Coronary artery calcium, or CAC, is one of the most reliable markers available for predicting future cardiovascular events. Calcium deposits build up in the walls of the coronary arteries over years, and the amount present, quantified by the well-established Agatston score on a non-contrast computed tomography scan, correlates closely with the likelihood of heart attack and stroke. Despite this diagnostic power, dedicated cardiac CT for calcium scoring is not performed universally, largely because it is costly, involves radiation exposure, and is generally reserved for patients already considered at intermediate risk. The vast reservoir of cardiovascular information embedded in images already being acquired, the chest X-rays gathering dust in hospital archives, remains largely untapped.

The research team set out to test whether a general-purpose, vision-capable large language model could bridge that gap. Their retrospective, single-center diagnostic accuracy study included 349 patients who had undergone both a non-contrast cardiac CT for calcium scoring and a chest radiograph within a six-month window. Using the CT-derived Agatston classification as the reference standard, the researchers divided the cohort into two clinically meaningful groups: those with a score of 100 or below, and those exceeding 100, the threshold generally regarded as indicating clinically significant coronary calcification that warrants aggressive risk management. Of the 349 patients, 131 fell into the high-calcium category.

The technical approach was striking in its simplicity. Rather than training a bespoke neural network on thousands of labeled images, as is typical for medical AI studies, the team fed preprocessed chest radiographs to GPT-5.5 in a zero-shot setting, meaning the model received no task-specific training whatsoever. Each image passed through a preprocessing pipeline that included contrast-limited adaptive histogram equalization, a technique that enhances local image contrast and can make faint calcific densities more visible. The model was given a structured prompt instructing it to estimate the probability that clinically significant calcification was present, assign a binary classification, and describe any visible calcification findings, mimicking the reasoning process of a human reader.

The results were cautiously encouraging. Using the model’s native binary output, the system correctly classified 292 of the 349 patients, for an overall accuracy of 83.7 percent. Sensitivity, the ability to detect patients who truly had high calcium burden, reached 72.5 percent, while specificity, the ability to correctly clear patients with low calcium, was higher at 90.4 percent. The positive predictive value stood at 81.9 percent and the negative predictive value at 84.5 percent. Perhaps most notably, the area under the receiver operating characteristic curve, a threshold-independent measure of discriminative ability, was 0.771, a figure comparable to some dedicated deep-learning systems developed for similar opportunistic screening tasks after extensive training.

The picture became more complicated, however, when the researchers examined the model’s continuous probability scores rather than its binary judgments. Applying a prespecified threshold of 0.50 to the raw scores yielded dramatically asymmetric performance: sensitivity dropped to 36.6 percent while specificity climbed to 94.0 percent. In other words, when the model did flag a radiograph as suspicious, that flag was very likely to be correct, but it missed the majority of high-calcium patients at that cutoff. The model-generated scores were indeed significantly higher in patients with CT-confirmed CAC above 100, confirming a genuine statistical association, yet the scores overlapped substantially in the intermediate range, blurring the boundary between low and high risk for a meaningful fraction of patients.

Calibration emerged as another significant weakness. A well-calibrated diagnostic score should correspond reliably to actual probability, with a reported risk of 70 percent meaning the finding is truly present in roughly seven of ten such cases. The calibration slope in this study was 2.49, far from the ideal value of 1, indicating that the model’s expressed confidence was poorly aligned with reality. For clinicians accustomed to interpreting validated risk tools, this miscalibration matters enormously, because clinical decisions often hinge not merely on whether a test is positive or negative, but on how much weight the numeric estimate can bear. The authors were transparent about this limitation, framing their results as exploratory and hypothesis-generating rather than clinically actionable.

The concept underlying the study, sometimes called opportunistic screening, has been gaining momentum across medical imaging. The idea is elegant: rather than recruiting patients for dedicated cardiovascular examinations, extract additional diagnostic value from examinations already performed for other reasons. CT scans done for lung screening have been mined for calcium scores, bone density, and fatty liver indicators. Chest radiographs, being the most common imaging examination in medicine, represent perhaps the largest potential reservoir of such incidental information. If an AI system could reliably flag patients with significant coronary calcification from these routine images, those patients could be referred for confirmatory CT calcium scoring, targeted statin therapy, or other preventive interventions years before a first cardiac event.

What makes the use of a zero-shot vision-language model particularly noteworthy is the departure from the conventional AI development paradigm. Building a dedicated medical imaging classifier typically demands curated datasets, expert annotation, fine-tuning, and validation across institutions, investments that slow the translation of algorithms into practice. Foundation models, by contrast, acquire broad visual and linguistic capabilities from web-scale training and can be steered toward new tasks through prompting alone. The Istanbul study offers an early signal that this generality may extend to subtle radiological patterns, such as coronary calcification, that were never explicitly part of the model’s training objectives. It also raises intriguing questions about whether such models perceive image features that resemble human radiological signs or rely on entirely different statistical correlations invisible to human interpretation.

The authors are careful to emphasize the considerable distance between their findings and any bedside application. The cohort was selected, comprising patients who had already undergone cardiac CT, a population very different from the unselected stream of patients whose chest X-rays would be analyzed in a true screening scenario. Human-reader benchmarking was not performed, so it remains unknown how the model compares with radiologists deliberately searching for calcification. Prospective validation in diverse populations, assessment of calibration after refinement, and evaluation of downstream clinical outcomes would all be required before opportunistic AI flagging could enter practice. Still, the central result stands: a general-purpose vision-capable large language model, without a single example of task-specific training, produced image-derived scores significantly associated with CT-defined coronary artery calcium category on plain chest radiographs. That a conversation-ready AI can glance at a decades-old imaging technology and detect the chemical fingerprints of atherosclerosis hints at how much diagnostic potential remains hidden in the world’s existing medical images, waiting for the right tools to reveal it.

Subject of Research: Using a vision-capable large language model to classify coronary artery calcium on chest radiographs

Article Title: Association between a vision-capable large language model’s image-derived score and CT-defined coronary artery calcium category on routine chest radiographs: a single-center retrospective study

Article References: Ayyildiz, H., Banaz, T., Kazanbas, M. F., Gunes, S. T., Arslan, A., Kartal, Y. C., & Ozkok, S. (2026). Association between a vision-capable large language model’s image-derived score and CT-defined coronary artery calcium category on routine chest radiographs: a single-center retrospective study. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02757-7

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02757-7

Keywords: coronary artery calcium, chest radiograph, large language model, GPT-5.5, zero-shot learning, diagnostic accuracy, opportunistic screening, cardiovascular risk, Agatston score, medical imaging AI, computed tomography, BMC Medical Imaging

The research team, led by emergency physicians at Van Training and Research Hospital, University of Health Sciences, enrolled 412 consecutive patients aged 0 to 17 years who presented with acute elbow trauma between December 2025 and May 2026. This was not a retrospective exercise on curated images; it was a prospective, single-center diagnostic accuracy study designed to reflect the real conditions of an emergency department. De-identified anteroposterior and lateral radiographs were exported from the hospital’s picture archiving and communication system as PDF files and presented to GPT-5.2 Instant through the ordinary ChatGPT web interface, once per case, using a fixed prompt and no task-specific fine-tuning. In other words, the model was tested exactly as a curious clinician might use it today, without any specialized medical training or optimization.

To establish the truth against which the AI’s answers were judged, the researchers built a composite reference standard. Two emergency physician readers assessed each case independently, and when their assessments agreed, that consensus served as the final diagnosis. When the two readers disagreed, the case went to an independent radiologist for review and adjudication. This layered approach is important because pediatric elbow radiographs are notoriously tricky: growth plates, ossification centers that appear at different developmental stages, and subtle supracondylar fractures can all mimic or mask injuries. The final adjudicated reference standard identified fractures in 41 of the 412 patients, a prevalence of 10.0 percent, meaning that roughly nine out of every ten children in the cohort had no fracture at all.

