The research, led by Dr. Georgia M. Vasilakis of the UPMC Heart & Vascular Institute alongside Dr. Kaitlyn E. McSurdy, Dr. Kathryn Berlacher, and Dr. Natalie Stokes, is a cross-sectional survey study — a snapshot in time of attitudes and behaviors across a single academic medical community. Cross-sectional designs like this one capture prevalence and perception rather than causation, but they are precisely the instrument needed when a technology has spread faster than the literature describing it. Graduate medical education, the period when medical school graduates train as residents and fellows under faculty supervision, is an especially revealing setting, because it pairs the least experienced clinicians with the most experienced ones and asks how both groups respond to the same unfamiliar technology.

The urgency of the question is hard to overstate. Generative AI systems — large language models that can draft prose, summarize evidence, answer clinical questions, and simulate patient encounters — have penetrated professional life at a pace with few precedents in medicine. Surveys of physicians conducted by the American Medical Association in 2026, cited in the study’s bibliography, document rapidly shifting sentiment toward augmented intelligence in clinical practice. Yet formal guidance has lagged well behind adoption. A 2025 report from the Association of American Medical Colleges found that artificial intelligence curricula remain unevenly implemented across U.S. and Canadian medical schools, meaning many trainees arrive in residency having never received structured instruction in what these tools can do, when they fail, or how their outputs should be verified.

That gap between use and instruction is the study’s central concern. The title’s nautical metaphor — sailing by the stars — is deliberate. Generative AI offers trainees a navigational aid: a quick answer on a busy call night, a draft of a difficult patient message, a summary of a dense guideline document. But stars can mislead. Large language models generate text by predicting plausible word sequences, not by verifying truth, and they can produce confident, fluent, entirely fabricated content — the phenomenon known as hallucination. A resident who consults an AI system at three in the morning may be consulting something that sounds authoritative and is wrong in ways that are difficult to detect without exactly the expertise the trainee is still developing.

The study’s authors situate their work within a growing but still thin body of evidence. Previous research by Cervantes and colleagues, published in the Journal of Investigative Medicine in 2024, examined medical educators’ perceptions of generative AI and found a profession divided — intrigued by the technology’s potential, wary of its risks to rigor and to the development of clinical reasoning. A 2025 review in the New England Journal of Medicine by Abdulnour, Gin, and Boscardin went further, proposing educational strategies for the clinical supervision of AI use itself, arguing that supervising how trainees engage with AI should become as routine as supervising how they prescribe medications or interpret electrocardiograms. The Pittsburgh survey adds empirical ground truth to that theoretical framework by asking what is actually happening in one large academic center.

The stakes of the question extend beyond individual competence into the architecture of medical training itself. Residency is built on graduated responsibility: interns perform supervised tasks, seniors direct teams, attendings hold ultimate accountability. Generative AI scrambles this hierarchy in an unfamiliar way, because it gives the most junior member of the team access to something resembling expert-level output without any guarantee of expert-level accuracy. A 2023 randomized vignette study published in JAMA by Jabbour and colleagues measured this directly, testing whether AI assistance improved diagnostic reasoning for hospitalized patients — and found that the benefits were neither automatic nor uniformly distributed, underscoring that how the tool is deployed matters as much as whether it is deployed at all.

There is also a subtler cognitive risk that the study’s framework acknowledges: the phenomenon of cognitive offloading. Research published in the journal Societies in 2025 by Gerlich examined how reliance on AI tools correlates with diminished engagement in critical thinking, a finding with obvious resonance for a profession whose core skill is precisely disciplined, skeptical reasoning. If a trainee learns to accept an AI-generated differential diagnosis without interrogating it, the technology may quietly erode the very habit of mind that residency exists to build. Conversely, if trainees are taught to treat AI output the way they treat a consultant’s recommendation — as an informed opinion to be weighed, checked, and integrated — the technology could sharpen rather than soften clinical judgment. Which outcome prevails depends heavily on the norms that faculty model and programs codify, which is exactly why measuring both groups’ behaviors matters.

The Pittsburgh team’s design reflects that dual focus. By surveying trainees and faculty within graduate medical education at the same institution, the researchers could capture the full supervision chain: the people most likely to reach for these tools and the people responsible for evaluating the work those tools touch. Portions of the work were presented as a poster at the University of Pittsburgh Graduate Medical Education Connect Conference in February 2026 and as a plenary session at the medical center’s Department of Research Day in May 2026, before the full paper was accepted on August 13, 2026 and published online on August 31, 2026. The authors report no funding and no competing interests, and the study involved no patients or protected health information, sparing it the privacy entanglements that often complicate clinical AI research. In a move that strengthens the study’s transparency, the datasets generated and analyzed are publicly available in the Harvard Dataverse repository, allowing other researchers to interrogate the findings independently.

What makes this study resonant beyond academic medicine is the broader economic and social evidence it aligns with. A widely discussed 2023 field experiment by Dell’Acqua, McFowland, Mollick, and colleagues described the performance of AI on knowledge work as a jagged technological frontier: tasks on one side of an invisible boundary are transformed by AI assistance, while superficially similar tasks just beyond it see little benefit or outright harm. Medicine sits squarely astride that frontier. Drafting administrative text and surfacing literature are tasks where generative models excel; nuanced diagnostic reasoning under uncertainty remains stubbornly human. The problem, as the Pittsburgh survey implicitly frames it, is that trainees and faculty may not agree on where the boundary lies — or may not know it exists at all.

The study’s ultimate contribution may be less a single statistic than a reframing of institutional responsibility. Sailing by the stars served generations of navigators well, but only because the stars’ positions were meticulously cataloged, their small motions understood, and their limitations respected. Generative AI in medical training deserves the same treatment: not prohibition, not blind adoption, but rigorous charting of where it helps, where it misleads, and what competencies trainees must retain regardless of how capable the tools become. The Pittsburgh team — clinicians embedded in the very training environment they studied — have supplied one of the careful navigational fixes the field now needs. As residency programs across the country confront trainees already carrying these tools in their pockets, the question the study poses is no longer whether graduate medical education will engage with generative AI, but whether it will do so with a map or by guesswork.

Subject of Research: Generative AI use among graduate medical education trainees and faculty physicians

Article Title: Sailing by the Stars: A Cross-Sectional Survey Study of Generative AI Use Among GME Trainees and Faculty

Article References: Vasilakis, G. M., McSurdy, K. E., Berlacher, K., & Stokes, N. (2026). Sailing by the Stars: A Cross-Sectional Survey Study of Generative AI Use Among GME Trainees and Faculty. Journal of General Internal Medicine. https://doi.org/10.1007/s11606-026-10751-6

Image Credits: AI Generated

DOI: 10.1007/s11606-026-10751-6

Keywords: generative artificial intelligence, graduate medical education, medical education, residency training, survey study, clinical supervision, physicians, large language models, cognitive offloading, academic medicine, Journal of General Internal Medicine, hallucination

The study, published in Frontiers in Digital Health, tackles a well-known weakness of vision-language models, the class of artificial intelligence systems designed to connect what a machine sees with what it can say. When these models generate radiology reports, they rely entirely on the representations they learned during training. If the relevant contextual information is not captured, the model may produce statements that are incorrect, unsupported, or outright fabricated, a phenomenon researchers call hallucination. Even state-of-the-art systems evaluated by trained radiologists have been shown to generate reports containing missing or false references, a serious problem in a domain where precision of diagnostic terminology matters.

The new framework, developed by Anandhu Saji, Fousia M. Shamsudeen, Chinnu Jacob, Cristina Popirlan, Claudiu-Ionut Popirlan, and D. Jude Hemanth, builds on an established vision-language model called R2Gen, which was designed specifically for medical report generation. The workflow operates in three stages. First, R2Gen processes the input chest X-ray through a ResNet-101 convolutional encoder that converts the image into a dense visual feature vector, and its language decoder produces a preliminary report. Second, a retrieval stage searches an external repository of radiology reports for cases similar to the one at hand. Third, a refinement stage powered by the Zephyr 7B large language model combines the preliminary report with the retrieved textual context and produces the final, polished report.

