Search

To assess the capabilities of large language models (LLMs), including Open AI (GPT-4.0) and Microsoft Bing (GPT-4), in generating structured reports, the Breast Imaging Reporting and Data System (BI-RADS) categories, and management recommendations from free-text breast ultrasound reports. In this retrospective study, 100 free-text breast ultrasound reports from patients who underwent surgery between January and May 2023 were gathered. The capabilities of Open AI (GPT-4.0) and Microsoft Bing (GPT-4) to convert these unstructured reports into structured ultrasound reports were studied. The quality of structured reports, BI-RADS categories, and management recommendations generated by GPT-4.0 and Bing were evaluated by senior radiologists based on the guidelines. Open AI (GPT-4.0) was better than Microsoft Bing (GPT-4) in terms of performance in generating structured reports (88% vs. 55%; p < 0.001), giving correct BI-RADS categories (54% vs. 47%; p = 0.013) and providing reasonable management recommendations (81% vs. 63%; p < 0.001). As the ability to predict benign and malignant characteristics, GPT-4.0 performed significantly better than Bing (AUC, 0.9317 vs. 0.8177; p < 0.001), while both performed significantly inferior to senior radiologists (AUC, 0.9763; both p < 0.001). This study highlights the potential of LLMs, specifically Open AI (GPT-4.0), in converting unstructured breast ultrasound reports into structured ones, offering accurate diagnoses and providing reasonable recommendations. [ABSTRACT FROM AUTHOR]

Pediatric stroke is an important cause of morbidity in children. Although research can be challenging, large amounts of data have been captured through collaborative efforts in the International Pediatric Stroke Study (IPSS). This study explores the use of an advanced artificial intelligence program, the Generative Pre-trained Transformer (GPT), to enter pediatric stroke data into the IPSS. The most recent 50 clinical notes of patients with ischemic stroke or cerebral venous sinus thrombosis at the UTHealth Pediatric Stroke Clinic were deidentified. Domain-specific prompts were engineered for an offline artificial intelligence program (GPT) to answer IPSS questions. Responses from GPT were compared with the human rater. Percent agreement was assessed across 50 patients for each of the 114 queries developed from the IPSS database outcome questionnaire. GPT demonstrated strong performance on several questions but showed variability overall. In its early iterations it was able to match human judgment occasionally with an accuracy score of 1.00 (n = 20, 17.5%), but it scored as low as 0.26 in some patients. Prompts were adjusted in four subsequent iterations to increase accuracy. In its fourth iteration, agreement was 93.6%, with a maximum agreement of 100% and minimum of 62%. Of 2400 individual items assessed, our model entered 2247 (93.6%) correctly and 153 (6.4%) incorrectly. Although our tailored generative model with domain-specific prompt engineering and ontological guidance shows promise for research applications, further refinement is needed to enhance its accuracy. It cannot enter data entirely independently, but it can be employed in tandem with human oversight contributing to a collaborative approach that reduces overall effort. [ABSTRACT FROM AUTHOR]

Large language models (LLMs) such as OpenAI's ChatGPT, Google's Bard, and Microsoft's Bing Chat have shown potential as educational tools in the medical and dental fields. This study evaluated their effectiveness using questions from the Japanese national dental hygienist examination, focusing on textual information only. We analyzed 73 questions from the 32nd Japanese national dental hygienist examination, conducted in March 2023, using LLMs ChatGPT-3.5, GPT-4, Bard, and Bing Chat. Each question was categorized into one of nine domains. Standardized prompts were used for all LLMs, and Fisher's exact test was applied for statistical analysis. GPT-4 achieved the highest accuracy (75.3%), followed by Bing (68.5%), Bard (66.7%), and GPT-3.5 (63.0%). There were no statistically significant differences between the LLMs. The performance varied across different question categories, with all models excelling in the 'Disease mechanism and promotion of recovery process' category (100% accuracy). GPT-4 generally outperformed other models, especially in multi-answer questions. GPT-4 demonstrated the highest overall accuracy among the LLMs tested, indicating its superior potential as an educational support tool in dental hygiene studies. The study highlights the varied performance of different LLMs across various question categories. While GPT-4 is currently the most effective, the capabilities of LLMs in educational settings are subject to continual change and improvement. [ABSTRACT FROM AUTHOR]

