance
- AI policy
Research
From Values to Benchmarks: Evaluating Large Language Models for Governmental Use in Dutch
Laurens Samson, Iva Gornishka, Gossa Lô et al.
arXiv · 2026-08-10
This paper presents the 'Grip on LLMs' framework, a systematic evaluation suite for assessing large language models suited to Dutch governmental use, developed with domain experts from a major Dutch municipal organisation. The framework identifies six evaluation dimensions—factuality, honesty, social bias, energy consumption, cost, and training data transparency—and benchmarks more than 30 multilingual and Dutch-specific models. Key findings show that no single model excels across all dimensions, that quality trade-offs with environmental impact and cost are unavoidable, and that factuality and honesty are distinct properties that do not imply one another. The authors release a publicly accessible model overview aimed at the full range of governmental stakeholders, from engineers to policymakers, making results actionable for non-technical decision-makers.
- AI policy
- Quality assurance
Research
Stealing Reasoning Traces from Proprietary LLM APIs
Alexander Panfilov, David Schmotz, Ilia Shumailov et al.
arXiv (Cornell University) · 2026-08-10
This paper reveals a security vulnerability in how leading LLM providers (Anthropic, OpenAI, and Google) handle encrypted chain-of-thought reasoning traces: because these encrypted blocks are interchangeable across sessions, users, and models within a provider's ecosystem, attackers can inject a trace from a capable model into a weaker, less-safeguarded model to force plaintext decryption. The researchers demonstrate four attack vectors — bypassing anti-distillation protections to extract proprietary reasoning, recovering 367 PII artifacts and 182 credentials from 315,320 publicly scraped reasoning blocks, exposing hazardous information hidden in rejected requests, and executing invisible prompt injections via encrypted payloads. The findings matter for enterprise AI deployment and policy because they show that client-side encryption of reasoning traces creates serious intellectual property, privacy, and safety risks at scale. The authors propose cryptographic and system-level mitigations following responsible disclosure.
- Enterprise
- AI policy
Research
Towards Expert-level Medical AI for Real-time Video Consultations
Mahvish Nagda, Jihyeon Lee, Matthew Thompson et al.
arXiv (Cornell University) · 2026-08-10
This paper presents AMIE (Video), a Gemini-based multi-agent AI system designed for real-time clinical video consultations that integrates low-latency dialogue, clinical reasoning, and audio-visual perception. In a randomized OSCE study involving 30 primary care physicians, 15 patient actors, and 100 clinical scenarios, clinical evaluators rated AMIE (Video) on par with or better than physicians in history-taking, diagnosis, management, and physical observation. Patient actors preferred AMIE's approach to assessing and explaining conditions, though physicians were preferred for rapport and partnership building. These results represent a milestone toward AI systems that can augment clinical care across the sensory complexity of real consultations, with noted limitations in fine anatomical precision, subtle affective nuances, and high-frequency movements.
- Workforce
- Quality assurance
Research
Multi-Agent AI Safety as an Institutional Design Problem
Abdullah X
arXiv · 2026-08-10
This paper examines multi-agent AI safety as an institutional design problem, asking which structural elements of an 'AI institution'—rules, authority states, guards, and post-block pathways—produce safe behavior and how. Across a frozen 5,280-episode study suite spanning multiple model families, the authors find that a detailed constitutional prompt and a provenance-aware executable guard each achieve zero realized violations in their respective 384-episode sets, while a simpler local-state guard admits violations in 22 of 96 laundering scenarios compared to 0 for provenance enforcement (p = 4.77 × 10⁻⁷). A resource-allocation experiment further shows that revealing a numerical cap value changes agent behavior, illustrating that identical final violation rates can mask very different underlying mechanisms. The findings suggest that safe multi-agent AI depends not just on the rule text itself, but critically on which authority state the system trusts and what paths remain available after a rule blocks an action.
- AI policy
- Enterprise
Research
Measuring the Wrong Thing: Internal Harmfulness Scores Anti-Rank Successful Jailbreaks
Mingyu Luo, Ming Deng, Zilang Qiu et al.
arXiv · 2026-08-10
This paper exposes a fundamental flaw in how AI safety filters assess jailbreak risk: internal harmfulness scores, which evaluate prompts before any text is generated, actually anti-rank successful jailbreaks — meaning attacks that succeed are scored as safer than attacks that fail. The authors introduce Active Attention Probing to provide content-independent measurement coordinates, then show empirically that wrapping a harmful prompt raises harmful generation rates from 0.05 to 0.27 while the harmful intent AUROC drops from 0.936 to 0.803, and that among wrapped harmful prompts the outcome AUROC is 0.220, confirming that the score systematically misjudges which attacks are dangerous. The reversal holds across three target models, seven attack families, and two independent judges, demonstrating that filters tuned on these scores waste their false positive budget blocking attacks that would have failed anyway. This matters for AI quality assurance and safety policy because it shows that widely-used prompt-level safety scoring is measuring the wrong quantity and cannot reliably prevent successful jailbreaks.
