Daily digest - 2026-07-17

154 new items since yesterday's digest, grouped by topic; tier 1 first within each section. One queued item was excluded as out of scope (a computational pharmacology paper with no AI content). Heavy day for domain-application journal papers; the headline news is Kimi K3, which shows up independently in two practitioner sources.

Open Models

• Kimi K3, and what we can still learn from the pelican benchmark - Named practitioner synthesis. Moonshot AI announced Kimi K3, calling it their most capable model to date at 2.8 trillion parameters, available now via their site and API with an open-weight release promised by July 27. Simon Willison runs it through his informal generative benchmarks and uses the occasion to argue what small, personal benchmarks still reveal about frontier releases. Corroborated by Latent Space's coverage the same day. link
• [AINews] Kimi K3 2.8T-A50B: the largest open model ever released; Opus 4.8-class at Sonnet 5 pricing - Named practitioner synthesis. Latent Space frames K3 as the largest open model ever released, a 2.8T-parameter mixture-of-experts with 50B active parameters, and positions its capability near Opus 4.8 at Sonnet 5 pricing. The active-parameter-to-total ratio is the mechanism to watch: sparse MoE is how open labs are buying frontier capability at servable cost. link

Multi-Agent Systems

• Multi-LLM Prototype: Open-source software for adaptive orchestration, conflict resolution, and cost-aware routing across heterogeneous large language models - Primary research, in SoftwareX. An open-source system for running several heterogeneous LLMs behind one interface, with adaptive orchestration, conflict resolution between disagreeing models, and cost-aware routing. Useful as teaching material precisely because it is published, inspectable software rather than a closed demo. link
• CoLMAgent: A collaborative agentic AI system that integrates large and small models for industrial intelligent human-machine interactions - Primary research, in Journal of Manufacturing Systems. Pairs large models with small specialized ones inside a collaborative agent system for industrial human-machine interaction, a concrete example of the big-model-plans, small-model-executes division of labor. link
• Multi-agent human-AI collaborative framework for converting hand-drawn floor plans to 3D BIM - Primary research, in Automation in Construction. A multi-agent pipeline, with humans in the loop, that turns hand-drawn floor plans into 3D building information models. A clean case study in decomposing a fuzzy perception-to-structured-output task across agents. link
• Intelligent Multi-Agent System for Research Automation - Primary research, in IJRASET. Proposes a domain-aware multi-agent system that routes user queries to specialized agents across biomedical, legal, market, academic, and programming domains, arguing that routing beats a single generalist model for research tasks. link
• AgentSearch: Indexing, Retrieval, and Ranking of AI Agents - Primary research, at SIGIR 2026. Treats agents themselves as the retrieval corpus: indexing, retrieving, and ranking agents so a system can find the right agent for a task. If agent ecosystems keep growing, agent discovery becomes its own IR problem, and this stakes out that ground. link

Agent Governance

• PromptStudio: A Governance-Aware Agentic AI Framework for Automated Prompt Generation, Optimization, Evaluation, and Lifecycle Management - Primary research, in IJRASET. Argues manual prompt design is inconsistent, hard to reproduce, and weakly governed, and builds an agentic framework that generates, optimizes, evaluates, and version-manages prompts under governance controls. Prompts treated as governed artifacts with lifecycles, not strings. link
• Runtime assurance for enterprise agentic AI systems: A policy-gated control model with quantitative autonomy-risk scoring - Primary research, in World Journal of Advanced Research and Reviews. Identifies a control gap as enterprises move from passive text generation to autonomous tool use: governance that evaluates outputs cannot control actions. Proposes policy gates plus a quantitative autonomy-risk score that throttles how much an agent may do unsupervised. link
• Designing Deterministic AI Agents in Enterprise Platforms: A Schema-Guided Reasoning Approach - Primary research, at the International Conference on Artificial Intelligence Management and Trends. Uses schema-guided reasoning to make enterprise agents behave deterministically, trading generative flexibility for predictability, the same tradeoff behind structured outputs and typed tool calls. link

Enterprise Agentic AI

• The renaissance of management: When regular employees become multi-level managers of AI agents - Primary research, in Leadership Quarterly. Argues that as agents proliferate, ordinary employees become multi-level managers of AI reports, reviving management skills as core individual-contributor competencies. Directly relevant to what an AI curriculum should teach beyond prompting. link
• Agentic AI in Enterprise Business Processes: A Systematic Review and Practitioner Survey on Adoption Readiness - Primary research, at the International Conference on Artificial Intelligence Management and Trends. Pairs a systematic review with a practitioner survey to gauge how ready enterprises actually are to put agents inside business processes. link
• The AI productivity-governance frontier: A theoretical model for enterprise value creation under agentic automation - Primary research, in International Journal of Science and Research Archive. Models why enterprise value from AI is uneven: technical capability outpaces organizational redesign, workforce adaptation, and governance maturity, so value is set by the weakest of the three. link
• Adaptive talent aligner: A large language model with dynamic hierarchical analysis and bias-corrective memory pool for personalized human resource management - Primary research, in Information Processing & Management. An LLM for personalized HR matching that adds hierarchical analysis of roles and a bias-corrective memory pool intended to keep past biased matches from compounding. link
• Leadership After AI: What Will Still Make Humans Indispensable? - Primary research, in Scholedge International Journal of Business Policy & Governance. Asks what remains of managerial work once forecasting, scheduling, evaluation, and parts of strategy are automated, and locates the durable residue in judgment and legitimacy rather than analysis. link
• From data to decisions: A narrative review of business intelligence and predictive analytics framework for enhancing SME competitiveness and economic resilience in the United States - Primary research, in Magna Scientia Advanced Research and Reviews. Reviews why US small and mid-size businesses remain underserved by advanced analytics infrastructure and sketches a framework for closing that gap. link