The retrieval mechanism is the technical heart of the system. The researchers used FAISS, an efficient library for similarity search developed for exactly this kind of vector-based lookup. For every image-report pair in the repository, the visual feature vector extracted by the R2Gen encoder is stored as the searchable vector in the FAISS index, alongside a fixed-dimensional semantic embedding of the corresponding report generated by the sentence transformer model all-MiniLM-L6-v2. At inference time, the query image yields both an image embedding and, via its preliminary report, a text embedding. A joint scoring function based on cosine similarity combines visual and semantic similarity, and the top-k most similar cases are retrieved. Their associated report information is then handed to the refinement model as textual context.

One of the most practical aspects of the design is what it does not require: retraining. The retrieval and refinement stages operate entirely at inference time, meaning the pretrained R2Gen component is left untouched. There is no additional fine-tuning of the underlying report generation model, and the retrieval repository can be extended or updated independently simply by rebuilding the retrieval index when needed. In a field where most recent advances have come from computationally expensive fine-tuning of large language models, this modular, plug-in approach offers a flexible alternative. The authors describe the two specialized modules, a Retrieval Agent and a Report Generation Agent, as functionally specialized components within a predefined pipeline rather than an autonomous multi-agent reasoning system.

To test the framework, the team used two benchmark datasets. IU X-Ray contains 7,470 chest X-ray images and 3,955 radiology reports, each structured into indication, comparison, findings, and impression sections. CheXpert Plus is far larger, with 223,000 image-report pairs from 64,700 patients and annotations for 14 chest conditions. In both cases the data were split 80:20, with the larger portion used to build the FAISS vector database and the remaining 20 percent reserved for evaluation. The retrieval databases were constructed independently for each dataset, patient-level separation was maintained, and the ground truth reports of the evaluation cases were used only for scoring, never as input to the generation process.

The results were evaluated with standard natural language generation metrics, each capturing a different aspect of text quality. BLEU measures the precision of n-gram matches, ROUGE-L focuses on recall-oriented overlap of the longest common sequences, METEOR accounts for stemming and synonym matching, and token-overlap precision, recall, and F1-score quantify how closely generated tokens match the reference. The gains over the R2Gen baseline were substantial. On CheXpert Plus, the baseline achieved a BLEU-4 score of 0.081, while the proposed framework reached 0.2085, with ROUGE-L rising from 0.246 to 0.3622 and METEOR from 0.318 to 0.3917. On IU X-Ray, BLEU-4 climbed from 0.165 to 0.3939, ROUGE-L from 0.371 to 0.5070, and METEOR from 0.187 to 0.5382.

The comparison extended to previously published state-of-the-art methods, including R2GenGPT, R2GenCSR, MambaXray-VL-L, METTransformer, Token-Mixer, and BootstrappingLLM. On CheXpert Plus, the proposed method’s BLEU-4 of 0.2085 and F1-score of 0.4546 exceeded the 0.112 and 0.335 reported for MambaXray-VL-L, the strongest prior result in that comparison. On IU X-Ray, the framework’s BLEU-4 of 0.3939, ROUGE-L of 0.5070, and METEOR of 0.5382 surpassed the 0.206, 0.401, and 0.215 reported for R2GenCSR, while also achieving precision, recall, and F1 values of 0.5997, 0.6976, and 0.6338. Qualitative analysis reinforced the numbers: the generated reports captured key diagnostic elements such as clear lung fields, absence of effusions or consolidations, a normal cardiac silhouette, and intact bony structures.

The work sits within a rapidly growing body of research on retrieval-augmented generation, or RAG, in radiology. A recent scoping review of 45 RAG-based studies found that retrieval consistently improves report generation and question answering performance, but also flagged hallucinations, inconsistent evaluation practices, and limited clinical validation as persistent challenges. Other recent approaches have explored semantically guided retrieval, factual evidence fusion, and multi-agent conversational systems. The present study distinguishes itself less by inventing a new retrieval algorithm and more by its modular organization: it shows that simply pairing an established report generator with inference-time retrieval and a separate language model for refinement can deliver measurable gains without touching the trained model.

The authors are careful about the scope of their claims. Automatic metrics like BLEU and ROUGE measure textual similarity, not clinical correctness, and improvements in these scores should not be read as direct evidence of diagnostic accuracy. The retrieval stage is also only as good as the repository behind it, so the characteristics of the stored reports inevitably shape the context supplied during refinement. Future work, they suggest, should incorporate dedicated clinical factuality measures, extend evaluation to additional datasets and imaging modalities, and explore more clinically informed retrieval strategies. Still, the message is compelling: sometimes the best way to make an AI a better writer is not to train it harder, but to give it a reference library and let it check its work against the accumulated experience of the reports that came before.

Subject of Research: Retrieval-augmented chest X-ray report generation using an agent-based framework combining R2Gen, FAISS similarity search, and Zephyr 7B report refinement

Article Title: Enhanced vision language framework for medical report generation using agent-based retrieval

Article References: Saji, A., Shamsudeen, F. M., Jacob, C., Popirlan, C., Popirlan, C.-I., & Hemanth, D. J. (2026). Enhanced vision language framework for medical report generation using agent-based retrieval. Frontiers in Digital Health, 8, Article 1989756. https://doi.org/10.3389/fdgth.2026.1989756

Image Credits: AI Generated

DOI: 10.3389/fdgth.2026.1989756

Keywords: chest X-ray, radiology report generation, retrieval-augmented generation, vision-language models, R2Gen, FAISS, Zephyr 7B, medical imaging, artificial intelligence, hallucination, natural language processing, health informatics

The research team, consisting of Kanishkavardhan P, Gowtham S, and Maheswari N, built their system around a multi-stage Retrieval-Augmented Generation, or RAG, methodology. RAG is an approach that grounds the outputs of a generative language model in externally retrieved documents, reducing the tendency of such models to fabricate plausible-sounding but incorrect information. In the legal domain, where a single misattributed citation can undermine an entire argument, this grounding is not a luxury but a necessity.

The framework begins when a user submits a legal query. The system first transforms the question into a vector embedding, a numerical representation that captures the semantic meaning of the text rather than just its surface-level words. At the same time, the query is classified into an appropriate legal category, which helps route the request toward the most relevant body of documents. This dual processing of meaning and category allows the system to understand, for example, that a question about property disputes should be directed toward relevant statutes and case law rather than unrelated procedural texts.

Once the query has been classified, the system deploys a hybrid retrieval mechanism that draws on two complementary search algorithms. The first, BM25, is a well-established keyword-based ranking function that scores documents according to the frequency and distribution of query terms. The second, FAISS, short for Facebook AI Similarity Search, is a library for fast similarity search over high-dimensional vectors, enabling the system to find documents whose meanings are close to the query even when they share few exact words. By running both methods in parallel, the framework captures the strengths of lexical matching and semantic understanding simultaneously.

The results from the two retrieval channels are then merged using Reciprocal Rank Fusion, a technique that combines multiple ranked lists into a single consensus ranking. Rather than simply concatenating results, RRF rewards documents that appear high in both lists, effectively filtering out items that only one algorithm found relevant. This fusion step is critical in legal research, where precision matters enormously; surfacing the wrong precedent at the top of a results list can mislead a user who lacks the expertise to evaluate it critically.

After retrieval, the selected documents are passed to large language models that generate structured, formatted answers aligned with the relevant legal citations. The multi-stage prompting design means the model does not simply produce free-form text; instead, it is guided through staged reasoning that ties each claim in the answer back to the retrieved evidence. The authors report that this pipeline yields improved answer relevancy, reliability, and contextual accuracy compared with conventional approaches, positioning the system as a practical tool for legal research and assistance.