Against that benchmark, GPT-5.2 Instant classified 74 cases, or 18.0 percent of the cohort, as fracture-positive, a substantial overcall. The model correctly identified 23 of the 41 true fractures but missed 18, yielding a sensitivity of 56.10 percent with a 95 percent confidence interval of 39.75 to 71.53. That figure means the model failed to detect nearly half of the actual fractures. Its specificity was 86.25 percent, which sounds respectable until one considers that in a population with only 10 percent prevalence, even modest overcalling produces many false positives. The positive predictive value, the probability that a fracture-positive call was truly a fracture, was just 31.08 percent, meaning roughly two out of three fracture alerts from the model were wrong.

The full statistical picture reinforces the sobering conclusion. The negative predictive value was 94.67 percent, which might tempt some to suggest the model could serve as a rule-out tool, but the authors explicitly warn against that interpretation. The negative likelihood ratio of 0.51 indicates that a negative result from the model reduces the probability of fracture only modestly, far from the near-certainty required for a safe rule-out strategy in pediatric trauma. The positive likelihood ratio of 4.08 is similarly unimpressive by diagnostic standards, where values above 10 are typically considered useful. Overall accuracy was 83.25 percent, and agreement with the reference standard, measured by Cohen’s kappa, was limited at 0.312, a level conventionally described as fair at best.

Why does a model capable of passing medical licensing examinations and generating fluent clinical summaries struggle so visibly with pediatric elbow X-rays? The technical answer likely lies in the nature of the task and the way the model was deployed. Pediatric elbow radiography is one of the most nuanced domains in musculoskeletal imaging. The developing skeleton presents multiple ossification centers whose appearance and fusion follow a well-known but variable sequence, and distinguishing a normal growth plate from a subtle fracture line requires pattern recognition honed on thousands of similar cases. General-purpose multimodal large language models are trained on broad internet-scale data rather than curated radiology archives, and they were not optimized here with task-specific fine-tuning. The PDF-based input pathway through a web interface may also have introduced compression and resolution constraints that degrade fine bony detail, though the study’s design deliberately reflected realistic, non-specialist access conditions.

The study’s secondary analyses add further caution. The researchers explored whether the model could classify fracture subtypes, including the three grades of supracondylar fracture, and whether it could offer sensible recommendations about surgical importance and immobilization. The supplementary data show that agreement on these management-related classifications was similarly limited, and the authors note that the surgical-importance analysis was exploratory because it included only ten physician-positive cases and three true-positive GPT classifications, making the resulting positive likelihood-ratio estimate unstable. Nursemaid’s elbow, a common pediatric injury that is primarily a clinical rather than radiographic diagnosis, was not evaluated as a radiographic target. These caveats matter because the appeal of a general-purpose AI assistant in the emergency department is precisely that it might help with triage and management, not just detection.

The authors are careful about what their findings do and do not show. They conclude that under the reported access and input conditions, GPT-5.2 Instant demonstrated limited sensitivity and agreement for pediatric elbow fracture detection, insufficient to support independent diagnostic or rule-out use. They also emphasize that clinician-AI collaboration was not evaluated in this study and requires prospective investigation. That distinction is crucial: the question is not whether AI can replace the radiologist’s eye, but whether a human-machine partnership could improve safety and efficiency. A model that misses nearly half of fractures on its own might still add value if it prompts a tired clinician to look more carefully, but proving that requires different study designs that measure outcomes when the model operates alongside, rather than instead of, human judgment.

The study also carries methodological strengths worth noting. It was prospective, meaning patients were enrolled according to a pre-specified protocol rather than selected retrospectively from archives. It used consecutive sampling, reducing selection bias. The ethics approval from the Non-Interventional Clinical Research Ethics Committee of Van Training and Research Hospital, granted on 24 October 2025, covered the radiographic image-input and reference-standard adjudication procedures before enrollment began, and written informed consent was obtained from a parent or legal guardian of every participant. The work was conducted in accordance with the Declaration of Helsinki, and the authors received no external funding, declaring no competing interests. Confidence intervals for predictive values were calculated using the logit-binomial method, likelihood-ratio intervals on the logarithmic scale, and F1-score intervals by nonparametric bootstrap resampling, reflecting careful statistical practice.

For the broader conversation about AI in medicine, the message is neither dismissal nor endorsement but calibration. The study was accepted for oral presentation at EUSEM 2026 in Paris, signaling that the emergency medicine community considers the question timely. Multimodal large language models are evolving rapidly, and future versions, or fine-tuned medical derivatives of them, may perform substantially better on this task. But this prospective, real-world evaluation demonstrates that the gap between impressive general capabilities and reliable diagnostic performance in a specialized pediatric imaging domain remains wide. Until models can demonstrate high sensitivity and strong agreement under realistic conditions, the reading of children’s elbow radiographs should remain firmly in human hands, with AI serving, at most, as an unproven assistant rather than an independent diagnostic authority.

Subject of Research: Diagnostic accuracy of the GPT-5.2 Instant multimodal large language model for detecting pediatric elbow fractures on radiographs

Article Title: Diagnostic performance of GPT-5.2 ınstant for pediatric elbow fracture detection on radiographs: a prospective single-center diagnostic accuracy study

Article References: Taş, O., Yorgun, M., Aktaş, R. S., Büyükkaya, M. Ş., & Şahin, M. (2026). Diagnostic performance of GPT-5.2 ınstant for pediatric elbow fracture detection on radiographs: a prospective single-center diagnostic accuracy study. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02788-0

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02788-0

Keywords: GPT-5.2 Instant, pediatric elbow fracture, radiography, diagnostic accuracy, large language model, emergency department, artificial intelligence, BMC Medical Imaging, supracondylar fracture, ChatGPT, medical imaging, prospective study

The core insight of the research is a deliberate shift in what gets measured. Previous smartphone-based stress studies have concentrated on indicators such as total screen time, the number of times apps are launched, the timing and frequency of use by app category, and typing behavior. Those metrics are easy to collect, but they carry a structural weakness: whenever the apps a person uses change, or when an app’s interface is redesigned, the underlying data change as well. A stress-detection model trained on one set of apps can therefore lose stability over long periods, precisely the timescale over which mental health monitoring is most valuable. Longitudinal reliability, the ability of a model to mean the same thing in month six as it did in week one, has been a persistent challenge for the field.

The KAIST team’s answer was to stop asking which app a person is using and instead ask what the person is doing. Even if a user abandons KakaoTalk for LINE or WhatsApp, the activity itself remains messenger chatting. The analysis is anchored to the purpose of the behavior rather than the name of the app. The researchers formalized this idea as Smartphone Human Interaction Routine Behavior Tasks, abbreviated SHIRBTs: repeated, purposeful units of smartphone interaction that persist across different applications and interface updates. Because these behavioral units are defined at the level of human intent, they are far less fragile than app-level usage counts.

Identifying these routines automatically required a two-stage technical pipeline. First, the team developed a technique called Multi-Level Sequential Pattern Mining, or ML-SPM, which scans screen-transition sequences recorded across multiple days and extracts recurring patterns of navigation. These patterns capture the characteristic ways a person moves through screens when performing a familiar task. Second, the system uses a large language model to assign human-readable labels to the mined patterns, producing categories such as messenger chatting, web browsing, financial payment, and reading or writing email. The labeling step matters because it converts raw interaction traces into behavioral units that both the model and a human observer can understand, an important property for any technology that aspires to transparency in mental health applications.

The classification of these behaviors proved highly reliable. In evaluations testing whether the same behavior could be consistently recognized across different email clients, web browsers, and messenger apps, the system achieved an accuracy of more than 95 percent. That consistency across app families is the technical foundation on which the rest of the study rests: if the behavioral labels are stable, then models built on them should inherit that stability even as individual applications come and go from a user’s phone.

Privacy considerations shaped the design from the outset. The system does not store the actual text or message content entered by users. It records only behavioral information: whether text was entered, how often it was entered, and how screens transitioned during use. Information about message recipients or call partners is likewise obscured through hashing, a process that converts the original identifiers into values that do not directly reveal them. This means the framework can characterize the rhythm and shape of a person’s digital day without ever reading a message or learning who a person communicated with, a distinction that matters for any technology intended to touch on sensitive mental state data.