AI Software Engineering

• Automatic translation of natural language requirements into CTL specifications using Large Language Models: A multi-approach evaluation - Primary research, in Journal of Systems and Software. Evaluates several approaches to having LLMs translate natural-language requirements into computation tree logic specs, a bridge between informal requirements and formally checkable properties. link
• A comparative analysis of the role of Large Language Models and Low-Rank Adaptation in Automated Program Repair - Primary research, in Information and Software Technology. Compares base LLMs against LoRA-adapted ones for automated program repair, isolating how much cheap fine-tuning actually buys in a well-benchmarked SE task. link
• A Closed-Loop Agentic AI Framework for Self-Configuring, Self-Healing, and Optimized CI/CD Automation - Primary research, in IJRASET. Argues static CI/CD pipelines cannot keep up with changing repos, dependencies, and deploy targets, and closes the loop with agents that reconfigure and heal pipelines as conditions change. link
• Auto-Devops GPT: An Agentic AI Framework for Self-Healing CI/CD Pipelines Using LLM-Based Root Cause Analysis and Reinforcement Learning - Primary research, in IJRASET. A second self-healing CI/CD framework in the same run, this one combining LLM root-cause analysis of build failures with RL over remediation choices. Two independent papers on agentic CI/CD in one day is a signal about where DevOps research is heading. link
• A Conceptual Framework for the Evolution of Application Development: From Manual Coding to Agentic Development - Primary research, in International Journal of Science and Research Archive. Periodizes application development from manual coding through low-code to agentic development and proposes a framework for reasoning about the transition. link

Reasoning and Planning

• PathSymphony: Harmonizing symbolic planning and Large Language Models for curriculum-guided mathematical reasoning - Primary research, in Knowledge-Based Systems. Couples a symbolic planner with an LLM and sequences problems curriculum-style for mathematical reasoning, the neurosymbolic pattern applied to math pedagogy. Notable for this project because the curriculum is the mechanism, not just the application. link
• Progressive subgoal-aggregated long-sequence decision-making with large language models - Primary research, in Engineering Applications of Artificial Intelligence. Tackles long-horizon decision-making by progressively aggregating subgoals rather than planning the whole sequence at once, trading global optimality for tractable stepwise commitment. link
• Discovering Ordinary Differential Equations with LLM-Based Qualitative and Quantitative Evaluation - Primary research, arXiv (cs.AI). Argues symbolic regression for equation discovery over-relies on quantitative fit metrics and adds LLM-based qualitative judgment of candidate equations, using the model as a scientific taste function alongside the numbers. link
• HABIB_TAZ at SemEval-2026 Task 11: Disentangling Formal Logic from Content via Synthetic Training and Multi-Objective Optimization - Primary research, arXiv (cs.CL), SemEval-2026 system paper. Targets content effects, the measurable bias of LLMs toward real-world plausibility over formal validity, using synthetic training data and multi-objective optimization to separate logic from content. link

Knowledge Graphs and RAG

• Integrating domain-specific knowledge graph with large language model for question-answering of construction laws and regulations: The case of China - Primary research, in Engineering Applications of Artificial Intelligence. Grounds regulatory QA in a domain knowledge graph so answers about construction law trace to actual provisions rather than parametric memory. link
• Orchestrating large language models with an expert-driven knowledge graph for causal reasoning and management of construction project risks - Primary research, in Advanced Engineering Informatics. Uses an expert-built knowledge graph to steer LLM causal reasoning about project risks, keeping the causal structure human-authored and the language work model-driven. link
• RealKGC: Relation-constrained large language models for inductive knowledge graph completion - Primary research, in Knowledge-Based Systems. Constrains LLM generation with relation schemas for inductive knowledge graph completion, so the model can propose links for unseen entities without inventing ill-typed ones. link
• Empowering building emergency response with large language models and BIM-Based knowledge graphs - Primary research, in Advanced Engineering Informatics. Combines building information models rendered as knowledge graphs with LLMs so emergency-response queries can be answered against the actual structure of a building. link
• A pavement maintenance decision-making method based on a retrieval-augmented generation framework with large language models - Primary research, in Advanced Engineering Informatics. RAG over maintenance records and standards to recommend pavement interventions, a representative example of RAG as the default grounding pattern in civil engineering. link
• A Human-in-the-Loop conceptual design framework jointly driven by large language models and knowledge graphs - Primary research, in Advanced Engineering Informatics. Puts a designer between an LLM proposing concepts and a knowledge graph constraining them, with each iteration updating both. link
• MKDS: Multi-source knowledge-driven data synthesis framework for effective domain adaptation of large language models - Primary research, in Knowledge-Based Systems. Synthesizes fine-tuning data from multiple structured knowledge sources so domain adaptation does not depend on scarce human-written domain text. link
• Text2Onto-Agent: An LLM Agent-Based End-to-End Automated Ontology Construction Method and a Case Study of Green Building Domain Modeling - Primary research, in Journal of Construction Engineering and Management. An agent that builds domain ontologies end to end from text, demonstrated on green-building codes, aiming to replace the expert-intensive ontology engineering bottleneck. link
• LLM4KGen: A framework for developing KG-based semantic applications with LLMs, RAG and AI agents - Primary research, in Data & Knowledge Engineering. A development framework that packages the LLM-plus-KG-plus-RAG-plus-agents stack so semantic applications can be assembled rather than hand-built. link
• Version 6.6.1 - AI-KM: Agent skills and ontology-driven knowledge modeling - Primary research, in SoftwareX. A software release pairing agent skills with ontology-driven knowledge modeling; notable that "agent skills" is now vocabulary in archival software journals. link
• DualHNIE: Dual-Channel Hypergraph Learning for Node Importance Estimation in Heterogeneous Knowledge Graphs - Primary research, arXiv (cs.LG). Estimates node importance in heterogeneous knowledge graphs with dual-channel hypergraph learning, arguing pairwise message passing misses the higher-order structure that determines importance. link

LLM Evaluation

• Can large language models recognize complex language errors such as zeugma? - Primary research, in Engineering Applications of Artificial Intelligence. Probes whether LLMs detect subtle rhetorical-grammatical errors like zeugma, a niche but sharp test of whether linguistic competence extends past surface fluency. link
• Growing a Tail: Increasing Output Diversity in Large Language Models - Primary research, arXiv (cs.CL). Measures response diversity on questions with many valid answers against human response distributions, finding model outputs concentrated relative to humans, and studies interventions to regrow the distribution's tail. Relevant to any assignment where students expect varied model outputs. link