A particularly valuable aspect of the study is its comparative evaluation of three prominent language models operating within the same framework. LLaMA-3.3-70B achieved the highest scores in Answer Relevancy at 0.91 and Faithfulness at 0.92, demonstrating an ability to generate accurate, intent-aligned responses that remain firmly grounded in the retrieved evidence with minimal hallucination. Faithfulness, in this context, measures how closely the generated answer adheres to the source documents, a crucial metric for any system expected to cite legal authority.

Gemini-2.0-Flash showed the strongest Context Recall at 0.78, indicating that it was particularly effective at drawing on the retrieved material when constructing its answers. Mistral-Large, by contrast, excelled in BLEU score, reaching 0.769, along with high semantic similarity, reflecting fluent and natural-sounding text generation. However, the authors note that Mistral-Large exhibited weaker evidence grounding, meaning its polished prose was less reliably tied to the underlying legal documents. This trade-off between fluency and faithfulness illustrates a central challenge in applying generative AI to high-stakes domains: a model that writes beautifully but cites loosely can be more dangerous than one that writes plainly but accurately.

The implications for the Indian judicial system are significant. India’s courts face enormous caseloads, and litigants, law students, and even practitioners often struggle to navigate the sheer volume of legal material. A system that can classify a query, retrieve the most pertinent documents through hybrid search, and generate a structured answer with proper citations could dramatically lower the cost and time of legal research. The authors emphasize that the framework is intended as a research and assistance tool, augmenting rather than replacing professional legal judgment.

The study also contributes to the broader field of legal AI by demonstrating how carefully engineered retrieval pipelines can tame the hallucination tendencies of large language models. Rather than relying on a single model or a single search technique, the multi-stage architecture layers semantic classification, hybrid retrieval, rank fusion, and grounded generation into a coherent whole. As generative AI continues to spread into domains where accuracy is non-negotiable, frameworks of this kind, tested and benchmarked with transparent metrics, offer a template for building systems that are both powerful and trustworthy. The research, received in July 2026 and accepted in September 2026, arrives at a moment when courts and legal institutions worldwide are actively debating how AI should be integrated into the practice of law.

Subject of Research: Retrieval-augmented generation for AI-assisted legal research in the Indian judicial system

Article Title: A retrieval-augmented multi-stage prompting framework for legal analysis in the Indian judicial system

Article References: P., K., S., G., & N., M. (2026). A retrieval-augmented multi-stage prompting framework for legal analysis in the Indian judicial system. Frontiers in Big Data, 9, Article 1933830. https://doi.org/10.3389/fdata.2026.1933830

Image Credits: AI Generated

DOI: 10.3389/fdata.2026.1933830

Keywords: legal AI, retrieval-augmented generation, large language models, Indian judiciary, information retrieval, BM25, FAISS, reciprocal rank fusion, legal reasoning, hallucination, LLaMA-3.3-70B, semantic search

The framework, called IPF-RAG, was described this week in Frontiers in Digital Health by researchers from Anhui University of Science and Technology, Anhui University, Shantou University Medical College, Peking University Shenzhen Hospital, and Shenzhen People’s Hospital. It belongs to a family of techniques known as retrieval-augmented generation, or RAG, which addresses a well-known weakness of large language models: their knowledge is frozen at the moment of training. A model asked about antifibrotic therapy may know nothing of the latest guidelines or newly approved drugs, and its generalist training gives it only shallow grip on the subtle, longitudinal phenotypes that define a rare disease like IPF. RAG solves the staleness problem by letting the model consult an external database before answering—but, as the new study shows, applying that idea to clinical medicine exposes deeper problems than most textbook descriptions suggest.

The first problem is time. Electronic health records store clinical events asynchronously: a progress note may be written hours after the examination it describes, and medication orders arrive on their own schedule. Conventional RAG systems retrieve text by semantic similarity alone, which means they can happily surface an anachronistic case or an irrelevant note that shares keywords with the query but belongs to a completely different phase of illness. The second problem is hallucination. Even when good evidence is retrieved, an unconstrained language model may weave it into a confident-sounding plan that no guideline would endorse. The team distilled these difficulties into two questions: how can longitudinal clinical data be structured and retrieved so that disease trajectories, rather than isolated snapshots, are preserved; and how can retrieved evidence be constrained so that the model’s reasoning follows rigorous medical logic?

Their answer begins with data. Drawing on 2,265 patients with interstitial lung disease seen at Peking University Shenzhen Hospital between 2014 and 2024, and applying international diagnostic criteria that require high-resolution CT findings consistent with usual interstitial pneumonia, the researchers curated 58 complete IPF clinical trajectories totaling more than 470,000 words. Each record was formalized in a relational schema keyed by patient identifier, timestamp, and record type, so that every clinical event carries its position in the disease’s chronology. Alongside this knowledge base they built a medication database covering 72 pulmonary fibrosis-related drugs, with dosage, administration routes, and therapeutic effects for each. Raw narratives were cleaned with a hybrid chunking algorithm that combines structural parsing with semantic segmentation, and all data were de-identified before analysis under an ethics approval from the hospital’s research committee.

On top of this foundation sits the framework’s first core component, the Temporal Retrieval Module. It works in three layers. A semantic similarity layer uses a domain-specific text embedding model and the FAISS nearest-neighbor library to find historical notes resembling the current patient’s symptoms. A temporal clinical logic association layer then aligns those notes with the medical orders that followed them—crucially, using temporal proximity rather than exact timestamp matching, since real records are asynchronous. Because timestamps increase monotonically within each patient, the team cast this alignment as a nearest-neighbor search over a sorted array and implemented it with binary search, cutting the localization cost from linear to logarithmic time. A dynamic sliding time-window captures adjacent orders so that retrieved treatment regimens are complete. Finally, a pharmaceutical traceability layer cross-references the retrieved orders against the drug database, attaching precise pharmacological constraints to every recommendation the system will later consider.

The second component, the Structured Enhancement Module, takes the temporally aligned cases and formats them as dynamic demonstrations within a few-shot prompting scheme. Because annotated datasets for rare diseases are small, the researchers chose in-context learning over fine-tuning: each retrieved case becomes an exemplar input-output pair drawn from real clinical records, embedded alongside the target patient’s current status in a structured medical prompt template. This conditions the language model to focus on the decision points that matter and to reason along chains that mirror actual clinical practice. The architecture is deliberately model-agnostic and decoupled, with a plug-and-play input layer that can accept text features extracted from chest X-rays by a separate vision-to-text model—extensibility the team put to the test in their external validation.

The results are striking. In automated evaluations scored by GPT-o3 across six criteria grouped into clinical safety, medical accuracy, and clinical utility, IPF-RAG beat standard RAG with every underlying model tested: DeepSeek-V3 gained 0.43 points on a five-point scale, DeepSeek-R1 gained 0.46, and GPT-o3 gained 0.51, all statistically significant with Holm-adjusted p-values as low as 0.0013. The largest gains appeared in evidence-based medicine and in warnings and monitoring—the dimensions most directly tied to patient safety. Human evaluation reinforced the pattern. In a two-round review by physicians, IPF-RAG raised average scores over the base models by 0.47 to 0.67 points, and in a head-to-head preference test, thirty physicians judged its plans better than standard RAG in 86.7 percent of comparisons and better than a multi-query RAG baseline in 68.3 percent, with only 5 percent judged worse in either case.

An ablation study confirmed that both modules are doing real work. Removing the temporal module dropped the overall score by 0.66 points, with significant declines in risk minimization, medical logic, and personalization—evidence that knowing where a patient sits in the disease’s arc is what makes a plan safe and individualized. Removing the structured module cost 0.62 points, showing that even perfectly retrieved evidence needs logical scaffolding to become actionable. Standard RAG, lacking both, fell 1.12 points below the full system. A sensitivity analysis over retrieval depth revealed another advantage: as more historical chunks were fed to standard RAG, its scores slid monotonically from roughly 3.75 to 3.35 as noise accumulated, while IPF-RAG stayed above 3.9 across every setting, peaking near 4.37 with just a single retrieved case. Temporal structure, it turns out, is itself a filter against information overload.