To test whether the framework could actually detect stress, the team analyzed real-life data from 26 university students and office workers in their 20s and 30s, all Android users. Among the participants for whom both daily self-reported stress assessments and smartphone records were available, 16 short-term participants contributed an average of 30.5 days of data each, while 10 long-term participants contributed an average of 105.1 days. Anchored to the participants’ daily self-assessments, the researchers built personalized artificial intelligence models that distinguished between days with stress and days without. In total, 29 distinct types of repeated behaviors were identified across the group, with each individual routinely repeating an average of about seven of them.

The results showed a consistent advantage for behavior-level analysis over app-level analysis. The best-performing personalized models achieved an average accuracy of 75.9 percent for short-term participants and 73.8 percent for long-term participants, meaning the system was correct roughly three times out of four about whether a given day was a stressed day. Those figures exceeded the conventional approach, which fed app usage data into the same AI technique, by 5.5 and 3.5 percentage points respectively. Notably, performance did not depend on observing every behavior: with the same personalized model, using only two behaviors, messenger chatting and financial payment, or even messenger chatting alone, produced higher average classification accuracy than the conventional method. In the long-term data, the indicators derived from repeated behaviors also changed less over time and maintained a more stable structure than those used in the conventional approach, confirming the central design hypothesis that behavioral routines outlast the apps that host them.

The practical implications extend in two directions. In this experiment, the researchers deliberately collected a broad range of smartphone usage records so that their method could be compared fairly with conventional approaches. In future deployments, once an individual’s repeated behaviors have been identified, data collection and analysis could be focused on just those relevant behaviors, reducing the burden of continuous logging. The study also suggests a path toward digital mental health monitoring that is both interpretable and durable: because the system classifies stress using labeled human behaviors rather than opaque app statistics, it is easier to understand which behaviors drove a given classification, and the model remains relatively stable even when a user’s app environment changes. Professor Uichin Lee framed the vision this way: the smartphone should be viewed not simply as a device for measuring usage time, but as a space in everyday life where an individual’s communication, information seeking, and financial activities are repeatedly carried out. Converting complex smartphone records into human-understandable behavioral units, he noted, provides a foundation for more reliable long-term mental health tracking.

The researchers are careful about the limits of the findings. The study demonstrated associations between behavior and stress, not a causal relationship showing that any particular behavior causes or reduces stress; the models detect a state, they do not explain it. The participant pool, drawn from Android users in their 20s and 30s, also leaves open questions about generalizability, and the team plans to expand the research to broader age groups, occupations, and smartphone environments while verifying long-term stability at larger scales. The study, with Hansoo Lee as first author, Professor Uichin Lee as corresponding author, and Taehyeon Park, Youngji Koh, and Professor Jae-Gil Lee of the School of Computing as co-authors, was published in December 2025 in the international journal Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, and is scheduled to be presented at ACM UbiComp/ISWC 2026 in Shanghai, China, from October 13 to 15, 2026. The work was supported by the Institute of Information and Communications Technology Planning and Evaluation through its ICT Convergence Digital Inclusion Technology Development project and by the National Research Foundation of Korea’s Sejong Fellowship project, funded by the Ministry of Science and ICT. If the approach proves durable across wider populations, the routine taps, searches, and payments that fill an ordinary day could become a continuous, privacy-conscious window into mental well-being, no additional hardware or questionnaires required.

Subject of Research: Detection of daily stress from repeated smartphone interaction behaviors using behavior-task mining and personalized AI models

Article Title: KAIST reads “stress” from smartphone chatting, searching, and payment habits

Article References: KAIST reads “stress” from smartphone chatting, searching, and payment habits. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: KAIST, stress detection, smartphone behavior, digital mental health, machine learning, behavior mining, large language model, UbiComp, privacy, mobile computing, personalized AI, passive sensing

The study, published in Frontiers in Communications and Networks by Usman Tariq of Prince Sattam Bin Abdulaziz University in Saudi Arabia and Kamran Shaukat of Torrens University in Australia, tackles a problem that goes far beyond a single lost drone. In a decentralized swarm, drones continuously share their positions with one another to maintain formation. If even one aircraft is compromised by spoofed navigation data, those false coordinates propagate through the group’s communication links, potentially shattering formation stability, triggering collisions, or steering the entire fleet off its mission path. In dense urban environments like Riyadh, where signal reflections and occlusion create favorable conditions for spoofing dominance, the threat is amplified.

OrbitGuardNet’s detection pipeline fuses three complementary analytical stages. First, a lightweight transformer encoder with four stacked self-attention blocks processes windowed sequences of BeiDou observables, including pseudorange, carrier phase, carrier-to-noise density ratio, and Doppler shift, alongside inertial measurement unit data and inter-UAV ranging measurements. The transformer extracts temporal drift signatures, capturing the gradual position shifts characteristic of drag-off attacks or the intermittent jumps of signal replay. Second, a graph neural network treats the swarm as a network of nodes and edges, aggregating neighbor embeddings so that temporal evidence can be checked against the swarm’s actual communication topology. This spatial reasoning exposes inconsistencies between what a drone reports and what its neighbors observe, catching spoofers that target only a subset of the fleet. Third, a Kalman filter compares the fused estimate against the predicted navigation state, producing residual evidence that flags anomalies beyond adaptive thresholds.

The most distinctive component is OrbitGuard-LLM, a domain-specialized large language model built as a compact decoder-only transformer with roughly 120 million trainable parameters. Rather than serving as a conversational assistant, this model functions as a contextual validator. It ingests a structured token sequence combining the joint transformer-GNN embedding, discretized Kalman residual tokens, and categorical mission descriptors such as geofences, no-fly zones, target waypoints, and formation spacing limits. The model then produces a plausibility score that asks a fundamentally different question from conventional detectors: is the observed motion consistent with the declared mission? In one revealing test case, a meaconing attack produced gradual drift that slipped past threshold triggers, but the LLM flagged the trajectory as implausible because it would carry the drone through a no-fly zone. This constraint-aware reasoning distinguishes benign multipath effects, which manifest as short-lived oscillations consistent with mission intent, from coherent, mission-agnostic spoofing drifts.

Detection alone, however, does not save a swarm. The researchers paired their detector with an active mitigation layer built on trust-weighted consensus control. Each UAV broadcasts anomaly metrics, corrected state vectors, and trust weights derived from spoofing confidence scores. When spoofing is suspected, the control law attenuates the influence of potentially compromised drones, weighting neighbor contributions by trust rather than treating all data equally. Leader selection follows a distributed ranking that multiplies trust by remaining energy, ensuring that a spoofed leader cannot misdirect the group while avoiding disruptive leadership churn in energy-critical missions. A Priority-Aware Reliable Mesh Protocol guarantees that time-critical spoofing evidence and control updates dominate the shared communication channel, using hop-by-hop acknowledgments and controlled flooding when partitions threaten connectivity.

The validation campaign combined high-fidelity simulation with a limited real-world pilot. The simulation modeled a five-square-kilometer urban area modeled on Riyadh’s structures, complete with multipath propagation, signal occlusion, and a Gilbert-Elliott channel model capturing bursty packet losses. Four threat scenarios stressed the system: a static single-point spoofer, a mobile pursuit spoofer mounted on a chasing UAV, multi-transmitter coordinated spoofers broadcasting different falsified positions to separate drone subsets, and time-shift injection attacks that replay genuine signals with delay. Each scenario was repeated ten times with varied noise and attacker timing, and offline training followed a strict five-fold cross-validation protocol to prevent overfitting to any single favorable split.

The results were striking. Across simulated urban missions, the full system achieved a macro-averaged detection probability of 98.1 percent with false alarm rates under 2.5 percent, with per-scenario detection ranging from 97.5 to 97.8 percent. Alerts arrived within approximately 1.8 seconds in simulation and 2.3 seconds in real flights, fast enough to trigger countermeasures before navigation errors spread through the swarm. Formation deviation stayed under 1.5 meters in simulation, roughly 95 percent cohesion, and 2.2 meters in real flights. By contrast, detection-only baselines that stop at alert generation, leaving formation control unchanged, produced deviations exceeding 10 meters and frequent formation breaks under identical spoofing conditions. All simulated missions completed successfully, and 90 percent of field missions achieved their objectives.