LLM Cognition

• Left-right asymmetry in predicting brain activity from LLMs' representations emerges with their formal linguistic competence - Primary research, arXiv (cs.CL). Shows that as LLMs train, their representations predict human brain activity with a left-right asymmetry that emerges alongside formal linguistic competence, tying a hallmark of human language lateralization to a measurable training milestone. link
• How LLMs Might Think - Primary research, arXiv (cs.CL). Responds to Stoljar and Zhang's argument from rationality that LLMs do not think, contending the argument fails and leaves open a positive account of machine thought. Useful reading for a philosophy-of-AI seminar week. link
• A Conceptual Framework for Artificial Intelligence Combining Buddhism and the Free Energy Principle - Primary research, in Digital Humanities, Social Science and Cultural Preservation. Sketches an AI framework joining the free energy principle account of brain function with Buddhist theory of mind. Speculative, but a data point on how far afield AI theorizing now reaches. link

Privacy and Security

• LLMs Leak Training Data Beyond Verbatim Memorization: Extraction via Membership Decoding - Primary research, in Proceedings on Privacy Enhancing Technologies. Shows the generate-then-audit extraction paradigm undercounts leakage: membership decoding steers generation toward training-set content, extracting data that never surfaces verbatim under greedy decoding. Materially expands what counts as memorization risk, and strong teaching material for a privacy unit. link
• SHIELD: Advanced persistent threats detection and intelligent explanation using large language models - Primary research, in Engineering Applications of Artificial Intelligence. Uses LLMs both to detect advanced persistent threats and to explain detections to analysts, pairing the classifier with a narrator. link
• Vulnerability detection of smart contracts by integrating large language models and dependency analysis - Primary research, in Engineering Applications of Artificial Intelligence. Combines LLM code understanding with dependency analysis to catch smart-contract vulnerabilities that pattern matching alone misses. link
• Auto Threat AI: An Agentic and Explainable Framework for Automated Cyber Threat Intelligence Extraction - Primary research, in IJRASET. Automates extraction of structured cyber threat intelligence from unstructured reports with an agentic, explainable pipeline, targeting the manual-analysis bottleneck in security operations centers. link
• Defending against AI-driven social engineering: a conceptual framework - Primary research, in AI and Ethics. A conceptual defense framework for social engineering attacks that are themselves AI-generated, mapping the attack surface before proposing countermeasures. link
• Overcoming Language Barriers: Multilingual Analysis of the 2023 Swiss Privacy Law's Impact - Primary research, arXiv (cs.CL). Uses multilingual NLP to measure the real-world impact of the 2023 Swiss privacy law, extending empirical privacy-regulation research beyond the EU and California. link

Systems and Efficiency

• FlexiTensor: Adaptive Multi-Task Deployment of LLMs on Resource-Constrained Heterogeneous Edge Devices - Primary research, in IEEE Transactions on Parallel and Distributed Systems. Deploys LLMs across heterogeneous, resource-constrained edge devices for multiple tasks at once, driven by privacy and latency requirements that rule out the cloud. Solid systems-side reading for a local-inference unit. link
• Artificial Intelligence-Based Predictive Approximation of Belady's Optimal Page Replacement Algorithm Using Deep Learning and Reinforcement Learning - Primary research, in IJRASET. Approximates Belady's clairvoyant page-replacement algorithm with learned predictors, a classic teaching example of ML approximating an optimal-but-unrealizable policy. link
• Column Generation with Domain-Independent Dynamic Programming - Primary research, arXiv (cs.AI). Brings domain-independent dynamic programming to the pricing problem inside column generation and branch-and-price, aiming to get exact large-scale optimization without hand-crafted pricing solvers. link
• Towards quantum machine learning for assessing the resilience of post-quantum cryptography - Primary research, arXiv (cs.LG). Explores using quantum machine learning to stress-test post-quantum cryptographic schemes, probing the defenses built for future quantum adversaries with today's quantum-adjacent tools. link
• A short review on the maximum clique problem algorithms with classical, AI, and quantum methods - Primary research, arXiv (cs.AI). Surveys maximum-clique algorithms across classical, AI-based, and quantum methods, a compact map of one NP-hard problem across three solver paradigms. link

Robotics and Embodied AI

• Agile perceptive multi-skill locomotion for quadrupedal robots in the wild - Primary research, arXiv (cs.AI). Integrates multiple motor skills, smooth gait transitions, and high-speed perceptive locomotion on quadrupeds using only onboard sensing, pushing lab locomotion out into unstructured terrain. link
• Pretraining in Actor-Critic Reinforcement Learning for Locomotion - Primary research, arXiv (cs.LG). Imports the pretraining-finetuning paradigm into actor-critic RL for robot locomotion, where skills are usually learned from scratch per task, testing whether locomotion has its own transferable foundation. link
• Koopman-driven grip force prediction through EMG sensing - Primary research, arXiv (cs.AI). Predicts grip force from surface EMG using Koopman operator theory for rehabilitation robotics, giving patients with stroke or MS a control signal from residual muscle activity. link
• Enhancing ROS Debugging: A User-Centric Diagnostic Framework - Primary research, in IEEE Robotics and Automation Letters. A diagnostic framework for the Robot Operating System aimed at the comprehension barriers created by its distributed architecture and messaging system. link
• Science-intent-driven embodied intelligent solar telescope: concept design (invited) - Primary research, in Laser & Optoelectronics Progress. An invited concept design for a solar telescope operated by embodied AI driven by scientific intent, an early example of LLM-era autonomy proposed for major scientific instruments. link