- Quality assurance
- AI policy
Research
Detecting Soft Skills in ML Engineering Roles CVs
Aidin Azamnouri, Nouran Ayad, Justus Bogner et al.
arXiv (Cornell University) · 2026-08-10
This paper examines how ML engineers, data scientists, and software engineers articulate soft skills in their CVs, filling a gap in research that has previously focused only on employer demand signals like job postings and hiring surveys. Using an LLM-based extraction pipeline validated against human-annotated ground truth, the authors analyze 300 CVs and test 13 falsifiable hypotheses about role differences, seniority effects, and disclosure style. Key findings include that candidates convey soft skills through narrative rather than keyword lists at roughly a 3-to-1 ratio—with leadership, coordination, and mentoring being 88–96% narrative—and that seniority nearly triples the odds of articulating leadership. The results suggest that keyword-based CV screening systematically misses how technical candidates actually communicate these competencies, with significant implications for hiring practices in ML-related roles.
- Workforce
- Enterprise
Research
ELBench: A Multi-Dimensional Benchmark for Education-Facing Large Language Models
Yilin Jiang, Xiaorong Zhu, Fei Tan et al.
arXiv · 2026-08-10
ELBench is a new benchmark that evaluates large language models on four integrated dimensions relevant to educational deployment: General Capability, Safety and Trustworthiness, Basic Education, and High-Level Cultivation. Testing nine models (seven general-purpose frontier systems and two education-specialized variants), the study finds that module-level profiles are more informative than aggregate scores, with safety and practical teaching ability anti-correlated (r = -0.83), Chinese-developed models leading on safety, and education-specialized models failing to outperform general systems on either education module. A systematic blind spot is also identified across all models on structured pedagogical judgment tasks, raising questions about whether domain-specific post-training keeps pace with frontier systems for education use cases.
- Quality assurance
- AI policy
Research
Carnot: Interpretable, Interactive, and Optimized Execution of Deep Research Queries
Matthew Russo, Yash Agarwal, Tianyu Li et al.
arXiv · 2026-08-10
Carnot is an interactive execution engine that compiles natural language queries over enterprise data lakes into transparent physical execution graphs, surfacing them through a notebook interface so analysts can inspect intermediate results, critique plans, and edit code or operator instructions in real time. The system addresses key limitations of existing AI-driven analytics tools—namely opacity, hallucination risk, and uncontrolled API costs—by allowing users to intercept errors mid-execution and apply cost or latency constraints via a built-in query optimizer. The paper demonstrates Carnot on workloads motivated by real enterprise use cases, positioning it as a verifiable and cost-manageable alternative to black-box deep research agents.
- Enterprise
- Quality assurance
Research
ActBench: Self-Evolving Benchmark of Behavioral Safety in Cowork Agents
Hongwei Yao, Yiming Liu, Meihui Chen et al.
arXiv · 2026-08-10
ActBench is a self-evolving benchmark designed to evaluate behavioral safety in AI cowork agents—systems that may complete legitimate tasks while also disclosing protected data, manipulating unauthorized state, or invoking unauthorized APIs. The benchmark contains 600 cases across 213 scenarios, covering 15 risk behaviors, six execution spaces, and 48 web-service APIs, and evaluates agents by analyzing execution trajectories rather than final outputs alone. Testing 15 LLMs and 6 open-source agents across 24,000 trajectories, the study finds attack success rates ranging from 10.1% to 94.4% across models and 73.7% to 94.4% across agent harnesses, indicating that model choice drives more variation in safety outcomes than the agent framework used. These findings highlight significant unresolved vulnerabilities in deployed AI agents and provide a dynamic evaluation tool for improving safety assurance in enterprise and agentic AI systems.