Crucially, the researchers are candid about the boundaries of their evidence. No open-air RF spoofing was employed at any point, for ethical and safety reasons. All spoofing signals were generated in software or within RF-shielded testbeds, and the live trials with three quadcopters injected counterfeit navigation data through the interface rather than radiating it. This means the receiver’s RF front-end, tracking loops, and acquisition stages were never exercised by a radiated adversary, so physical-layer effects like capture dynamics, automatic gain control saturation, and loss-of-lock remain outside the current field evidence. The authors explicitly frame their flight results as pilot feasibility validation of the onboard pipeline and coordination logic, not as proof of robustness against real radiated spoofers in open environments.

Computational feasibility emerged as a central design constraint. The core detector path, excluding the LLM, requires about 17.37 MFLOPs per tick per UAV, roughly 173.7 MFLOPs per second at the 10 Hz detector rate, running on an NVIDIA Jetson TX2 module with 8 GB of RAM. Average CPU usage measured 55 percent, drawing approximately 15 watts, well within the platform’s limits. The LLM validator, however, occasionally spiked CPU use to 70 percent, prompting the team to offload part of that inference to a ground station. The core transformer-GNN-Kalman stack remains fully onboard, preserving continuous detection and mitigation triggers even when the ground link is unavailable, so the offload functions as an accelerator rather than a single point of failure.

Compared against recent alternatives, including CNN-based classifiers on cross-ambiguity function maps, LSTM detectors optimized by genetic algorithms, compact tiny deep networks, and transformer-based acquisition-stage detectors, OrbitGuardNet stands apart by combining accurate detection with swarm-wide active response. Prior systems largely operate on single drones, provide detection without countermeasures, or demand GPU acceleration unsuited to embedded platforms. The new framework’s trust reweighting, leader re-election, and adaptive formation scaling represent a qualitative shift from passive alerting to coordinated resilience. Future work will pursue fully onboard LLM inference through compression and quantization, larger and more diverse field trials with regulator-approved controlled interference testing, and expanded attack libraries covering combined cyber-GNSS threats. Until those radiated-signal trials are complete, the reported figures stand as carefully bounded evidence of what is possible when artificial intelligence meets swarm coordination in contested skies.

Subject of Research: AI-based detection and mitigation of BeiDou navigation spoofing in decentralized UAV swarms

Article Title: Mission-aware BeiDou spoofing defense in UAV swarms with LLM-assisted context validation

Article References: Tariq, U., & Shaukat, K. (2026). Mission-aware BeiDou spoofing defense in UAV swarms with LLM-assisted context validation. Frontiers in Communications and Networks, 7, Article 1760543. https://doi.org/10.3389/frcmn.2026.1760543

Image Credits: AI Generated

DOI: 10.3389/frcmn.2026.1760543

Keywords: BeiDou, UAV swarms, GNSS spoofing, large language model, graph neural network, transformer encoder, Kalman filter, drone security, swarm coordination, smart cities, navigation resilience, OrbitGuardNet

The research team, led by Meng Sheng and Yalin Liu of Changde Vocational Technical College in China, together with colleagues at Lingnan University and Changde First Hospital of Traditional Chinese Medicine, built their intervention around a framework known as cognitive apprenticeship. The idea, borrowed from the learning sciences, is that expert thinking should be made visible to novices in stages: the instructor first models how an expert approaches a problem, then coaches students as they attempt it themselves, and finally fades support as competence grows. The researchers layered a generative AI learning agent onto this scaffold, creating a model they call AI-CAS-PBL, and tested it against conventional problem-based learning and lecture-based instruction in a biochemistry unit on lipid metabolism.

The technical heart of the study is the AI agent itself. The team developed it on the Coze platform, powering it with the Doubao 1.5 Pro 32K large language model, a system with a context window large enough to hold an entire clinical case, the relevant metabolic pathways, and a running dialogue with the student. Rather than simply answering questions, the agent was positioned within the cognitive apprenticeship structure to provide one of two complementary forms of scaffolding: it could respond to student questioning at any hour, offering explanations calibrated to the learner’s current understanding, while the human instructor retained the modeling role, demonstrating expert reasoning in front of the class. The design deliberately avoided the most common failure mode of educational chatbots, which is to hand students finished answers that short-circuit the very thinking the curriculum is meant to build.

Methodologically, the study took care to hold everything constant except the instructional approach. The same instructor taught all three groups, and the learning objectives, core content, scheduled contact time, case materials, and assessment procedures were standardized across conditions. What differed was the mode of support: the AI-CAS-PBL group received the AI agent alongside the cognitive apprenticeship structure, the conventional problem-based learning group worked without AI support, and the lecture-based group received traditional instruction. This quasi-experimental design, while not a randomized trial, substantially strengthens the claim that the AI integration itself, rather than differences in teaching quality or content coverage, drove the observed effects.

The results were striking. Students in the AI-CAS-PBL group achieved significantly higher post-test scores on both lipid metabolism knowledge and case-based clinical application analysis than students in either comparison group, with all differences reaching statistical significance at the conventional threshold. In other words, the AI-supported approach did not merely help students memorize metabolic pathways better; it improved their ability to apply biochemical reasoning to clinical scenarios, which is precisely the transfer of learning that problem-based learning was invented to promote and that lectures so often fail to deliver.

Equally important were the findings on how the experience felt to learners. Using the NASA Task Load Index, a validated instrument originally developed for aviation that measures perceived mental demand, effort, frustration, and related dimensions of workload, the researchers found that students in the AI-CAS-PBL group reported significantly lower cognitive workload than their peers in conventional problem-based learning. This matters because cognitive load theory, one of the most influential frameworks in instructional psychology, holds that working memory is a severely limited resource, and that learning fails when instruction imposes demands that exceed it. The AI agent appears to have absorbed some of the extraneous load, the wasted mental effort of searching for information and struggling alone, freeing capacity for the productive kind of load that builds durable knowledge structures.

The psychological benefits extended beyond workload. Students working with the AI agent reported higher self-efficacy, measured with the New General Self-Efficacy Scale, reflecting greater confidence in their own ability to succeed at the task. Self-efficacy is not a soft outcome; decades of research link it to persistence, engagement, and ultimately achievement. A struggling novice who can test half-formed ideas against a patient, endlessly available AI interlocutor, without fear of embarrassment in front of peers, may be more willing to attempt difficult reasoning, and that willingness compounds over a course. The students also rated the agent highly on perceived usefulness and perceived ease of use, the two central constructs of the Technology Acceptance Model, indicating that the tool was not merely tolerated but genuinely embraced.

To probe the learning environment itself, the team drew on instruments tied to the cognitive apprenticeship experience, including the Maastricht Clinical Teaching Questionnaire and measures of critical thinking disposition. The theoretical claim underlying the intervention is that AI-supported questioning and instructor modeling provide complementary forms of instructional support, a dual-scaffolding arrangement in which the machine handles on-demand explanation while the human demonstrates expert strategy. The authors are appropriately careful here: because the study did not directly examine the interaction between AI support and instructor scaffolding, they present this dual-scaffolding framework as a theoretically informed interpretation rather than a confirmed causal pathway. That honesty is a model of how AI-in-education research should be reported, resisting the temptation to overclaim mechanism from outcome data.

The implications reach well beyond one biochemistry unit. Generative AI is flooding into classrooms faster than the evidence base can support, with institutions adopting chatbots on enthusiasm rather than data. This study offers something rarer: a controlled comparison showing that AI integration, when embedded in a principled pedagogical framework rather than bolted onto an existing course, can outperform both the traditional alternative and the reform method it was meant to enhance. The lesson may be that the framework matters as much as the technology. Cognitive apprenticeship gave the AI a defined pedagogical role, preventing it from becoming either a crutch that replaced thinking or an oracle that students passively consulted.