Clinical LLMs

• Do bots provide correct and adequate guidance regarding acidity: A blinded comparison rated by patients and physicians - Primary research, in World Journal of Methodology. Blindly compares LLM guidance on gastrointestinal acidity, rated by both patients and physicians, measuring accuracy, empathy, actionability, and readability rather than accuracy alone. link
• Evaluating large language models for structuring cardiology reports: a real-world clinical study on patient subtyping and trial recruitment - Primary research, in International Journal of Medical Informatics. A real-world study using LLMs to structure free-text cardiology reports for patient subtyping and clinical-trial recruitment, moving evaluation from benchmarks to a live clinical pipeline. link
• Home-based sarcopenia diagnosis via multimodal gait analysis and personalized large language model intervention - Primary research, in Engineering Applications of Artificial Intelligence. Combines multimodal gait analysis for at-home sarcopenia screening with an LLM that personalizes the intervention, sensing plus counseling in one loop. link
• EEG-AI: An agentic system for AI-assisted semi-automated EEG preprocessing and artifact removal - Primary research, in Journal of Neuroscience Methods. An agentic system that semi-automates EEG preprocessing and artifact removal, targeting the low signal-to-noise drudgery that precedes any neural-marker analysis. link
• MediCARE: Medical Collaborative Agents REasoning over Interpretable Heterogeneous Graphs - Primary research, in Artificial Intelligence in Medicine. Collaborative medical agents that reason over interpretable heterogeneous graphs, keeping the reasoning substrate inspectable where clinical accountability demands it. link
• Personalized AI Cardiovascular Risk Twin Using Explainable Reinforcement Learning - Primary research, in IJRASET. Builds a personalized cardiovascular digital twin with explainable RL to predict disease progression and recommend interventions. link
• Artificial Intelligence-Generated Electronic Medical Record Summarization in Breast Surgical Oncology - Primary research, in Annals of Surgical Oncology. Evaluates AI-generated EMR summaries in breast surgical oncology, a specialty-specific test of the most widely deployed clinical LLM use case. link
• Why health information remains hard to read: The curse of knowledge as a structural barrier to patient education materials - Primary research, in Patient Education and Counseling. Diagnoses the persistent unreadability of patient materials as a curse-of-knowledge problem: experts structurally cannot simulate novice readers. Frames exactly the gap LLM rewriting tools claim to fill. link
• Patient- and Caregiver-Informed Considerations for the Design and Implementation of Generative AI-Supported Patient-Centered Clinical Decision Support: Qualitative Study - Primary research, in Journal of Medical Internet Research. Qualitative study surfacing what patients and caregivers actually want from GenAI-supported clinical decision support before such systems are designed around them. link
• Generative large language models in the clinical management of Alzheimer's disease and mild cognitive impairment - Primary research, in Neurological Sciences. Reviews LLM use in managing Alzheimer's and mild cognitive impairment, where clinicians must integrate uncertain evidence across neuropsychology, imaging, and biomarkers. link

Legal and Financial LLMs

• ClauseMiner: Prompt-engineered large language model for accurate, scalable legal clause extraction - Primary research, in Engineering Applications of Artificial Intelligence. Prompt-engineered clause extraction at contract scale, the unglamorous end of legal NLP where accuracy requirements are strict and volume is high. link
• Identifying central bank multi-objective preferences using large language models: Evidence from China - Primary research, in Finance Research Letters. Uses LLMs to read central bank communications and recover the bank's implicit weighting across competing objectives, with Chinese monetary policy as the case. link
• Financial news sentiment meets market data: A large language model-based approach to stock price prediction - Primary research, in Information Sciences. Fuses LLM-extracted news sentiment with market data for stock prediction, the standard test of whether language signals add alpha over price history. link
• Large Language models for banking supervision: reliability evidence from European systemic banks - Primary research, in Finance Research Letters. Tests LLM reliability on supervisory tasks over European systemic banks, evidence that matters because supervisory errors are asymmetric and expensive. link
• AI: Stare Non Decisis - Primary research, in Open Access Journal of Artificial Intelligence and Technology. Argues the deep danger of AI in legal practice is not hallucinated citations but erosion of precedent itself: if lawyers reason through models rather than through cases, stare decisis quietly stops doing its work. link