Perhaps most compelling is the case study the authors highlight. An elderly man presented with two months of worsening breathlessness and oxygen saturation falling to 89 percent—signs of an acute exacerbation. Standard RAG, matching on keywords like chronic cough, retrieved unrelated chronic bronchitis records and generated a pirfenidone prescription that deviated from what the treating physician actually chose. IPF-RAG, tracking the progression to hypoxemia, located a genuinely comparable exacerbation case with its medication timeline, and recommended nintedanib 150 milligrams twice daily along with omeprazole for the patient’s reflux—matching the ground-truth decision and covering the comorbidity. External validation on the public MIMIC-IV database, in which radiology reports were deliberately masked and chest X-ray features were extracted by a bridging vision model, extended the finding across cohorts and modalities: IPF-RAG outperformed standard RAG by 0.50 points on average, with a Holm-adjusted p-value below 0.01.

The authors are careful about what remains to be done. The internal evaluation set is small, the temporal knowledge base is still curated semi-manually, and early failures in semantic matching can cascade into the temporal alignment downstream. Prospective, multi-center clinical trials will be needed before any real-world deployment. But the direction is clear: by teaching retrieval systems to respect chronology and teaching generation to obey medical logic, IPF-RAG offers a template that could extend well beyond one disease—toward any chronic condition where the decisive evidence lives not in a single document, but in the sequence of events that led to today.

Subject of Research: A temporal-aware retrieval-augmented generation framework for clinical decision support in idiopathic pulmonary fibrosis

Article Title: IPF-RAG: temporal-aware retrieval-augmented generation for idiopathic pulmonary fibrosis care

Article References: Zheng, X., Rui, X., Long, T., Wang, L., Lv, Z., Wang, J., Jin, Z., Huang, W., Dong, X., Wang, Q., & Dai, Y. (2026). IPF-RAG: temporal-aware retrieval-augmented generation for idiopathic pulmonary fibrosis care. Frontiers in Digital Health, 8, Article 1883739. https://doi.org/10.3389/fdgth.2026.1883739

Image Credits: AI Generated

DOI: 10.3389/fdgth.2026.1883739

Keywords: idiopathic pulmonary fibrosis, retrieval-augmented generation, large language models, clinical decision support, electronic health records, temporal retrieval, medical AI, hallucination, digital health, few-shot learning, MIMIC-IV, pulmonary fibrosis drugs

The study, published in Frontiers in Communication by Areej Albawardi of Imam Abdulrahman Bin Faisal University and Alyaa Olwi of King Abdulaziz University, is the first to trace the full remediation chain — original advertisement to AI-generated text to AI-generated image — through the lens of multimodal discourse analysis. The researchers selected six advertisements from Almarai, one of the largest dairy companies in Saudi Arabia and among the most valuable companies in the Middle East, recently posted on the company’s official Instagram page with more than ten thousand views each. The ads were chosen precisely because they carry a rich variety of Saudi cultural and social markers: traditional dress, family configurations, gender and age representation, and distinctly Saudi settings.

The experimental design was deceptively simple but technically revealing. Seven AI models — ChatGPT, Gemini, Meta AI, Grok, Qwen, Leonardo AI and Recraft V3 — were each given the same instruction, “turn it into a text prompt,” producing forty-two textual descriptions. Then, using fresh accounts so the models would not rely on stored memory or context, the researchers fed those generated prompts back with the instruction “create an image.” Because Google and Leonardo AI each operate two image-generation submodels — Google Imagen 4 and Gemini 2.5 Flash (Nano Banana), and Leonardo’s Phoenix 1.0 and Lucid Origin — the image stage involved nine generators, yielding fifty-four images. Data collection took place between October 1 and November 1, 2025, using models ranked on Artificial Analysis’ Text-to-Image Leaderboard for accessibility and performance.

To ensure the qualitative judgments had a reliable empirical basis, the two authors independently coded each output against a shared framework, recording whether each model, advertisement and semiotic feature fulfilled a criterion. The coding produced 906 paired decisions — 420 for the image-to-text stage and 486 for the text-to-image stage. Observed agreement reached 91.94 percent, and Cohen’s kappa, which corrects for chance agreement, was 0.8319 — within the “almost perfect” band of the Landis and Koch categories. Alternative coefficients, including Krippendorff’s alpha, Scott’s pi and Gwet’s AC1, all exceeded 0.82, confirming that the coding framework was applied consistently in both directions of transformation.

The analytical framework drew on social semiotic technology, an extension of multimodal discourse analysis rooted in Halliday’s socio-semiotic approach and Kress and Van Leeuwen’s grammar of visual design. Rather than treating AI as a neutral tool, the researchers viewed the models as semiotic technologies that actively shape meaning — echoing Lucy Suchman’s observation that AI works as a “disclosing agent for assumptions about the human.” The analysis identified recurring semiotic resources across the data: color, setting and location, people, action, cultural artifacts, language, intertextuality, logos, and modality, the degree of realism an image communicates through lighting, saturation and filters.

The results split cleanly along a technical fault line. Features that are visually salient and semantically explicit — color, people, action, setting — were transferred with high fidelity. Eight of the nine image generators produced correct colors, and most models accurately rendered the number, gender and activities of the people described in the prompts. Action verbs such as “placing,” “smiling” and “handing,” enriched with adverbs like “gently,” translated reliably into the final images. But features that are implicit, layered or linguistically encoded collapsed in transit. Camera angle and gaze direction were rarely described in the prompts, so generated images showed people looking the wrong way. Screen layering — the advertisement’s graphic plane superimposed on the photographic scene — was misread by several models, which rendered the Almarai sign as if it were physically hanging inside a supermarket rather than added as a logo.

Cultural representation proved the most consequential failure. The researchers found that AI models operate on concentric geographic circles — Muslim, Arab, Middle Eastern, Gulf, Saudi — and dress is represented or misrepresented according to the word chosen. Traditional Saudi clothing in the original advertisements was frequently replaced with Gulf, Palestinian or Pakistani garments. One model described “checkered keffiyehs” as “traditional Saudi Arabian attire,” although that specific garment belongs to Palestinian culture — a contradiction the authors link to what Zoe Hurley has called “image hallucination,” a distortion of reality in which cultural norms are misrepresented. In one generated image, a woman wears a red-and-white shemagh, a headpiece normally worn by men. Race was inferred from dress alone, so when the prompts lacked explicit racial description, models substituted Asian features for Saudi ones.

Language was an equally stark marker of the technology’s limits. Arabic script — the brand name in the logo and the Arabic slogan — was the lowest-scoring feature across all models. Most generators either removed the Arabic text entirely or filled the linguistic void with meaningless gibberish. Only ChatGPT correctly included Arabic textual elements in nearly all of its generated images, and only Qwen3-Max provided an accurate translation of the Arabic text while identifying the language as Arabic. Logos fared little better: none of the nine image generators reproduced the Almarai logo correctly, and several omitted it even when the prompt described it. Notably, Qwen3-Max produced the best textual prompts — averaging 159 to 303 words, rich in cultural and intertextual detail — yet failed to generate accurate images, while ChatGPT, whose prompts lacked detail, produced the best images, possibly because the model stores the input image in its database or draws on additional internet input.

The study also introduces a new conceptual tool: “text hallucination,” derived from Hurley’s taxonomy of image hallucination. Where image hallucination distorts visual reality, text hallucination describes inaccurate textual descriptions of objects or actions in uploaded images — as when one model described a man “holding a hard yellow hat” that does not exist in the original advertisement, which then materialized in the generated image. The authors identify three types: textual, cultural-perspective, and spatial, the last involving misplaced objects and anatomically distorted faces. Hallucinations appeared in the majority of models even when the textual prompts contained no misinformation, suggesting that the image generators themselves add, alter or misrepresent visual content. Only ChatGPT and Google Imagen produced images without any hallucinations.