Caveats remain, and the authors name them. The study was conducted at a single institution with a single instructor, and the quasi-experimental design cannot rule out all confounding. The authors themselves call for evaluation in larger, multicenter, and multi-class studies before the model is widely adopted. Questions about long-term retention, transfer to other subject areas, and the effects of AI dependence on independent problem-solving all await future work. Yet as a proof of concept, the study lands at a propitious moment. It suggests a concrete, testable architecture for human-AI collaborative teaching, one in which the machine’s tireless availability and the teacher’s expert modeling are not competing for the same instructional space but reinforcing each other. If larger trials confirm these results, the image of the medical student lost in a case discussion, silently overwhelmed, may give way to something better: a novice with an expert’s questions at their fingertips and an expert’s thinking demonstrated before them, learning biochemistry not by memorizing pathways but by reasoning through them with less strain and more confidence.

Subject of Research: Integrating generative AI into cognitive apprenticeship-based problem-based learning in medical biochemistry education

Article Title: Integrating generative AI into cognitive apprenticeship–based problem-based learning in biochemistry: a quasi-experimental study

Article References: Sheng, M., Gong, Y., Hu, W., & Liu, Y. (2026). Integrating generative AI into cognitive apprenticeship–based problem-based learning in biochemistry: a quasi-experimental study. BMC Medical Education. https://doi.org/10.1186/s12909-026-10504-3

Image Credits: AI Generated

DOI: 10.1186/s12909-026-10504-3

Keywords: medical education, problem-based learning, generative artificial intelligence, cognitive apprenticeship, cognitive load, biochemistry, large language model, self-efficacy, quasi-experimental study, clinical reasoning, NASA-TLX, human-AI collaboration

The study’s central innovation is a new interview format called the Multi-Topic Emotion Interview, or MTEI. Rather than asking patients to complete a resilience scale, the researchers guided them through conversations spanning multiple emotional topics, recording both audio and momentary self-assessments during the session. Between December 2024 and November 2025, the team collected multimodal data from 248 patients diagnosed with major depressive disorder and 101 healthy controls. To test whether their models would hold up beyond the original cohort, they also ran a four-week longitudinal follow-up with 106 participants, creating a temporal validation set that mimicked the real-world challenge of applying a model to new patients weeks after it was built. The study was retrospectively registered with the Chinese Clinical Trial Registry under identifier ChiCTR2500100966 on April 17, 2025, and approved by the ethics committee of Beijing Anding Hospital.

At the heart of the analysis sits an expert-guided large language model, a design choice that distinguishes this work from most speech-based mental health studies. Instead of letting a general-purpose language model freely interpret transcripts, the researchers steered it with expert knowledge to extract two families of features. The first is a Semantic Resilience Index, or SRI, a quantitative score capturing how resilient a patient’s language sounds in terms of coping strategies and adaptive thinking. The second is a three-dimensional emotional profile known in the literature as VAD, standing for valence, arousal, and dominance, which characterizes the emotional tone of what a patient says along dimensions of pleasantness, intensity, and sense of control. These language-derived features were then combined with Self-Assessment Manikin ratings, a validated pictorial instrument in which patients report their momentary emotional state during the interview itself.

To turn these features into a continuous resilience score, the team employed a Voting Regressor ensemble, a machine learning technique that combines the predictions of multiple regression models to produce a more stable estimate. The target variable was the Connor-Davidson Resilience Scale, or CD-RISC, the standard clinical questionnaire for measuring resilience. The best-performing multimodal fusion model, which integrated the semantic resilience index, the VAD emotional features, and the SAM state assessments, achieved an R-squared of 0.555 and a Concordance Correlation Coefficient of 0.705. In practical terms, the model explained more than half of the variance in questionnaire-measured resilience, and its predictions agreed closely with the actual scores patients reported. Those are respectable figures for a behavioral biomarker, a field where effect sizes are often far weaker.

What makes the result scientifically interesting is how well the model traveled through time. When evaluated on the four-week follow-up data from 106 participants, performance barely moved: R-squared of 0.557 and a Concordance Correlation Coefficient of 0.701. In machine learning for medicine, models frequently degrade when applied to data collected later, a phenomenon driven by drift in recording conditions, patient populations, or clinical workflows. The near-identical metrics here suggest that the speech-derived resilience signal is stable rather than an artifact of a particular recording session. For a digital biomarker intended to be used repeatedly across the course of treatment, temporal stability is arguably as important as raw accuracy, and this study provides unusually direct evidence for it.

The researchers then asked a harder question: was the model actually measuring resilience, or was it merely detecting how depressed someone was? Because depression severity and resilience are related, a model could score well on resilience prediction while really just tracking mood. To disentangle the two, the team performed residual analysis, statistically removing the contribution of depression severity and examining what the model still captured. The answer was that the model explained 9.6 percent of the variance in pure resilience residuals, meaning it genuinely decoded resilience-specific information that exists independently of current depressive symptoms. This is the study’s most consequential finding, because it demonstrates that a speech-based system can identify a patient’s coping capacity even when that patient is in the depths of a depressive episode, when self-report instruments are least reliable.

Interpretability, often the Achilles heel of large language model applications in medicine, received explicit attention through SHAP analysis, a technique that quantifies how much each feature contributes to a model’s predictions. The analysis revealed that the most critical predictors were three LLM-derived constructs: agency and control, reflecting the degree to which patients describe themselves as capable of influencing outcomes; cognitive nuance, capturing the sophistication and flexibility of their thinking about difficult situations; and contextual emotional dominance, a measure of perceived control within emotionally charged content. These are precisely the psychological mechanisms that resilience theory says should matter, which lends the model a degree of face validity that purely data-driven features often lack. A clinician looking at the model’s output would not receive an opaque number but a profile grounded in recognizable cognitive dimensions.

Equally notable is what failed. Traditional acoustic features, the pitch, energy, and rhythm measures that have dominated speech-based depression research for decades, did not provide independent incremental validity during late fusion, the stage at which different feature streams are combined. In other words, once the model had access to the deep semantic features extracted by the expert-guided language model, the raw sound of the voice added essentially nothing. This challenges a common assumption in the digital psychiatry field that paralinguistic vocal cues are the primary carriers of diagnostic information. The finding suggests that for resilience specifically, what patients say and how they frame their experiences may matter more than how their voices sound, a conclusion that could redirect research funding and design choices across the field.

The clinical implications are considerable. A scalable, objective resilience measure could change how treatment is personalized in psychiatry. Two patients with identical symptom scores may have very different recovery trajectories depending on their coping resources, and today those differences are largely invisible to standard assessments. A speech-based biomarker could be administered during routine visits, requiring only a short interview and no additional hardware beyond a microphone, and could track whether an intervention is actually strengthening resilience rather than just suppressing symptoms. The authors suggest that the identified predictors, particularly agency and control and cognitive nuance, point to actionable targets for personalized interventions, since these are constructs that psychotherapies such as cognitive behavioral therapy explicitly aim to modify.

Cautions remain before such a tool reaches the clinic. The cohort was recruited at a single Chinese psychiatric hospital, and cross-cultural validation will be essential, since both language use and the expression of coping differ across populations. The follow-up period of four weeks is short relative to the relapsing course of major depressive disorder, and the study measured resilience against a self-report gold standard, albeit one that the model demonstrably exceeded in objectivity. The expert-guided design also raises questions about how sensitive the system is to the specific prompting framework used. Still, the combination of a large multimodal cohort, temporal validation, residual analysis, and mechanistic interpretability makes this one of the more rigorous demonstrations to date that language, parsed intelligently, can carry clinically meaningful psychiatric signal. As large language models continue to mature, the quiet details of how patients talk about their lives may become one of psychiatry’s most informative vital signs.