Industrial LLM Applications

• A semi-supervised multi-stage pipeline for low-resource industrial condition monitoring by injecting domain knowledge via cross-modal multimodal large language model fine-tuning - Primary research, in Engineering Applications of Artificial Intelligence. Injects domain knowledge into condition monitoring through cross-modal LLM fine-tuning where labeled industrial data is scarce. link
• A bicycle traffic prediction framework using pretrained large language models - Primary research, in Transportation Research Part C. Adapts pretrained LLMs to bicycle traffic prediction, part of the wave testing whether language-pretrained representations transfer to numeric urban time series. link
• LKG-STNet: A large language model-assisted knowledge graph-guided spatiotemporal network for aero-engine remaining useful life prediction - Primary research, in Advanced Engineering Informatics. LLM-assisted, KG-guided spatiotemporal prediction of aero-engine remaining useful life, stacking three modeling paradigms on one high-stakes maintenance problem. link
• Large language models in mechanical design of mechatronic systems: A review - Primary research, in Advanced Engineering Informatics. Reviews LLM applications across mechanical design of mechatronic systems, a field map for where design automation actually stands. link
• Large language model-driven information extraction and causal chain reasoning from chemical accident investigation reports for process safety intelligence - Primary research, in Journal of Loss Prevention in the Process Industries. Extracts causal chains from chemical accident reports so process-safety lessons become queryable rather than buried in PDFs. link
• Visual question answering for bridge damage inspection using a multi-modal large language model - Primary research, in Automation in Construction. Multimodal VQA over bridge inspection imagery, letting engineers interrogate damage photos in natural language. link
• A dual-layer bilevel optimization model to complete algorithm evolution in agile satellite task scheduling problem using Large Language Models - Primary research, in Engineering Applications of Artificial Intelligence. Uses LLMs inside a bilevel optimization loop to evolve scheduling algorithms for agile satellites, LLM-as-algorithm-designer rather than LLM-as-scheduler. link
• XFD-LVLM: An explainable multimodal framework for aviation hydraulic pump intelligent fault diagnosis with large Vision-Language models - Primary research, in Advanced Engineering Informatics. Explainable vision-language fault diagnosis for aviation hydraulic pumps, where explanation is a certification requirement, not a nicety. link
• Semantic metadata extraction from crop model components with large language models to facilitate platform interoperability - Primary research, in Computers and Electronics in Agriculture. LLM extraction of semantic metadata from crop model code so agricultural modeling platforms can interoperate. link
• LightLLM4FDD: Domain-specific lightweight large language models for fault detection and diagnosis in building HVAC systems - Primary research, in Advanced Engineering Informatics. Lightweight domain-specific LLMs for HVAC fault detection, sized for building-management hardware rather than datacenter inference. link
• On-device large language model-powered intelligent copilot for on-site construction tutorials - Primary research, in Advanced Engineering Informatics. An on-device LLM copilot delivering construction tutorials at the job site, where connectivity and data-sensitivity rule out cloud calls. link
• An adaptive industrial large language model for mechanical fault diagnosis under variable operating conditions - Primary research, in Advanced Engineering Informatics. An industrial LLM that adapts fault diagnosis across variable operating conditions instead of assuming the fixed regimes most diagnosis models train on. link
• A novel framework for user-centric smart product-service system development leveraging large language models and the computational grounded theory - Primary research, in International Journal of Industrial Ergonomics. Pairs LLMs with computational grounded theory to derive product-service system requirements from user data. link
• Construction Quality Risk Factors Extraction Based on LLM-Assisted Text Mining from Big Data in Court Cases: Aligned with IDI Liability Clauses - Primary research, in Journal of Management in Engineering. Mines construction court cases with LLM assistance to extract quality risk factors and aligns them with inherent-defect-insurance liability clauses, turning litigation records into a risk knowledge base. link
• Remote sensing change captioning meets large language and vision models - Primary research, in ISPRS Journal of Photogrammetry and Remote Sensing. Applies large vision-language models to describing changes between satellite images in natural language. link
• Construction Safety Knowledge Reasoning-Capable Large Language Model for Hydropower and Water Conservancy Engineering - Primary research, in Journal of Construction Engineering and Management. A construction-safety LLM specialized for hydropower and water conservancy projects, domain adaptation aimed at a single hazardous sector. link
• Agent-orchestrated domain-specific large language model for large-scale database-driven geotechnical site characterizations - Primary research, in Automation in Construction. An agent orchestrates a domain LLM over large geotechnical databases to characterize sites, replacing manual borehole-log synthesis. link
• Polyaniline nanocomposite matrices in next-generation agri-sensing: A material-to-algorithm review of interfacial mechanisms, metaheuristic optimization, and agentic AI pathways - Primary research, in Biosensors and Bioelectronics: X. A materials-to-algorithm review of polyaniline agri-sensors whose forward-looking section sketches agentic AI pathways for sensor networks; the AI content is prospective rather than demonstrated. link
• An Agentic AI-Empowered product family configuration approach for personalizing rehabilitation assistive devices with Multi-Faceted wearing experience - Primary research, in Advanced Engineering Informatics. Agentic configuration of rehabilitation assistive device families around individual wearing experience. link
• Towards fully automated city operations: Integrating agentic AI with urban digital twins - Primary research, in Computers, Environment and Urban Systems. Couples agentic AI to urban digital twins as a path toward automated city operations, with the twin as the agent's world model. link
• Engineering Challenges toward Level-5 Autonomous Fixed Telecommunications Networks - Primary research, in IJRASET. Maps the engineering gaps between today's manually operated fixed telecom networks and fully autonomous self-monitoring, self-healing ones. link
• Modern technologies in the diagnosis and management of plant-pathogenic fungi: From CRISPR to the microbiome and Artificial Intelligence - Primary research, in International Journal of Science and Research Archive. Reviews diagnosis and management of plant-pathogenic fungi across CRISPR, microbiome, and AI methods; AI is one strand among several. link
• GenTL: A General Transfer Learning Model for Building Thermal Dynamics - Primary research, arXiv (cs.LG). A general transfer-learning model for building thermal dynamics that cuts the data a target building needs by transferring from source buildings. link
• From physical surfaces to human-centric heat stress: LST and UTCI heat mapping reveals nonlinear effects of urban morphology - Primary research, arXiv (cs.LG). Maps land-surface temperature against human-centric thermal comfort indices to show urban morphology drives heat stress nonlinearly, evidence for climate-adaptive planning. link

Recommendation and Personalization

• Harmonized recommendation with large language model guidance and reinforcement learning - Primary research, in Engineering Applications of Artificial Intelligence. Uses LLM guidance to shape an RL recommender, letting language-level knowledge steer exploration that pure interaction data cannot. link
• TipRec: Time-interval-aware prompting for Recommendation with large language models - Primary research, in Knowledge-Based Systems. Encodes time intervals between user actions into prompts so LLM recommenders reason about pacing, not just sequence. link
• Bridging Personalization and AI: From RAG to Agent - Primary research, at SIGIR 2026. Charts the migration of personalization from retrieval-augmented pipelines to agentic architectures that act on a user's behalf. link

Consumer AI Interaction

• How do AI service agents impact tourists' pro-environmental behavior? The effects of childlike anthropomorphism and moral disengagement - Primary research, in Journal of Hospitality and Tourism Management. Finds the design of AI service agents, specifically childlike anthropomorphism, shifts tourists' pro-environmental behavior through moral disengagement. link
• How do consumers perceive digital nudging by agentic AI in E-commerce? - Primary research, in Journal of Retailing and Consumer Services. Examines consumer perception when the nudger is an autonomous agent rather than a static interface, a question regulators will inherit. link
• Who Is Shopping With You? How Persona Design Shapes Cognitive and Social Engagement in AI Shopping Agents - Primary research, at SIGIR 2026. Shows shopping-agent persona design changes both cognitive and social engagement, evidence that persona is a functional parameter, not cosmetic. link
• SOMIN: Agentic AI for Automating Professional Visibility and Countering Content Homogenization - Primary research, at SIGIR 2026. An agent that automates professional social-media visibility while explicitly countering the content homogenization that automation causes, an interesting self-aware design constraint. link
• Artificially sincere? How chatbot language style shapes consumer perceptions across brand contexts - Primary research, in European Journal of Marketing. Shows chatbot abbreviations, formality, and message length shift perceived sincerity and effort differently across brand contexts. link