This study aimed to determine the accuracy of large language models (LLMs) in answering oral and maxillofacial surgery (OMS) multiple choice questions. A total of 259 questions from the university's question bank were answered by the LLMs (GPT-3.5, GPT-4, Llama 2, Gemini, and Copilot). The scores per category as well as the total score out of 259 were recorded and evaluated, with the passing score set at 50%. The mean overall score amongst all LLMs was 62.5%. GPT-4 performed the best (76.8%, 95% confidence interval (CI) 71.4–82.2%), followed by Copilot (72.6%, 95% CI 67.2–78.0%), GPT-3.5 (62.2%, 95% CI 56.4–68.0%), Gemini (58.7%, 95% CI 52.9–64.5%), and Llama 2 (42.5%, 95% CI 37.1–48.6%). There was a statistically significant difference between the scores of the five LLMs overall (χ2 = 79.9, df = 4, P < 0.001) and within all categories except 'basic sciences' (P = 0.129), 'dentoalveolar and implant surgery' (P = 0.052), and 'oral medicine/pathology/radiology' (P = 0.801). The LLMs performed best in 'basic sciences' (68.9%) and poorest in 'pharmacology' (45.9%). The LLMs can be used as adjuncts in teaching, but should not be used for clinical decision-making until the models are further developed and validated. [ABSTRACT FROM AUTHOR]

The integration of teledermatology and artificial intelligence (AI) marks a significant advancement in dermatologic care. This study examines the synergistic interplay between these two domains, highlighting their collective impact on enhancing the accuracy, accessibility, and efficiency of teledermatologic services. Teledermatology expands dermatologic care to remote and underserved areas, and AI technologies show considerable potential in analyzing dermatologic images and performing various tasks involved in teledermatology consultations. Such integration facilitates rapid, precise diagnoses, personalized treatment plans, and data-driven insights. Our explorative study involved designing a GPT-based chatbot named "Dr. DermBot" and exploring its performance in a teledermatologic consultation process. The design phase focused on the chatbot's ability to conduct consultations autonomously. The subsequent testing phase assessed its performance against the backdrop of current teledermatologic practices, exploring the potential of AI and chatbots to simulate and potentially enhance teledermatologic health care. Our study demonstrates the promising future of combining teledermatology with AI. It also brings to light ethical and legal concerns, including the protection of patient data privacy and adherence to regulatory standards. The union of teledermatology and AI not only aims to enhance the precision of teledermatologic diagnoses but also broadens the accessibility of dermatologic services to previously underserved populations, benefiting patients, health care providers, and the overall health care system. [ABSTRACT FROM AUTHOR]

Heart failure requires complex management, and increased patient knowledge has been shown to improve outcomes. This study assessed the knowledge of Chat Generative Pre-trained Transformer (ChatGPT) and its appropriateness as a supplemental resource of information for patients with heart failure. A total of 107 frequently asked heart failure-related questions were included in 3 categories: "basic knowledge" (49), "management" (41) and "other" (17). Two responses per question were generated using both GPT-3.5 and GPT-4 (i.e., two responses per question per model). The accuracy and reproducibility of responses were graded by two reviewers, board-certified in cardiology, with differences resolved by a third reviewer, board-certified in cardiology and advanced heart failure. Accuracy was graded using a four-point scale: (1) comprehensive, (2) correct but inadequate, (3) some correct and some incorrect, and (4) completely incorrect. GPT-4 provided 107/107 (100%) responses with correct information. Further, GPT-4 displayed a greater proportion of comprehensive knowledge for the categories of "basic knowledge" and "management" (89.8% and 82.9%, respectively). For GPT-3, there were two total responses (1.9%) graded as "some correct and incorrect" for GPT-3.5, while no "completely incorrect" responses were produced. With respect to comprehensive knowledge, GPT-3.5 performed best in the "management" category and "other" category (prognosis, procedures, and support) (78.1%, 94.1%). The models also provided highly reproducible responses, with GPT-3.5 scoring above 94% in every category and GPT-4 with 100% for all answers. GPT-3.5 and GPT-4 answered the majority of heart failure-related questions accurately and reliably. If validated in future studies, ChatGPT may serve as a useful tool in the future by providing accessible health-related information and education to patients living with heart failure. In its current state, ChatGPT necessitates further rigorous testing and validation to ensure patient safety and equity across all patient demographics. [ABSTRACT FROM AUTHOR]