The implications reach beyond the laboratory. Generative AI is projected to become a 1.3 trillion dollar market within a decade, and advertising agencies are already deploying AI-generated imagery in campaigns for brands like Casablanca, Valentino and Moncler. Advertisements do not merely communicate marketing messages; they create cultural experiences that carry social and ideological perspectives. In a market like Saudi Arabia, where companies deliberately ensure that advertising adheres to local traditions and represents culture accurately, the technology’s inability to render Arabic script, correct logos and authentic cultural dress is not a cosmetic flaw but a functional barrier. The authors argue that their findings carry weight for theories of multimodality and culture, and practical weight for AI developers, who must pay closer attention to scripts, languages and cultural specificity. As the models continue to evolve, the study offers a benchmark: fluency in color and composition is no longer enough. The next frontier for generative AI is cultural precision — the ability to see, and to show, the difference between an Arab advertisement and a Saudi one.

Subject of Research: AI and multimodal communication

Article Title: Cultural and multimodal realization in various AI models: text to image, image to text

Article References: Albawardi, A., & Olwi, A. (2026). Cultural and multimodal realization in various AI models: text to image, image to text. Frontiers in Communication, 11, Article 1959385. https://doi.org/10.3389/fcomm.2026.1959385

Image Credits: AI Generated

DOI: 10.3389/fcomm.2026.1959385

Keywords: generative AI, multimodality, advertising, cultural representation, text-to-image, image-to-text, hallucination, semiotics, Cultural, multimodal, realization, various

The core problem RAG addresses is well understood. Because large language models rely on static parametric knowledge, they suffer from knowledge staleness, hallucinations, limited domain adaptability, weak traceability, and high updating costs in knowledge-intensive scenarios. Retraining a model to refresh its facts is expensive and slow, and even the largest models cannot guarantee that every generated sentence is grounded in verifiable evidence. RAG sidesteps this by pulling relevant documents from an external knowledge source at inference time and feeding them into the generation process, improving factual grounding, knowledge timeliness, and evidence traceability without touching a single model weight.

The new survey, led by Xiaoling Shu, Xiao Zhang, Ruixin Wang, Ziyi Yang, Xiaonan Zhao, and Shujie Han, organizes the field along three major dimensions: the system stack and paradigm evolution of RAG, evaluation methods and application evidence, and practical deployment and security challenges. Using a structured and transparent review protocol, the authors screened studies that proposed, improved, evaluated, or applied RAG-related methods, excluding work that merely touched on prompting or fine-tuning without retrieval or external evidence use. The result is a taxonomy-oriented, evaluation-aware, and deployment-conscious reference intended to give researchers and practitioners a systematic map of a field that has become increasingly fragmented across architectures, strategies, benchmarks, and risks.

At the heart of every RAG system lies the retriever, and the survey dissects the mathematical machinery behind it with unusual rigor. Sparse retrieval methods such as BM25 score documents through term frequency and inverse document frequency, rewarding exact lexical matches while remaining efficient and interpretable. Dense retrieval takes a different route, encoding queries and documents into continuous vectors with dual encoders and ranking them by dot product or cosine similarity, which excels at semantic matching when the query and the relevant document use different wording. Hybrid retrieval fuses both signals through weighted combinations with score normalization, a pattern the authors identify as especially valuable in enterprise question answering, legal documents, and technical manuals where both precise terminology and semantic relevance matter.

Beyond these foundations, the survey catalogues a family of retrieval-enhancement patterns that mark the field’s maturation. Graph-based retrieval represents knowledge as entities, edges, and relation types, scoring candidate paths by relation relevance, node relevance, and path coherence to support multi-hop reasoning over knowledge graphs. Iterative retrieval refines queries through multiple retrieval-generation cycles; recursive retrieval decomposes complex questions into sub-queries; adaptive retrieval adjusts retrieval depth based on intermediate feedback; and self-corrective retrieval evaluates its own evidence and regenerates when quality falls short. Neuro-symbolic retrieval combines neural semantic matching with rule-based constraints, while retrieval-augmented reasoning fetches evidence for intermediate reasoning steps rather than retrieving once before generation. Together, these templates show how retrieval has shifted from a one-shot preprocessing step to an active, dynamic component of reasoning itself.

The survey frames the field’s evolution as a progression from Naive RAG to Agentic RAG. Early systems simply retrieved passages and appended them to the prompt, emphasizing context relevance, answer relevance, and faithfulness. Later methods grew modular, graph-enhanced, self-corrective, multimodal, and increasingly agentic, with capability orientations shifting toward information integration, noise robustness, negative rejection, counterfactual reasoning, and deployment-oriented control. A central theoretical insight formalized in the survey’s appendix captures why retrieval quality dominates: answer quality is bounded by a function of retrieval recall, evidence use, and generation ability, meaning that even a powerful generator will fail if the retriever supplies insufficient, irrelevant, or unreliable evidence.

Measurement receives equally careful treatment. The survey assembles a multi-dimensional evaluation framework spanning retrieval metrics such as Precision@K, Recall@K, Hit Rate@K, mean reciprocal rank, mean average precision, and normalized discounted cumulative gain; generation metrics including exact match, accuracy, BLEU, and ROUGE-L; and semantic similarity measures like BERTScore and cosine similarity. Crucially, it goes beyond surface overlap to faithfulness and attribution metrics that check whether generated claims are actually supported by retrieved evidence, alongside robustness measures for noise, counterfactual evidence, and the ability to refuse unanswerable questions. Deployment-oriented metrics covering latency, retrieval cost, and end-to-end cost round out the framework, reflecting the reality that a RAG pipeline’s expense spans embedding, indexing, search, reranking, generation, verification, and tool calls.

Stress-testing emerges as a recurring theme. The survey examines evaluation protocols designed to probe whether systems remain reliable under noisy, incomplete, contradictory, or unanswerable conditions, and it synthesizes benchmark datasets covering open-domain question answering, multi-hop reasoning, fact verification, dialogue, summarization, information extraction, structured knowledge access, domain-specific QA, and dynamic evaluation. The authors caution that because different studies use different datasets, metrics, and experimental settings, their capability mapping of representative methods should be read as a qualitative synthesis rather than a strict statistical ranking, particularly for very recent work where only preprints may be available.

The survey’s most sobering contribution may be its treatment of security. As RAG systems are deployed in real environments, they inherit a distinctive attack surface: prompt injection can hijack the generation process, data poisoning can corrupt the knowledge base, retrieval hijacking can manipulate which evidence surfaces, privacy leakage can expose sensitive documents, and citation laundering can lend false credibility to fabricated claims by attaching plausible-looking references. By cataloguing these emerging risks alongside defensive mechanisms, the survey makes the case that RAG security deserves the same systematic attention as RAG capability, a point that will resonate as enterprises move from pilots to production.

For a technique that began as a pragmatic workaround for static model knowledge, retrieval-augmented generation has grown into a defining architecture of applied artificial intelligence, and this survey captures that trajectory with rare breadth and precision. By connecting paradigm evolution, component-level analysis, benchmark synthesis, application distribution, evaluation frameworks, and deployment and security analysis in a single coherent reference, the work offers both newcomers a structured entry point and practitioners a checklist of what mature RAG systems must now demonstrate: not just fluent answers, but grounded, traceable, robust, and safe ones. As the boundary between retrieval and generation continues to blur, the questions the survey raises about evaluation, cost, and trust are likely to define the next chapter of the field.