Human-AI Interaction

• Memory-Driven Self-Disclosure and Relational Turning Points: A Longitudinal Multimodal Study of Human-AI Interaction - Primary research, arXiv (cs.CL). A 24-participant, 10-session longitudinal study of a memory-augmented conversational agent, tracing how repeated interactions with memory become something participants treat as a relationship. Rare longitudinal evidence in a literature dominated by single-session studies. link
• Affective Computing in the Age of GenAI: Balancing Risks and Rewards - Primary research, in IEEE Computational Intelligence Magazine. Surveys the convergence of generative AI and affective computing, weighing emotionally responsive systems against manipulation and privacy risks. link

Creativity and Authorship

• A multi-modal framework for generating and evaluating diverse jokes using large language models - Primary research, in Engineering Applications of Artificial Intelligence. Generates and evaluates multimodal jokes with LLMs, treating humor diversity as a measurable target rather than a byproduct. link
• Text2Sign: A Single-GPU Diffusion Baseline for Text-to-Sign Language Video Generation - Primary research, arXiv (cs.CL). A single-GPU diffusion baseline for text-to-sign-language video, deliberately lowering the compute floor so accessibility research is not gated on datacenter budgets. link
• AI-Based Text-to-Video Generation Using Diffusion and Deep Learning Techniques - Primary research, in IJRASET. A text-to-video framework converting prompts into short animated sequences with diffusion methods. link
• Man, Machine, and Masterpiece: Artistic Ownership in the AI Era - Primary research, arXiv (cs.AI). Surveys the unsettled question of how ownership should be defined and attributed when human and AI contributions to creative work are intertwined. link

AI Pedagogy and Assessment

• Enhancing a large language model with a chain-of-metacognitive reasoning approach increases argumentative writing evaluation accuracy, student writing outcomes, and mental effort - Primary research, in Computers & Education. Adds chain-of-metacognitive reasoning to an LLM writing evaluator and measures gains on three fronts at once: evaluation accuracy, student writing outcomes, and student mental effort. One of the few items today with a measured student-outcome effect, which is the bar this project cares about. link
• Educators' Discussion with a Generative Artificial Intelligence Trainer - Primary research, in Innovations in Pedagogy and Technology. Studies K-12 educators in dialogue with a GenAI trainer, looking at tool development from the educator side rather than the usual student-impact side. link
• Estimating difficulty for the open-ended reading comprehension questions via Large Language Models - Primary research, in Engineering Applications of Artificial Intelligence. Uses LLMs to estimate difficulty of open-ended reading comprehension questions, a building block for adaptive assessment where classical item-response calibration needs data that open-ended items rarely have. link
• Harnessing LLMs for Reliable Academic Supervision: A Comparative Study - Primary research, arXiv (cs.CL). Frames academic supervision as a harness-engineering problem: single-shot prompts are fluent but unreliable, so the paper compares deliberate compositions of deterministic scaffolding around the model. Continues the harness-engineering thread from yesterday's digest into an education setting. link
• Conceptual Framework of Common Misconceptions Regarding Generative AI in Elementary School Students Using Concurrent Think-Aloud Protocols - Primary research, in Computers. Uses think-aloud protocols to catalog elementary students' misconceptions about GenAI, the raw material any early AI-literacy curriculum has to correct. link
• LLMs as Cognitive Partners in Shipping 4.0: An Extend AI Approach to Maritime Predictive Maintenance Training - Primary research, in Georgian Maritime Scientific Journal. Proposes an intelligent predictive-maintenance training framework combining generative and explainable AI for maritime education, casting the LLM as cognitive partner rather than answer engine. link
• Developing an Education Framework for MIS Professionals Aiming at Social Impact: A Systematic Review and Design-Oriented Synthesis - Primary research, in Sustainability. Synthesizes a socio-technical curriculum design framework for management information systems education oriented at social impact. link
• Artificial Intelligence in Education: Can AI Improve Teacher Performance or Will It Replace Them? - Primary research, in International Journal of Social Sciences, Language, and Education. Examines the augmentation-versus-replacement question for teachers directly, a framing students and faculty both bring to every AI-in-education conversation. link
• The Effect of ChatGPT Utilization on Students' Digital Literacy at SMA Negeri 1 2X11 Kayutanam - Primary research, in Journal Multidisciplinary Science. Measures how ChatGPT use affects digital literacy at one Indonesian secondary school, small-N but ground-level evidence. link
• The Role of Artificial Intelligence in Education: Opportunities, Challenges, and Implications for Formal and Non-Formal Learning - Primary research, in Journal of Non-Formal and Digital Education. Surveys AI's role across formal and non-formal learning after the generative-tools shift moved AI from research topic to routine classroom presence. link
• Teacher-regulated generative AI support, student agency, and perceived learning gains in higher education: the moderating role of perceived fairness - Primary research, in Frontiers in Artificial Intelligence. Finds the educational value of GenAI depends on how teachers regulate its use, with perceived fairness moderating the link between regulated support, student agency, and perceived gains. link
• The Effectiveness of Generative AI in Learning History - Primary research, in Proceedings of the International Conference on New Findings in Humanities and Social Sciences. A qualitative study of GenAI in history learning, deliberately departing from the quantitative-effect designs that dominate the area. link
• Psycholinguistic markers of a comfortable educational environment for students in the context of digital transformation - Primary research, in Vestnik of Samara State Technical University. Identifies psycholinguistic markers of student well-being as learning moves beyond conventional classroom formats. link
• From complexity to compassion: reimagining higher education through academic hospitality, integrity and systems thinking - Primary research, in Education Innovations: Systems and Future Learning. Argues higher education under volatility needs equity-driven, systems-thinking frameworks rather than purely technical fixes. link
• Etika Digital Guru dalam Pemanfaatan Teknologi Pembelajaran: Kajian Systematic Literature Review - Primary research, in Edutik. A systematic literature review of teachers' digital-ethics practices and challenges in using learning technology, including AI, in Indonesian education. link