Subject of Research: Speech-based digital biomarkers of psychological resilience in major depressive disorder using expert-guided large language models

Article Title: Interpretable speech biomarkers of psychological resilience in major depressive disorder via expert-guided large language models

Article References: Wang, S., Zhu, X., Li, N., Yang, Q., Liu, Y., Wang, J., Liu, P., Feng, L., & Wang, G. (2026). Interpretable speech biomarkers of psychological resilience in major depressive disorder via expert-guided large language models. BMC Psychiatry. https://doi.org/10.1186/s12888-026-08584-y

Image Credits: AI Generated

DOI: 10.1186/s12888-026-08584-y

Keywords: major depressive disorder, psychological resilience, digital biomarker, large language model, speech analysis, natural language processing, Connor-Davidson Resilience Scale, SHAP analysis, multimodal fusion, psychiatric assessment, semantic resilience index, machine learning

The motivation comes from a persistent gap in the automated machine learning landscape. Building an accurate deep learning model requires navigating a multi-stage pipeline: formulating the problem, splitting data into training, validation, and test sets, preprocessing temporal signals, selecting architectures such as convolutional and LSTM networks, tuning hyperparameters, and finally deploying the model. Errors at any stage can compromise validity, and most existing frameworks concentrate almost exclusively on the modeling phase, leaving problem formulation and data preparation underexplored. General-purpose tools also fail to capture domain-specific nuances, and their black-box nature hinders adoption in scientific fields where transparency and reproducibility are non-negotiable.

xAutoGA targets multivariate time-series data from wearable and laboratory sensors, a staple of biomechanics and movement disorder research. The software automates participant-level splitting of data into training, validation, and test sets, a strategy shown to give more realistic estimates of how well a model generalizes to new people. It then jointly optimizes the sliding window size and overlap used to segment sensor signals together with the neural network architecture itself, searching across both time-domain and frequency-domain representations. The frequency stream applies a short-time Fourier transform with a Hann window to convert one-dimensional signals into spectrogram-like images, which are processed by a two-dimensional convolutional network, while a parallel one-dimensional network extracts temporal features; the two streams are concatenated before classification.

The optimization engine is a genetic algorithm that encodes the entire pipeline into a fixed-length chromosome. Gene blocks govern data segmentation, with window durations ranging from 0.25 to 10 seconds and overlaps of 25 to 75 percent; convolutional blocks with binary existence flags that allow layers to be included or bypassed; filter counts, kernel sizes, strides, activations, and batch normalization choices; and training parameters including batch size, learning rate, and the loss function itself. To handle imbalanced classes, a common problem in activity recognition, the chromosome can select between weighted cross-entropy and focal loss with tunable class weights. Each candidate model is evaluated by its macro-averaged F1-score on validation data, averaged over three training trials, and populations of 50 models evolve over 30 generations through tournament selection, uniform crossover, and mutation.

Benchmarks against two established automated deep learning systems, McFly and DRAGON, showed that xAutoGA achieved the highest F1-scores on the AReM, WISDM, and Opportunity datasets, with statistically significant improvements over McFly on three of five public datasets according to Wilcoxon signed-rank tests. DRAGON retained the edge on PAMAP2, a shortfall the authors attribute to fixed genetic algorithm settings that were not retuned per dataset. In controlled comparisons of optimization strategies on a search space of roughly 16,000 configurations, the genetic algorithm showed the best convergence stability, with the lowest interquartile range of final scores, and its population-based structure allows parallel evaluation of candidates, unlike the sequential Bayesian optimization baseline.

The team also stress-tested the algorithm with a deliberately deceptive synthetic landscape they call the Synchronization Trap, where only perfectly matched dual-stream architectures reach the global optimum. Here the genetic algorithm’s survival-of-the-fittest mechanism became a liability: tournament selection quickly eliminated risky dual-stream candidates in favor of a mediocre single-stream local optimum, and the algorithm found the true optimum in only 58 percent of runs, compared with 100 percent for random search and 96 percent for Bayesian optimization. The authors note that real biomechanical datasets tend to have more continuous fitness landscapes, which explains why the genetic algorithm excelled in the primary experiments despite this theoretical weakness.

What sets xAutoGA apart is its educational machinery. Hovering over any hyperparameter in the interface reveals its definition, purpose, default values, and literature references, so users learn while configuring experiments. During optimization, the software displays gene domination plots that borrow the logic of natural selection: each model is an organism, its hyperparameters are genes, and hyperparameter values that confer higher fitness visibly spread through the population over generations, much as longer giraffe necks once did. In one example, batch size 64 started as one option among five and gradually dominated the population until it was the only survivor by generation 12, giving users an intuitive picture of why the final model chose what it chose.

Beyond visualization, the system mines its own optimization history. Every model the genetic algorithm produced becomes a data point, with hyperparameter settings as features and F1-score as the label, and a linear programming-based rule learner extracts interpretable if-then rules that separate top-performing from underperforming configurations. On the AReM dataset, for instance, smaller batch sizes, larger windows and overlaps, and shallower networks with fewer neurons distinguished the best models. A locally deployed language model, chosen as Qwen3-4B to keep user data private, then translates these machine rules into plain-language rationales, explaining for example that smaller batches allow more frequent parameter updates. The authors deliberately position the language model as an explanatory agent rather than a search engine, avoiding the computational overhead and hallucination risks seen when language models drive optimization directly.

A usability survey at Montana State University, involving 19 students from biomechanics, statistics, computer science, and other fields plus two machine learning experts, offered cautious encouragement. Most participants rated the visualizations as moderately to very effective at showing how the final model was created and at increasing trust, and System Usability Scale responses revealed strong agreement that the software is well integrated and not cumbersome, though responses were split on how much prior learning is needed to get started. Experts flagged that gene domination plots could mislead if the population converges prematurely, prompting plans for restart mechanisms and diversity alerts. Future work will also validate whether extracted rules transfer across datasets, integrate additional explainability techniques such as SHAP and Grad-CAM, and formally assess how much users actually learn. With the source code publicly available on GitHub, the researchers hope the framework will help biomechanics researchers, clinicians, and engineers build trustworthy deep learning models while understanding exactly what the machine did on their behalf.

Subject of Research: Explainable automated deep learning for multivariate time-series sensor data using genetic algorithms and large language models

Article Title: xAutoGA: an automated deep learning software to educate users

Article References: Mohasel, S. M., & Aghaei, A. A. (2026). xAutoGA: an automated deep learning software to educate users. Multimedia Tools and Applications, 85(10), Article 791. https://doi.org/10.1007/s11042-026-21942-y

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21942-y

Keywords: automated machine learning, explainable AI, genetic algorithm, neural architecture search, time-series classification, large language model, biomechanics, activity recognition, hyperparameter optimization, user education, wearable sensors, deep learning

At the heart of the project is a geospatial artificial intelligence, or GeoAI, framework that weaves together three components that are usually handled separately: traffic forecasting, spatial mapping and a large language model. The forecasting engine projects future traffic volumes from historical observations. The spatial engine maps those projections onto a hierarchical grid so that congestion can be examined at scales ranging from entire boroughs down to individual corridors and hotspots. The conversational layer, built on a large language model hosted locally rather than in the cloud, sits on top of both, allowing a planner to pose ordinary questions and receive answers grounded in the city’s own data. The system was tested using fifteen years of New York City traffic data, a dataset of unusual depth made possible by the city’s extensive open transportation data programs.

The motivation, Rozhkov explains, came partly from conversations with urban planning practitioners. Cities increasingly want to incorporate artificial intelligence into their daily work, but planning departments often lack the specialized technical expertise needed to build and maintain custom systems. There is also a deeper institutional reluctance at play: transportation data can be sensitive, revealing patterns about infrastructure, commerce and movement that agencies may not wish to hand over to commercial models and chatbots. That combination of aspiration and caution shaped the project from the start. The team set out to build a platform that could be hosted entirely on local infrastructure, secure enough for sensitive data, intuitive enough for planners to use comfortably in day-to-day work, and powerful enough to justify the effort of adoption.