Subject of Research: Retrieval-augmented generation architectures, evaluation, and security in large language models

Article Title: From knowledge retrieval to generation: a comprehensive study of RAG's architectures, evolution, and applications

Article References: Shu, X., Zhang, X., Wang, R., Yang, Z., Zhao, X., & Han, S. (2026). From knowledge retrieval to generation: a comprehensive study of RAG's architectures, evolution, and applications. Artificial Intelligence Review. https://doi.org/10.1007/s10462-026-11716-1

Image Credits: AI Generated

DOI: 10.1007/s10462-026-11716-1

Keywords: retrieval-augmented generation, large language models, RAG evaluation, agentic RAG, hallucination, knowledge retrieval, dense retrieval, BM25, prompt injection, data poisoning, knowledge-intensive tasks, Artificial Intelligence Review

The research, conducted by Mei Chen and Shibin Wang, tackles a blind spot in how we evaluate medical AI. Most assessments of large language models in health communication focus on hallucinations—the invention of facts that were never in the source text. That focus is understandable, since fabricated claims are dramatic and easy to spot once you know to look. Chen and Wang argue, however, that the opposite error deserves equal attention: the quiet, systematic omission of knowledge during text simplification. A summary that leaves out a drug’s key risk factor or a disease’s underlying mechanism does not announce its own incompleteness. It simply reads as complete.

To measure this hidden loss, the researchers developed a method they call knowledge triplet comparison, or KTC. The approach borrows a structure from knowledge graphs, the databases that represent factual information as connected triples of entities and relations. In the medical domain, a triplet might link a disease to its cause, a chemical to a disease it triggers, or a condition to one of its complications. By extracting such triplets from a source text and from the model’s adapted version, and then comparing the two sets, KTC turns the fuzzy question of whether meaning survived the rewrite into a countable, quantifiable measurement. Knowledge that appears in the original but not the adaptation counts as omitted; knowledge that appears only in the adaptation counts as added; and triplets that are altered between the two versions register as distortions of their own.

The team put this method to work on the Chemical-Disease Relations dataset, a benchmark corpus in which sentences from biomedical literature are annotated with relationships between chemicals and diseases. Six large language models were asked to adapt the material across eight different scenarios, each simulating a real-world use case: simplifying text for readers at a middle school reading level, shortening it, lengthening it, and other transformations of style, length, and audience. For each combination of model and scenario, the researchers measured how much of the original knowledge survived and how the surviving knowledge was distributed across categories of medical content.

The headline finding is stark. Across every adaptation scenario and every model tested, the systems retained only about half of the medical knowledge present in the source text. In other words, a patient who relies on an AI-generated plain-language version of a medical abstract is, on average, receiving roughly fifty percent of the factual content of the original. At the same time, the models inserted new material that was not in the source, adding roughly ten percent more knowledge than the original contained. This added fraction is the familiar territory of hallucination risk, and it is the smaller of the two problems the study documents.

The losses were not random, either. When the researchers categorized the omitted and distorted knowledge, a clear pattern emerged. Knowledge about etiology and pathological mechanisms—the explanations of why and how a disease develops—was among the most frequently lost categories, as was knowledge about complications and risk factors. These are not trivial details. Mechanistic understanding helps patients grasp why a treatment works or why a behavior matters, and risk-factor information often underpins prevention and early detection. A simplified text that systematically sheds exactly this kind of content may remain grammatically fluent and factually accurate in every sentence it keeps, while still leaving the reader dangerously under-informed.

Perhaps the most sobering result concerns the attempted fixes. The researchers tested whether targeted prompt engineering—explicitly instructing the models to preserve key knowledge, to avoid omissions, or to stay faithful to the source—could reduce the problem. It could not. Neither the quantity of knowledge lost nor its categorical distribution changed significantly in response to the corrective prompts. The distortion, in other words, is not a quirk of careless prompting but a persistent and regular property of how these models perform simplification. The authors suggest this regularity is itself valuable: if omissions follow predictable patterns, future research can characterize those patterns precisely and build detection and correction systems around them, rather than treating each summary as an unpredictable black box.

The implications reach well beyond the laboratory. Patient-friendly adaptation of medical literature is one of the most socially valuable applications proposed for language models, promising to close the gap between specialist publications and the general public. Hospitals, health websites, and patient advocacy organizations are already experimenting with AI tools that translate clinical jargon into accessible language. The study’s findings do not mean such tools are useless, but they do mean that fluency and readability are poor proxies for fidelity. A summary can score well on every readability metric while discarding half of what the source actually said, and current evaluation pipelines—built largely around hallucination detection and style assessment—would never flag it.

Chen and Wang frame their contribution as groundwork. By quantifying knowledge distortion systematically and showing that it follows stable, category-level patterns, the study lays the foundation for research into why models lose particular kinds of knowledge during simplification and how that loss might be prevented. The KTC method itself offers a template: rather than asking whether a summary sounds right, developers and regulators can ask, triplet by triplet, what survived the transformation. Until such safeguards mature, the study’s message to anyone using AI to make medicine readable is a cautionary one—the greatest danger of a simplified medical text may not be what it gets wrong, but everything it silently leaves out.

Subject of Research: Knowledge distortion in large language models during patient-friendly adaptation of medical literature

Article Title: Prevalent knowledge omissions: An unseen risk of LLMs in patient-friendly adaptation of medical literature

Article References: Chen, M., & Wang, S. (2026). Prevalent knowledge omissions: An unseen risk of LLMs in patient-friendly adaptation of medical literature. PLOS Digital Health, 5(10), e0001739. https://doi.org/10.1371/journal.pdig.0001739

Image Credits: AI Generated

DOI: 10.1371/journal.pdig.0001739

Keywords: large language models, medical literature, knowledge distortion, patient communication, hallucination, health informatics, text simplification, knowledge triplets, prompt engineering, PLOS Digital Health, AI safety, biomedical NLP

The study, published in Frontiers in Public Health, evaluated ChatGPT, DeepSeek, Perplexity, and Microsoft Copilot against 108 standardized heart failure questions, generating 432 responses that were then dissected by a team of clinicians and researchers. Heart failure affects approximately 64 million people worldwide and demands sustained patient education about medications, diet, fluid restriction, and symptom monitoring. The stakes for getting this information right could not be higher, because poor health literacy in heart failure patients is already linked to worse knowledge, poorer self-care, more hospitalizations, and higher mortality.

The research team, led by investigators from the General Hospital of Northern Theater Command in Shenyang, China, and several Chinese medical institutions, designed a rigorous testing protocol. Each of the 108 questions, drawn from a previously validated patient education question set and categorized into basic knowledge, management, and other heart failure-related issues, was submitted to each chatbot using an identical prompt template. The questions were reviewed by two cardiologists and a heart failure specialist nurse to ensure clinical appropriateness. All queries were conducted between August 1 and 5, 2025, from the same workstation, network, and operating system, with ChatGPT and DeepSeek tested with web browsing disabled and Perplexity and Copilot using their default online retrieval modes.

To assess content quality, three independent raters used two established patient-information instruments: the modified DISCERN scale, which measures reliability and clarity of treatment information, and the Ensuring Quality Information for Patients tool, which evaluates organization, understandability, and actionability. The raters were blinded to which chatbot had produced each response, with answers anonymized and presented in randomized order. Inter-rater reliability was strong, with intraclass correlation coefficients of 0.87 for DISCERN and 0.83 for EQIP, and the modified DISCERN instrument showed good internal consistency with a Cronbach’s alpha of 0.87.

The results revealed a clear frontrunner. DeepSeek consistently achieved the highest content quality scores across all three question domains, with mean modified DISCERN scores of 3.73 in the basic knowledge domain, 4.36 in management, and 4.43 in other heart failure-related issues. Perplexity consistently scored lowest, with EQIP percentages falling into the poor quality range in two of three domains. According to the predefined EQIP thresholds, ChatGPT, DeepSeek, and Copilot all achieved good quality with minor deficiencies across all domains, while no system reached the excellent quality category in any domain.

Readability told a more complicated story. In the primary analysis using one calculator, Perplexity generated the most linguistically complex text, with Simple Measure of Gobbledygook scores reaching 11.98 in the basic knowledge domain, well above the sixth to eighth grade reading level generally recommended for patient education materials. However, when the researchers ran a sensitivity analysis using a second calculator with different algorithms for tokenization, syllable counting, and rounding, the rankings shifted, with Copilot showing the highest complexity scores instead. This discrepancy underscores a critical methodological point: readability estimates depend heavily on the specific formula and calculator used, and no single number can capture how easily a patient will actually understand AI-generated health advice.