Language Learning

• How Well Does AI-Generated Feedback Work? Intrinsic and Extrinsic Evaluation across more than 20,000 EFL Essay Drafts - Primary research, arXiv (cs.CL). Evaluates LLM written corrective feedback on more than 20,000 EFL essay drafts, both intrinsically (feedback quality) and extrinsically (what happens to student writing). The scale makes it the strongest evidence base in today's education batch. link
• Sustainable L2 writing pedagogy in Turkish higher education: Effects of AI-mediated feedback on self-regulated learning and writing performance - Primary research, in PLoS ONE. Compares feedback modalities in Turkish higher education, measuring effects of AI-mediated feedback on self-regulated learning strategies and writing performance. link
• Future-Ready Pedagogy in English Language Teaching: Toward an Ethical and AI-Responsive Framework - Primary research, in International Journal of Language and Artificial Intelligence. Proposes an ethical, AI-responsive framework for English language teaching as generative tools reshape feedback, personalization, and interaction. link
• Artificial Intelligence in University English Classes in Taiwan: Opportunities, Risks, and a Context-Sensitive Framework for Integration - Primary research, in World Journal of Advanced Research and Reviews. Reviews GenAI in Taiwanese university English classes, where EFL learning intersects English-medium instruction and national bilingual policy, and offers a context-sensitive integration framework. link
• Artificial Intelligence in English Language Teaching and Learning: A Scoping Review of Intelligent Computer-Assisted Language Learning (2015-2025) - Primary research, in Arab World English Journal. A decade-spanning scoping review of intelligent computer-assisted language learning, useful for periodizing the pre- and post-generative eras. link
• Learning German from a Prompt-Based Perspective - Primary research, in EL.LE. A qualitative workshop study of upper-secondary learners of German interacting with GenAI during meaning-making, examining prompting as a language-learning activity in itself. link

AI in Medical Education

• Toward an AI-integrated nursing curriculum: A Kano model analysis of generative AI competency needs - Primary research, in Nurse Education Today. Applies Kano analysis to sort which GenAI competencies nurses consider must-have versus attractive, giving curriculum designers a prioritization tool instead of a wish list. Directly reusable method for this project's competency model. link
• The integration of generative artificial intelligence in nursing education from 2020 to 2025: A bibliometric analysis - Primary research, in Journal of Nursing Reports in Clinical Practice. Maps five years of research on GenAI in nursing education bibliometrically, showing where the field's attention actually went. link
• Eroding scholarly integrity: Confronting the misuse of generative AI in nursing education - Primary research, in Nurse Education Today. Confronts GenAI misuse in nursing education specifically, where integrity failures propagate into clinical competence. link
• AI literacy among healthcare professionals and students in the Americas - Primary research, in The Lancet Regional Health - Americas. Assesses AI literacy across healthcare professionals and students in the Americas as AI tools spread through medical education faster than training keeps up. link
• From parallel to aligned: a United States Medical Licensing Examination (USMLE) mapping and assessment-tagging framework for medical education - Primary research, in BMC Medical Education. Builds a mapping and tagging framework to close the "parallel curriculum" gap between what medical schools teach and what the USMLE tests. link
• Medical students' attitudes, usage patterns, and associated factors with DeepSeek adoption in education: a cross-sectional study in China - Primary research, in BMC Medical Education. A cross-sectional study of Chinese medical students adopting DeepSeek, a reminder that the model students actually use varies by region, and adoption research should too. link

Academic Integrity

• Understanding How University Guidelines Address Privacy and Security Issues of Generative AI in Academic Settings - Primary research, in Proceedings on Privacy Enhancing Technologies. Audits university GenAI guidelines for how they handle privacy and security given that the systems are controlled by a handful of private companies. A rigorous venue examining the documents most institutions wrote hastily. link
• The Misclassification of Autistic Writing as AI-Generated - Primary research, arXiv (cs.CL). Tests anecdotal claims that AI-text detectors disproportionately flag autistic writers, adding a disability-bias dimension to the already-shaky case for detection tools in assessment. Anyone still relying on detectors for integrity enforcement should read this. link
• The Flaw of Perfection: How University Professors Identify the Inauthentic Voice of AI - Primary research, in Journal of Academic Ethics. Studies how professors detect AI's inauthentic voice, finding suspicion attaches to flawless prose: the tell is perfection, not error. link
• Generative Artificial Intelligence and the Ambiguity of Academic Integrity in Higher Education - Primary research, in Education Sciences. Examines how LLMs blur the line between AI output appropriated as original authorship and legitimately assisted work, and why verification of independent work has become so hard. link
• The Polemics of Ethical AI Usage in Higher Education: A Case Study Investigating the Axiological Expectations Mismatch - Primary research, in Proceedings of the International Academic Conference on Education, Teaching and Learning. Introduces "axiological expectations mismatch" to name the gap between what institutions, faculty, and students each believe ethical AI use means. link
• Behavioral and ethical predictors of continuous intention to use generative AI responsibly in higher education - Primary research, in Contemporary Educational Technology. Extends the theory of planned behavior with ethical constructs to model what sustains responsible GenAI use over time, not just initial adoption. link
• Ethical concerns of generative AI in academic libraries: A cross-regional quantitative study of librarians' risk perceptions, trust, and governance practices - Primary research, in Journal of Librarianship and Information Science. Quantifies librarians' perceptions of GenAI risk across regions: privacy breaches, threats to intellectual freedom, and algorithmic bias, plus the governance practices they adopt in response. link
• From disclosure to accountability under generative and agentic AI in tourism research - Primary research, in Annals of Tourism Research. Argues disclosure norms are insufficient for research integrity once agentic AI contributes to studies, and accountability frameworks must replace them. link
• The algorithmic pharmakon: Derrida and generative AI in education - Primary research, in Educational Philosophy and Theory. Reads GenAI in education through Derrida's pharmakon, the remedy that is also poison, giving the integrity debate a philosophical vocabulary beyond cheating-versus-tools. link