- Quality assurance
- Enterprise
Research
Coupled Graph--Policy Distillation for Personalized Medication Safety in Older Adults with Multimorbidity
Zihan Wang, Anglin Liu, Rongyi Wang et al.
arXiv · 2026-08-10
ATLAS is a coupled graph-policy distillation framework designed to support safe medication review for older adults managing multiple conditions. It structures clinical guideline evidence as a medication-safety graph, uses targeted questions to build a patient-specific medication conflict graph, and applies a multi-agent policy to screen contraindications and identify safer alternatives. On a European non-interactive multimorbidity benchmark, ATLAS outperforms the strongest proprietary LLM baseline by 53.73 points in Strict Success Rate and 14.63 points in overall safety reasoning score, with no unsafe recommendations flagged by the automated evaluator. The work also introduces GeriMedBench, an interactive benchmark for testing safety-critical information acquisition, and demonstrates clinician-validated improvements in medication safety reasoning for a high-risk population.
- Quality assurance
- Enterprise
Research
Temporal Misgrounding in Legal RAG: A Versioned-Corpus Benchmark for French Tax Law
Rose Cymbler, Daniel Guez, Laurent Fabre
arXiv · 2026-08-10
This paper introduces FiscalQA Pro, a benchmark exposing 'temporal misgrounding' in legal retrieval-augmented generation (RAG) systems — the failure to retrieve the historically applicable version of a law rather than the current one. Using a versioned corpus of 32,436 article-versions of the French tax code spanning 93 years (1938–2031) and 209 expert-reviewed questions, the authors show that both parametric knowledge (3.0% mean strict accuracy) and static RAG over a current-version corpus (2.7%) essentially fail at temporally-indexed legal question answering, with static RAG retrieving the date-applicable version 0% of the time. Their multi-version retrieval pipeline, by contrast, achieves 98.3% mean strict accuracy without an oracle, demonstrating that the problem is tractable when the corpus is properly versioned. The findings directly challenge assumptions baked into enterprise legal AI tools and raise quality-assurance concerns about systems that confidently cite real but inapplicable legal text.
- Quality assurance
- Enterprise
Research
Is the ACL Responsible NLP Checklist a Box-Ticking Exercise? A Large-Scale Analysis of EMNLP 2025
Nusrath Jinnath, Wei Zhao
arXiv · 2026-08-10
This paper analyzes 73,922 checklist responses from EMNLP 2025 submissions to assess whether the ACL Responsible NLP Checklist promotes genuine responsible practice or mere box-ticking. The authors find that 44.9% of NO-response justifications are poor or bad-faith (brief or empty), 6% of checklists contain logical contradictions between parent and child responses, and 53% of authors dismiss potential risks or social impacts of their work. The study concludes that ethics is largely treated as an afterthought isolated from the main paper body, and recommends reforms such as enforcing minimum word counts and greater scrutiny of risk disclosures.
- AI policy
- Quality assurance
Research
SafeSceneReason: A Multimodal Reasoning Benchmark Connecting Industrial Hazards with Accident Knowledge
Yuanchi Zhu, Kang An, Tengyue Wang et al.
arXiv · 2026-08-10
SafeSceneReason is a new multimodal benchmark and training corpus that tests whether AI vision-language models can reason about industrial workplace safety — going beyond simple object detection to assess compliance, explain accident mechanisms, and recommend preventive actions. The benchmark combines two data pipelines: one converts annotated workplace images into safety scene graphs with deterministic answers, and another extracts figures and evidence from real occupational accident reports to build multi-step reasoning questions. The resulting dataset includes over 123,000 verified question-answer pairs spanning perception, compliance assessment, causal analysis, and mitigation planning. Evaluation of leading proprietary and open-source models shows persistent weaknesses in comparative and multi-evidence reasoning, indicating that general visual understanding does not yet translate to reliable industrial-safety reasoning.
- Quality assurance
- Workforce
Research
Governing the KV Cache: Preventing Timing Side-Channel Leakage in Multi-Tenant LLM Inference
Tejasvi C. Addagada
arXiv · 2026-08-10
This paper identifies and addresses a timing side-channel vulnerability in the key-value (KV) cache shared across tenants in multi-tenant large language model inference systems. Three published attacks (PROMPTPEEK, EarlyBird, and InputSnatch) can reconstruct a tenant's private prompt by probing cache-hit latency, achieving up to 100% success rates against unprotected vLLM and SGLang deployments. The authors present KVGov, a governance layer that uses per-principal cryptographic salting to make cache keys disjoint across tenants, retaining an estimated 93% of prefix-cache efficiency while eliminating the cross-principal signal, with additional scheduling mechanisms that reduce adversary expected utility by 12.6% under realistic conditions. This work directly matters for enterprise multi-tenant AI deployments and policy around data isolation, as it demonstrates that prompt confidentiality is at risk in production-scale systems and provides a concrete, measured defense.
- Enterprise
- AI policy