Traffic proved an ideal test case, and New York an ideal proving ground. The researchers analyzed traffic observations spanning 2009 through 2024 and benchmarked two forecasting methods against each other. The first was ARIMA, or AutoRegressive Integrated Moving Average, a conventional statistical forecasting technique that has anchored time-series analysis for decades. ARIMA models excel at capturing seasonality, the regular weekly and annual rhythms that make rush hours predictable. The second was a Long Short-Term Memory network, or LSTM, a recurrent neural architecture designed to learn patterns that extend across long sequences. Unlike simpler feedforward networks, LSTMs use gating mechanisms that let them retain information over many time steps, making them well suited to traffic data, where the useful signal is rarely a clean repetition of past cycles.

The benchmark results, reported by Motarwar, who trained the forecasting models, were decisive. On previously unseen data from 2021 through 2024, the LSTM outperformed ARIMA, achieving an average prediction error of roughly 343 vehicles per day compared with about 418 for ARIMA, an improvement of approximately 18 percent. The gap, Motarwar notes, reflects a conceptual difference between the two methods. ARIMA gives you seasonality, which is true but not the whole story; the LSTM picks up the parts of the pattern that do not repeat cleanly, and that is where most of the improvement came from. This distinction matters practically because the non-repeating component of traffic, driven by construction, shifting land use, weather and economic change, is precisely the component that most affects planning decisions.

With the forecasting engine validated, the team turned its projections toward the future. The model projects average daily traffic rising from 12,540 vehicles in 2025 to 19,680 in 2029, a substantial increase that the researchers are careful to frame as a forecast rather than a certainty. Those figures come with prediction intervals wide enough to demand humility, a caveat the authors emphasize deliberately. Forecasting in urban systems is an exercise in bounded uncertainty, not prophecy, and the framework is designed to present projections alongside the ranges in which they might plausibly fall, keeping the limits of the evidence visible rather than burying them beneath a single confident number.

Forecasting alone, however, was only one piece of the puzzle, and arguably not the most novel one. Patil led the spatial component, mapping traffic onto Uber’s H3 system, an open-source hexagonal hierarchical geospatial indexing scheme that has become popular for spatial analytics. Unlike irregular administrative boundaries, H3 divides geographic space into nested hexagonal cells that can be aggregated or subdivided consistently, allowing planners to examine congestion at multiple scales, from borough-wide patterns down to localized hotspots, without the distortions that arise when data is forced into political or census boundaries. Clustering algorithms were then applied to identify areas where heavy traffic repeatedly accumulated over time. The results were instructive: Manhattan emerged as the highest-congestion borough in the analysis, followed by Brooklyn and Queens, but the more useful finding was the temporal consistency of the problem corridors.

That consistency is exactly what makes the spatial layer valuable to practitioners. A citywide average, Patil points out, does not help city planners. What they need to know is which corridors are badly impacted, and those corridors turn out to be consistent year to year. The hexagonal mapping and clustering were designed precisely to expose that structure, converting thousands of individual traffic observations into a spatial picture of where intervention would matter most. When the same hotspots persist across years, they become targets for policy: bus lane additions, signal retiming, freight management or congestion pricing can be evaluated against the specific places where repeated accumulation of traffic is demonstrable rather than anecdotal.

The final component turns those analytical results into something closer to a conversation. Building on the forecast and the congestion maps, the researchers constructed a customized portal using Meta’s LLaMA model connected to a project-specific knowledge base containing forecasts, congestion locations and other traffic information. A planner could ask, for example, how a highway expansion near LaGuardia Airport might affect traffic in Queens, or what could happen if one-way streets in Lower Manhattan were converted to two-way. Because the language model is bound to the project’s own knowledge base, its answers are anchored in the city’s data rather than in generic statistical narratives. A general chatbot, Motarwar observes, gives you a reasonable-sounding generic paragraph about congestion; planners need an answer that comes from their own data. The team frames this as much a data problem as a model problem, a distinction that goes to the heart of how public institutions should think about generative AI.

Crucially, the entire system is designed to run locally. Agencies could keep sensitive transportation information behind their own firewall rather than sending it to an outside AI service, even if commercial models sometimes offer stronger capabilities. That trade-off, capability for control, is one the researchers consider worth making, and it explains why the paper appeared in a special issue devoted to ethical and explainable GeoAI. The team is not claiming the platform can solve congestion on its own, nor has it yet undergone a large-scale trial with working planners. Rozhkov hopes to test future versions with smaller municipalities and other planning datasets. Rather than asking AI to make urban decisions, the researchers envision it as an interface between planners and increasingly complicated datasets, a way of keeping the underlying data, assumptions and decisions under local control, which is ultimately what will benefit local residents.

Subject of Research: A geospatial artificial intelligence framework combining traffic forecasting, H3 hexagonal spatial mapping and a locally hosted large language model to support urban traffic congestion planning.

Article Title: A new AI framework could help cities plan for future traffic

Article References: A new AI framework could help cities plan for future traffic. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: GeoAI, traffic forecasting, LSTM neural network, ARIMA, urban planning, New York City traffic, Uber H3 hexagonal grid, large language model, LLaMA, congestion mapping, Transactions in GIS, explainable AI

The research, authored by Jun Yin of Beijing Cnoce Technologies Co., Ltd., addresses a persistent weakness in interactive intelligent systems. When requirements are complex, vague, or embedded in specialized background knowledge, conventional systems struggle to accurately interpret human needs or intentions. Rather than treating each user request as a one-shot prediction problem, the proposed algorithm treats interpretation as an iterative cycle: the system computes a response, collects real-time feedback about how well that response matched the demand, and continuously refines its behavior according to the demand response structure. The goal is to improve both the intelligent response capability of the system and the experience of the user on the other side of the interaction.

At the foundation of the method is an effective data collection mechanism. The system gathers real-time response information as interactions unfold, building up the raw material needed to evaluate and improve its own performance. In a known finite domain, the algorithm derives a corresponding feature statistical decision, which anchors the interpretation process in measurable properties of the data rather than in guesswork. This statistical grounding matters because human-computer cooperation networks are messy: nodes in such networks are not equally spaced or equally loaded, and the demands arriving at each node differ in both volume and character.

One of the paper’s central technical contributions lies in how it handles that unevenness. The algorithm extracts the load characteristics of non-equidistant nodes in the human-machine cooperation network using the knowledge of a large language model. In practical terms, the language model’s broad knowledge base is used to characterize how computational and communicative load is distributed across nodes that sit at irregular positions in the network topology. This allows the system to understand where demand concentrates, where it is sparse, and how those patterns shift as users interact with the system over time. Load characterization of this kind is essential for any cooperative system that must allocate responses fairly and quickly across many simultaneous users.

The second pillar of the method is attention. The algorithm constructs a knowledge framework for the large language model according to a two-level attention matrix, and then analyzes an iterative demand response strategy for human-machine cooperation in two stages. The two-level attention structure allows the model to weigh different pieces of information at different granularities, deciding which features of a request deserve emphasis and which contextual signals should modulate that emphasis. By organizing the response strategy into two stages, the system separates coarse interpretation from fine-grained refinement, letting each iteration pass progressively sharper approximations of user intent back into the loop.

The experimental results reported in the study offer concrete evidence that the design works as intended. The algorithm is able to remove abnormal data when the threshold is set to 10, and across three datasets the proportion of abnormal data reaches its lowest state under the proposed method. Anomaly removal at a fixed threshold is more than a housekeeping detail; noisy or malformed interaction data can poison the feedback loop that iterative systems depend on, so a reliable filtering stage protects the integrity of every subsequent optimization step. The fact that the algorithm achieves the lowest abnormal-data proportion across all three datasets suggests that the filtering mechanism is robust rather than tuned to a single data distribution.

Performance gains extend beyond data cleaning. After applying the algorithm, both throughput and utilization rate improved significantly, and the response speed became faster. Throughput measures how much useful work the system completes per unit of time, while utilization rate reflects how effectively the system’s resources are kept busy on productive tasks. Improvements in both metrics, combined with faster responses, indicate that the iterative loop does not merely reinterpret requests more accurately but also runs efficiently enough to lighten the operational burden on the underlying infrastructure. For users, faster response speed is the most visible benefit; for system operators, higher throughput and utilization translate directly into capacity savings.