Higher Ed Adoption

• The chat dilemma: Usefulness, risk, ethics, and student intentions to use generative AI in higher education - Primary research, in Technology and Society. Models student intention to use GenAI as a tradeoff among usefulness, risk, and ethics rather than a pure utility calculation. link
• Generative AI Use and Critical Thinking Dispositions in Higher Education: A Cross-Sample Study of the Sequential Role of Metacognitive Weakness and Epistemic Laziness - Primary research, in Journal of Intelligence. Across samples, links GenAI use to critical-thinking dispositions through a sequential path: metacognitive weakness feeding epistemic laziness. Names the mechanism behind the "AI makes students lazy" worry precisely enough to test. link
• Drivers and barriers to generative AI adoption in public higher education: a toe-based multi-case study of administrative stakeholders - Primary research, in Journal of Systems and Information Technology. A technology-organization-environment multi-case study of GenAI adoption among administrators at two Canadian public institutions, covering the staff side research usually skips. link
• Assessing the Baseline Landscape of Generative Artificial Intelligence (AI) Integration: An Empirical Study among University Students - Primary research, in Archives of Current Research International. Baselines GenAI integration among 115 students at an Indian agricultural college, a data point from outside the usual elite-university sampling frame. link
• Application of AI in learning and scientific research at private universities: opportunities and challenges - Primary research, in Tap chi Khoa hoc Truong Dai hoc Trung Vuong. Surveys generative AI's benefits and risks for learning and research at Vietnamese private universities. link

Practitioner Notes

• Firefox in WebAssembly - Named practitioner synthesis. Puter compiled Firefox to WebAssembly so the entire browser runs inside another browser. A vivid classroom demo of how far WASM as a compilation target has come. link
• Mermaid to ASCII art (mermaid-ascii) - Named practitioner synthesis. Willison follows up his LLM-generated Rust Mermaid-to-ASCII tool by finding an older, more complete Go library doing the same thing, a small honest note on the build-versus-search failure mode of coding with LLMs. link
• Spot birds not golf - Named practitioner synthesis. A tongue-in-cheek proposal that hyperscalers under fire for data-center water use convert golf courses to public parks and fund birdwatching. Light, but part of the ongoing datacenter-externalities conversation. link

Trends

The dominant pattern today is the industrialization of the LLM-plus-knowledge-graph recipe. Eleven knowledge-graph items and two dozen industrial applications, most in engineering journals like Advanced Engineering Informatics and Automation in Construction, follow the same template: a domain KG or database supplies ground truth, the LLM supplies language and reasoning glue, and often an agent orchestrates the two. Construction engineering alone accounts for nearly ten items. This is no longer a research frontier; it is a default architecture diffusing through every applied discipline at once.

Agentic DevOps is showing velocity from a standing start. Two independent self-healing CI/CD frameworks (A Closed-Loop Agentic AI Framework for CI/CD Automation, Auto-Devops GPT) landed on the same day, alongside agentic frameworks for prompt lifecycle management (PromptStudio), runtime assurance (Runtime assurance for enterprise agentic AI systems), and deterministic agent design. The governance side is arriving with the automation rather than after it, which is different from how the last decade of ML deployment went.

The education literature is visibly maturing past adoption surveys. Adoption and attitude studies still arrive daily, but today's stronger items measure mechanisms and outcomes: metacognitive weakness and epistemic laziness as the causal path from GenAI use to weaker critical thinking, teacher regulation and perceived fairness as moderators of learning gains, and feedback evaluated across 20,000 essay drafts. Meanwhile the integrity conversation is turning against detection tools, with The Misclassification of Autistic Writing as AI-Generated adding disability bias to the case, and The Flaw of Perfection showing human detection heuristics are themselves suspect.

On the model front, two independent practitioner sources converged on Kimi K3 within a day, framing it as the largest open release ever and near the frontier at a fraction of the price. Together with yesterday's GLM-5.2 coverage, the open-weight ecosystem is compressing its lag behind closed frontier models release by release.

Predictions

Signal: K3 ships open weights by July 27 at claimed Opus 4.8-class capability, a week after GLM-5.2 was called a step change for open agents. Inference: within 8-12 weeks, open-weight models become the default substrate for agentic coursework and research labs that need inspectability, and the interesting differentiation moves up the stack to harnesses and governance layers. Confirmed if the K3 weights actually land on schedule and benchmark near closed frontier models on agentic evals; killed if the release slips substantially or the open weights underperform the API version.

Signal: governance-aware agent frameworks (runtime assurance, policy gating, deterministic schema-guided design, prompt lifecycle management) are appearing in applied venues simultaneously with, not after, agentic automation papers. Inference: an "agent governance" tooling category consolidates within the quarter, and enterprise curricula will need a unit on it. Confirmed if vendor products and standards drafts citing autonomy-risk scoring or policy gates appear in the next two months; weakened if these remain isolated conceptual papers with no shared vocabulary.

Top picks

• LLMs Leak Training Data Beyond Verbatim Memorization: Extraction via Membership Decoding link - Top-tier privacy venue, a genuinely new extraction mechanism, and direct teaching value for any unit on memorization and privacy risk.
• Kimi K3, and what we can still learn from the pelican benchmark link - The day's most consequential release, corroborated across two independent practitioner sources, with a working lesson on personal benchmarks thrown in.
• How Well Does AI-Generated Feedback Work? Intrinsic and Extrinsic Evaluation across more than 20,000 EFL Essay Drafts link - Evaluation scale two orders of magnitude beyond typical feedback studies, and it measures downstream student writing, not just feedback plausibility.
• Enhancing a large language model with a chain-of-metacognitive reasoning approach increases argumentative writing evaluation accuracy, student writing outcomes, and mental effort link - Computers & Education venue and measured student outcomes, the standard most AI-in-education papers claim and few meet.
• The Misclassification of Autistic Writing as AI-Generated link - Tests a widely repeated anecdote with data and lands on an equity failure that should change how institutions use AI detectors.