The most alarming findings came from the hallucination audit. Two independent reviewers with cardiovascular surgical training cross-verified every one of the 432 responses against the 2022 American Heart Association, American College of Cardiology, and Heart Failure Society of America guideline and relevant clinical evidence. A response was classified as hallucinated if it contained at least one clinically meaningful statement that was guideline-inconsistent, contradicted established evidence, was unsupported, was overgeneralized beyond the relevant patient population, or could mislead patient education. The results were sobering: DeepSeek produced hallucinations in 5.56 percent of responses, ChatGPT in 7.41 percent, Copilot in 9.26 percent, and Perplexity in 14.81 percent.

Statistical analysis confirmed these differences were not due to chance. Cochran’s Q test showed a significant overall difference among the four systems, and a generalized estimating equation logistic regression model accounting for question-level clustering found that Perplexity had nearly three times the odds of producing a hallucinated response compared to DeepSeek. Pairwise comparisons after Bonferroni correction showed that the differences between ChatGPT and Perplexity, and between DeepSeek and Perplexity, remained statistically significant. Encouragingly, no high-severity hallucinations were identified; all hallucinated responses were classified as low or moderate severity, meaning none would reasonably encourage unsafe action or delay emergency care, though some could plausibly influence medication use or self-management decisions if followed without clinician review.

The researchers emphasize that their findings should be interpreted as a time-specific evaluation rather than a permanent ranking, given how rapidly these systems evolve. They also note important limitations: the single-turn testing design did not replicate real-world patient interactions, the platforms differed in their retrieval capabilities, and the study did not measure actual patient comprehension or clinical outcomes. Because the internal training data and retrieval mechanisms of these systems were not accessible, the team could not determine whether observed differences reflected model architecture, source selection, or underlying data bias.

What emerges from this study is a clear warning wrapped in cautious optimism. AI chatbots are not going away, and their potential to supplement heart failure education is real, particularly for patients who struggle to access reliable information. But the findings make equally clear that fluency, organization, and apparent completeness are not proxies for accuracy. Every system produced at least some hallucinated content, and even the best performer stumbled in roughly one in eighteen responses. Until these tools are validated in real patients and clinical workflows, the researchers conclude that they should be regarded as information-generating aids rather than trusted sources, with guideline-aligned verification and clinician oversight remaining non-negotiable before any patient-facing deployment.

Subject of Research: Comparative evaluation of four AI chatbots for heart failure patient education quality, readability, and hallucination rates

Article Title: Comparative quality, readability, and factual reliability of four publicly accessible chatbot systems for heart failure patient education: a cross-sectional evaluation

Article References: Qin, N., Shao, X., Wang, J., Lei, J., Wen, R., Wu, W., Liu, F., Wang, H., & Zhu, Y. (2026). Comparative quality, readability, and factual reliability of four publicly accessible chatbot systems for heart failure patient education: a cross-sectional evaluation. Frontiers in Public Health, 14, Article 1966612. https://doi.org/10.3389/fpubh.2026.1966612

Image Credits: AI Generated

DOI: 10.3389/fpubh.2026.1966612

Keywords: artificial intelligence, chatbots, heart failure, patient education, hallucination, large language models, health literacy, readability, clinical guidelines, digital health, medical misinformation, DeepSeek

The central problem the team set out to solve is hallucination, the well-documented tendency of language models to produce fluent but false answers. Research cited in the study, including a Nature paper showing that larger, more instructable language models become less reliable over time, underscores why this matters in a domain where a wrong answer can carry serious operational consequences. The researchers’ solution is architectural rather than statistical: instead of letting the model generate answers from internal knowledge, they make it act as a decision-making layer between the user and a library of nine aviation-validated tools, from trajectory prediction models to FAA document retrieval and an air traffic control simulator.

When a query arrives, the system first interprets the user’s intent, extracting key entities such as airport names, weather conditions, and traffic management initiatives. It then selects the appropriate tool or combination of tools, fills in required inputs from defaults or auxiliary plugins, executes the model, and synthesizes the output into a natural-language response that explicitly references the data sources, models, and assumptions used. If the user asks, for example, for the best runway configuration at JFK right now, the system calls an FAA API to fetch current and forecast weather and feeds that data into NASA’s rule-based runway configuration model, then explains the result with full transparency about what was used and what was assumed.

The nine-tool arsenal is a who’s who of operational aviation expertise. A machine learning model estimates aircraft performance parameters such as drag coefficients and weight from historical flight records to sharpen trajectory predictions. A rule-based model anticipates runway configuration changes driven by wind and closures. A DeBERTa-based natural language processing pipeline extracts traffic management decisions from FAA teleconference transcripts with near human-level F1 scores. Additional plugins tap the FAA’s ATCSCC operations plan advisories, an aviation route API, live METARs and TAFs from aviationweather.gov, NASA’s cloud-based live operational data services streaming thousands of flight updates per second, retrieval-augmented access to FAA manuals, and the open-source BlueSky air traffic control simulator.

To test whether an LLM could reliably navigate this constrained landscape, the team built a dataset of 1,000 manually drafted and verified aviation queries spanning five categories. Nearly 60 percent were straightforward single-tool queries, but the rest deliberately probed failure modes: 17.6 percent omitted required parameters, 10.7 percent presented edge cases or fictional scenarios, 7.3 percent phrased questions to resemble the wrong tool, and 4.6 percent required no tool at all. Aviation experts collaboratively established ground-truth labels, adopting the more conservative interpretation whenever they disagreed, and an independent professional drafted a separate 50-question held-out test set after prompt tuning was complete.

The results were striking. GPT-4o achieved an average correct selection rate of 95.2 percent and input selection accuracy of 96.1 percent across all nine tools, with several tools, including entity recognition, reaching a perfect 100 percent. The smaller GPT-4o-mini followed at 91.5 and 93.7 percent respectively. On the independent held-out set, the system scored 94 percent, suggesting the prompt engineering process did not overfit to the development data. The most persistent weakness was the weather tool, at 86.3 percent, where the models occasionally called the tool for general meteorological questions that should have been answered directly or declined.

Perhaps the most practically useful finding concerns when a smaller, cheaper model suffices. For tools with few operational constraints, such as weather, routes, live data, and FAA manuals, the performance gap between the two models was a negligible 0.8 to 1.5 percentage points. But for tools enforcing hard boundaries, like the runway configuration model that supports only three airports, the gap ballooned to 15.8 points, because GPT-4o was far better at declining out-of-scope queries while GPT-4o-mini tended to call the tool regardless. The authors suggest that lighter models may be viable for open-ended tool calling, while full-size models are preferable where scope enforcement is critical.

The team also ran a revealing baseline comparison against a plain GPT-4o without tool access. For real-time data sources, the baseline correctly declined everything, having no way to see current advisories or weather. Given identical raw flight data, it failed on structured queries like counting flights above a certain altitude or filtering by wake turbulence category, while the tool-grounded system answered correctly. Only on static regulatory questions did the baseline win on raw accuracy, at 95 percent versus 74 percent, but it produced one factual error by misidentifying the hijacking transponder code, precisely the kind of confident hallucination the grounded system avoids by returning no-information-found responses when content is absent from its corpus.

Multi-tool scenarios showcase the system’s fluidity. When a user asked about current weather and then the best runway configuration at an airport, the system retained the weather data in memory and fed it into the runway model. In another case, it fetched the latest FAA advisory affecting Boston, then used that context to recommend optimal routes from Orlando. When a user requested aircraft performance parameters for an Airbus A319 arriving at Dallas-Fort Worth, the system proactively retrieved weather first, then invoked the trajectory model, listing all its input assumptions. In a final demonstration, it translated an airline shorthand query into a live database count of airborne United flights, then instantiated one of those flights in the BlueSky simulator with live data and issued descent commands.