The study situates itself within a growing body of work on human-machine cooperation. Prior research has explored cooperation between human operators and machines in scheduling problems, man-algorithm cooperation in the intelligent design of clothing products across multiple links, and collision-free path planning for industrial robot manipulators that must account for safe human-robot interaction. More recent efforts have examined how autonomous vehicles can assimilate human feedback in reinforcement learning models and how human-centric digital twins can support human-machine collaboration. What distinguishes the new algorithm is its explicit integration of large language model knowledge into the cooperation loop itself, rather than using language models only as peripheral interfaces.

That integration reflects a broader trend. Large language models have already been applied to domain-specific text mining, as demonstrated by materials-science language models built for information extraction, and to collaborative requirements elicitation, where human-machine iterative approaches use language models to help elicit and model software requirements. Natural language processing systems have been deployed to extract structured events from clinical texts and to classify sensitive records in mental healthcare. The new work extends this lineage into the domain of demand response, where the

Subject of Research: An iterative demand response algorithm that integrates large language model knowledge to improve human-computer cooperation in interpreting complex user requirements.

Article Title: Iterative Demand Response Algorithm for Human-Computer Cooperation Integrating Knowledge of Large Language Model

Article References: Yin, J. (2026). Iterative Demand Response Algorithm for Human-Computer Cooperation Integrating Knowledge of Large Language Model. Applied Intelligence, 56(15), Article 436. https://doi.org/10.1007/s10489-026-07428-2

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07428-2

Keywords: demand response algorithm, human-computer cooperation, large language model, iterative optimization, attention matrix, load characteristics, anomaly detection, throughput, utilization rate, user intent, Applied Intelligence, computational intelligence

Traditional fault analysis for space TWTAs depends on manual statistics and empirical judgment. Engineers comb through accumulated fault documentation, match observed symptoms against remembered cases, and reason their way to a probable cause and remedy. This approach suffers from low efficiency, scales poorly as the fault knowledge base grows, and lacks any capacity for knowledge reasoning across the many dimensions of a fault event, such as the mission stage in which a failure occurred, the components involved, the phenomenon observed, the causal analysis, and the correction applied. Artificial intelligence techniques, particularly knowledge graphs and large language models, have shown substantial potential in fault diagnosis in other domains, but until now, research on effectively integrating the two and applying the combination to the specific fault diagnosis of space TWTAs remained unexplored.

Researchers from the School of Electronic Science and Engineering at the University of Electronic Science and Technology of China have addressed this gap. In a study published in Space: Science & Technology, the team proposed a fault diagnosis framework for space TWTAs built on the integration of a knowledge graph and a large language model. The framework consists of three cooperating modules: a named entity recognition module, a knowledge graph module, and a graph retrieval module. Together, these modules implement retrieval-augmented generation functionality. A user submits a fault description in natural language; the NER module extracts the relevant entities; the graph retrieval module automatically generates graph query language to search the knowledge graph; the retrieved contextual information is combined with the original query to form a prompt; and a general-purpose large language model generates the final diagnostic response. Notably, requirements analysis, graph retrieval, information judgment, and answer generation are all accomplished by a single large language model, which enhances the coupling among modules and the overall stability of the framework.

The named entity recognition module is the framework’s first line of interpretation. The research team constructed three types of corpora for data alignment: a component corpus, a stage corpus, and a fault description corpus. Implementation proceeds in three steps: data augmentation, data pair preparation, and model training. Data augmentation enriches the diversity of training data by replacing component name aliases, rephrasing fault descriptions, and adjusting sentence structures, so that a single raw data entry yields multiple augmented variants. The component name, stage, and fault description information are extracted to form the ASSISTANT portion of each example, which is combined with the PROMPT portion to constitute training data pairs. Model training employs QLoRA technology to fine-tune the Qwen2-7B-Instruct large language model. QLoRA freezes the 4-bit quantized pre-trained model and updates only the parameters of low-rank adaptation modules, achieving efficient fine-tuning under limited computational resources. The fine-tuned model automatically extracts three categories of entities from user inputs: stage, component, and fault.

The knowledge graph module defines the structured representation of TWTA fault knowledge. Five entity types are specified: stage, component, fault, analysis, and correction. Five relationship types connect them: causal relationship, corrective relationship, occurrence relationship, involvement relationship, and composition relationship. This schema comprehensively restores the details of real fault events, including the time of occurrence, the components involved, descriptions of the phenomenon, the causal analysis performed, and the corrective solutions applied. By organizing heterogeneous text records into this graph, the framework gains an explicit, queryable model of fault causality that a language model alone, relying purely on parametric memory, could not reliably provide.

Bridging natural language and the graph is the job of the graph retrieval module. The team adopted a two-step self-instruction method to generate training data pairs that map natural language questions to graph query language, enabling automatic translation from a user’s free-text description to a formal query. This alignment between the large language model and the knowledge graph is what allows the full pipeline to run without handcrafted query templates. Once the graph returns relevant substructures, a retrieval-augmented generation mechanism feeds that evidence, alongside the original user query, into the language model, which then produces fault causes and treatment solutions grounded in the curated knowledge base rather than in speculation.

Experimental validation targeted both individual modules and the end-to-end framework. In tests of the named entity recognition module, the QLoRA-fine-tuned Qwen2-7B-Instruct model achieved F1-scores of 97.56 percent for stage entities, 79.99 percent for component entities, and 90.00 percent for fault entities, yielding a macro-average F1-score of 89.18 percent. These results significantly outperformed both the non-fine-tuned version of the same model and GPT-4o. The graph retrieval module attained a hit rate of 89 percent across 100 natural language queries, confirming that the self-instruction alignment reliably converts questions into queries that locate the intended graph content.

Real-scenario testing demonstrated the practical value of the integrated system. Out of 100 fault descriptions, the framework provided responses to 92, and domain experts adopted 87 of those, corresponding to an adoption rate of 94.56 percent. By contrast, the non-fine-tuned baseline model, although it generated responses to all 100 descriptions, achieved only 23 adoptions. The framework also maintained semantic consistency when answering identical questions repeatedly, while the baseline model exhibited persistent instability across trials. In interference experiments introducing three types of noise, the framework achieved accuracies of 97.22 percent under spelling errors, 95.00 percent under distracting sentences, and 100 percent under irrelevant data, and it effectively rejected irrelevant queries, demonstrating robust stability in messy, real-world conditions.

Scale and speed matter for operational decision support, and the larger experiments addressed both. Across a test set of 1,618 data entries, the framework still maintained an accuracy of 88.13 percent, and the average response time was 19 seconds per query. Those figures validate the system as a usable decision-support tool rather than a laboratory curiosity, giving engineers a fast second opinion grounded in a structured fault knowledge base within the time frame of an interactive troubleshooting session.

The research offers an efficient technical solution for intelligent fault diagnosis of space traveling wave tube amplifiers, and its significance extends beyond a single component type. It shows how a domain-specific knowledge graph, aligned with a fine-tuned, resource-efficient large language model through retrieval-augmented generation, can convert decades of unstructured maintenance text into actionable reasoning. For spacecraft operators, the payoff is enhanced on-orbit reliability and improved operational maintenance efficiency; for the broader aerospace community, it provides a template for bringing natural language understanding and knowledge reasoning together wherever scattered fault records currently resist systematic exploitation.

Subject of Research: An intelligent fault diagnosis framework for space traveling wave tube amplifiers that integrates a knowledge graph with a fine-tuned large language model

Article Title: Fault diagnosis framework for space TWTAs by combining the knowledge graph and large language mode

Article References: Fault diagnosis framework for space TWTAs by combining the knowledge graph and large language mode. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: space TWTAs, fault diagnosis, knowledge graph, large language model, Qwen2-7B-Instruct, QLoRA, named entity recognition, retrieval-augmented generation, satellite communications, Space: Science & Technology, on-orbit reliability, decision support