The work required substantial human effort: early tools needed more than 20 revision cycles of expert feedback before prompts stabilized, though later tools converged far faster. The authors are candid about limitations, noting that the 1,000-query dataset is modest, that the system cannot extrapolate beyond hard tool constraints, and that only GPT-based models were tested. Still, the study makes a compelling case that the future of AI in safety-critical domains may look less like an oracle and more like a well-trained librarian, one that knows exactly which expert to consult, what data to fetch, and when to admit it does not know. As advanced air mobility fills the skies, that kind of disciplined humility may prove to be the most valuable feature an AI assistant can offer.

Subject of Research: LLM-based decision support grounded in validated aviation models and real-time data for air traffic management

Article Title: A large language model–based method for improving decision-making in aviation

Article References: Sharif Rohani, A., Clarke, S. S., & Das, A. (2026). A large language model–based method for improving decision-making in aviation. Frontiers in Aerospace Engineering, 5, Article 1808857. https://doi.org/10.3389/fpace.2026.1808857

Image Credits: AI Generated

DOI: 10.3389/fpace.2026.1808857

Keywords: large language models, aviation decision support, air traffic management, NASA Ames, GPT-4o, hallucination, tool calling, advanced air mobility, prompt engineering, trajectory prediction, retrieval-augmented generation, BlueSky simulator

The research team, led by Saiful Islam Sagor and Erina Baynojir Joyee of UNC Charlotte’s Department of Mechanical Engineering and Engineering Science, together with colleagues in electrical and computer engineering and an industry partner at Corvid Technologies, set out to answer a question that has become urgent across engineering disciplines. General-purpose models such as GPT-4 and LLaMA have absorbed vast amounts of internet text, yet additive manufacturing knowledge lives in places those models rarely see in depth: peer-reviewed papers, manufacturer documentation, ASTM and ISO technical standards, process guidelines and material datasheets. Earlier studies had already flagged the problem, finding that GPT-4 offered only high-level, superficial responses on manufacturing questions, and that domain-specific systems built around retrieval delivered far more accurate answers.

To run a fair comparison, the researchers first built a curated knowledge corpus of remarkable breadth. It includes more than 250 peer-reviewed research articles, over 30 book chapters, more than 150 additive manufacturing process guidelines, over 50 international technical standards and standard operating procedures, and more than 120 material datasheets. Four researchers manually reviewed and filtered every document, removing duplicates and low-quality material. The collection spans the major families of 3D printing processes, from fused deposition modeling and stereolithography to direct ink writing, directed energy deposition, binder jetting, two-photon polymerization and even 4D printing with shape-memory materials, along with subdomains such as material science, machine settings, design rules and failure analysis.

All three model configurations in the study were built on the same foundation: LLaMA-3-8B, an open-weight, eight-billion-parameter transformer model chosen for its balance of capability, computational accessibility and reproducibility. The baseline configuration used the pretrained model as-is, answering questions with only the general knowledge learned during pretraining. The fine-tuned version was adapted using QLoRA, a parameter-efficient technique that compresses the model to 4-bit precision and inserts small trainable low-rank matrices into the attention layers, allowing domain adaptation on a single GPU-scale budget. Training ran for nine epochs with the AdamW optimizer and early stopping based on validation loss, and the loss curves converged cleanly, indicating the poor later performance was not caused by overfitting.

The retrieval-augmented generation system took a fundamentally different approach. Rather than baking knowledge into the model’s weights, it stores the curated corpus as dense semantic embeddings in a FAISS vector database, built with sentence-embedding models from the Sentence-Transformers framework. At inference time, each user question is embedded and matched against the index to pull out the most relevant document chunks, which are then appended to the query as context. A prompt template explicitly instructs the model to ground its answer solely in the retrieved material, a constraint designed to suppress hallucination. The architecture mirrors the logic of DeepMind’s RETRO model, which matched GPT-3 performance with 25 times fewer parameters by leveraging large-scale retrieval.

Evaluation rested on a benchmark of 200 expert-designed questions spanning fundamental concepts such as typical layer heights in fused deposition modeling, intermediate queries about process and material-property relationships, and advanced problems in troubleshooting, parameter optimization and defect formation. Experts assessed responses in blinded pairwise comparisons, judging accuracy, relevance, harmfulness, contextual correctness, understandability and real-world applicability. The verdict for retrieval was emphatic: retrieval-augmented responses were judged more accurate than the baseline in 75.5 percent of comparisons, more relevant in 90.8 percent, and preferred overall in 85.2 percent. Every one of the retrieval system’s outputs was rated non-harmful, 91.3 percent were clearly understandable, and 74 percent offered guidance usable in real engineering contexts.

The fine-tuned model’s numbers told the opposite story. Trained on raw, unstructured domain text with a causal language modeling objective, the model produced more accurate answers than the baseline in only 5.6 percent of comparisons, and more relevant answers in just 32.5 percent. Only 8.5 percent of its responses were considered suitable for real-world use, and just 31.8 percent were judged understandable. The authors attribute this not to overfitting, which the loss curves rule out, but to a mismatch between the training objective and the task: predicting the next token from raw text teaches a model domain vocabulary, but it never teaches the model how to answer a structured technical question accurately or concisely. Exposure to knowledge, in other words, is not the same as learning to use it.

The implications extend well beyond 3D printing. Fine-tuned models are static; incorporating new standards, new materials or new process data means retraining, and their answers offer no transparent link to sources. Retrieval systems, by contrast, can absorb new documents simply by updating the index, and every answer can be traced back to retrieved reference material, a property that matters enormously in fields where correctness and traceability are non-negotiable. The findings align with a growing body of evidence from medicine, biomedical natural language processing and manufacturing showing that hybrid architectures pairing language models with external knowledge retrieval outperform parameter-based adaptation for knowledge-intensive tasks.

The authors are careful to note the limits of their conclusions. The benchmark contains 200 questions, which cannot capture the full diversity of real-world queries; expert judgment, though structured and blinded, introduces subjectivity; and the fine-tuning experiment used raw text rather than supervised question-answer pairs, which better instruction tuning might have improved. The study also tested only one model family at one size, leaving open whether larger or proprietary models would shift the balance. The researchers propose several next steps, including supervised fine-tuning on high-quality question-answer pairs, hybrid approaches combining retrieval with instruction tuning, comparisons against state-of-the-art proprietary systems, and integration of real-time process monitoring signals such as thermal and optical inspection data to support live manufacturing decisions.

Even with those caveats, the central message lands with unusual force for an AI benchmark study. For engineers hoping to deploy trustworthy AI assistants in specialized technical domains, the cheapest and most effective route may be to stop trying to stuff knowledge into model weights and instead give the model a well-organized library to consult. In additive manufacturing, where a single wrong parameter recommendation can warp a part, ruin a build or compromise a component, the difference between 85 percent and 5.6 percent is not a statistic. It is the difference between a tool an engineer can trust and one they cannot afford to use.

Subject of Research: Domain adaptation of large language models for additive manufacturing question answering

Article Title: Domain adaptation of large language models for additive manufacturing using retrieval-augmented generation and fine-tuning

Article References: Sagor, S. I., Haghighi, T., Rahatuzzaman, M., Khecho, A., Alam, M. N., Zorn, J., & Joyee, E. B. (2026). Domain adaptation of large language models for additive manufacturing using retrieval-augmented generation and fine-tuning. Frontiers in Manufacturing Technology, 6, Article 1875357. https://doi.org/10.3389/fmtec.2026.1875357

Image Credits: AI Generated

DOI: 10.3389/fmtec.2026.1875357

Keywords: large language models, retrieval-augmented generation, fine-tuning, additive manufacturing, 3D printing, LLaMA-3-8B, QLoRA, vector database, expert evaluation, artificial intelligence, manufacturing technology, hallucination