The rapidly increasing adoption of large language models in medicine has drawn attention to potential applications within the field of neurosurgery. This study evaluates the effects of various contextualization methods on ChatGPT's ability to provide expert-consensus aligned recommendations on the diagnosis and management of Chiari Malformation and Syringomyelia. Native GPT4 and GPT4 models contextualized using various strategies were asked questions revised from the 2022 Chiari and Syringomyelia Consortium International Consensus Document. ChatGPT-provided responses were then compared to consensus statements using reviewer assessments of 1) responding to the prompt, 2) agreement of ChatGPT response with consensus statements, 3) recommendation to consult with a medical professional, and 4) presence of supplementary information. Flesch-Kincaid, SMOG, word count, and Gunning-Fog readability scores were calculated for each model using the quanteda package in R. Relative to GPT4, all contextualized GPTs demonstrated increased agreement with consensus statements. PDF+Prompting and Prompting models provided the most elevated agreement scores of 19 of 24 and 23 of 24, respectively, versus 9 of 24 for GPT4 (p=.021, p=.001). A trend toward improved readability was observed when comparing contextualized models at large to ChatGPT4, with significant decreases in average word count (180.7 vs 382.3, p<.001) and Flesch-Kincaid Reading Ease score (11.7 vs 17.2, p=.033). The enhanced performance observed in response to ChatGPT4 contextualization suggests broader applications of large language models in neurosurgery than what the current literature indicates. This study provides proof of concept for the use of contextualized GPT models in neurosurgical contexts and showcases the easy accessibility of improved model performance. [ABSTRACT FROM AUTHOR]

Artificial intelligence (AI) has become increasingly used in neurosurgery. Generative pretrained transformers (GPTs) have been of particular interest. However, ethical concerns regarding the incorporation of AI into the field remain underexplored. We delineate key ethical considerations using a novel GPT-based, human-modified approach, synthesize the most common considerations, and present an ethical framework for the involvement of AI in neurosurgery. GPT-4, ChatGPT, Bing Chat/Copilot, You, Perplexity.ai, and Google Bard were queried with the prompt "How can artificial intelligence be ethically incorporated into neurosurgery?". Then, a layered GPT-based thematic analysis was performed. The authors synthesized the results into considerations for the ethical incorporation of AI into neurosurgery. Separate Pareto analyses with 20% threshold and 10% threshold were conducted to determine salient themes. The authors refined these salient themes. Twelve key ethical considerations focusing on stakeholders, clinical implementation, and governance were identified. Refinement of the Pareto analysis of the top 20% most salient themes in the aggregated GPT outputs yielded 10 key considerations. Additionally, from the top 10% most salient themes, 5 considerations were retrieved. An ethical framework for the use of AI in neurosurgery was developed. It is critical to address the ethical considerations associated with the use of AI in neurosurgery. The framework described in this manuscript may facilitate the integration of AI into neurosurgery, benefitting both patients and neurosurgeons alike. We urge neurosurgeons to use AI only for validated purposes and caution against automatic adoption of its outputs without neurosurgeon interpretation. [ABSTRACT FROM AUTHOR]

The use of artificial intelligence (AI) tools to aid in summarizing information in medicine and research has recently garnered a huge amount of interest. While tools such as ChatGPT produce convincing and naturally sounding output, the answers are sometimes incorrect. Some of these drawbacks, it is hoped, can be avoided by using programmes trained for a more specific scope. In this study we compared the performance of a new AI tool (the-literature.com) to the latest version OpenAI's ChatGPT (GPT-4) in summarizing topics that the authors have significantly contributed to. The AI tools were asked to produce a literature review on 7 topics. These were selected based on the research topics that the authors were intimately familiar with and have contributed to through their own publications. The output produced by the AI tools were graded on a 1–5 Likert scale for accuracy, comprehensiveness, and relevance by two fellowship trained consultant radiologists. The-literature.com produced 3 excellent summaries, 3 very poor summaries not relevant to the prompt, and one summary, which was relevant but did not include all relevant papers. All of the summaries produced by GPT-4 were relevant, but fewer relevant papers were identified. The average Likert rating was for the-literature was 2.88 and 3.86 for GPT-4. There was good agreement between the ratings of both radiologists (ICC = 0.883). Summaries produced by AI in its current state require careful human validation. GPT-4 on average provides higher quality summaries. Neither tool can reliably identify all relevant publications. • Literature review is a crucial, but labour intensive, part of research. • Artificial intelligence tools have the potential to realise significant time savings. • AI driven tools can in certain cases produce impressive literature reviews. • However, current AI literature review tools are too unreliable for routine use. [ABSTRACT FROM AUTHOR]

OpenAI's GPT-4 and similar large language models (LLMs) show impressive conversational capabilities. This Opinion asks: in what sense does a LLM have knowledge? The answer to this question extends beyond the capabilities of a particular AI chatbot and challenges our assumptions about the nature of knowledge and intelligence. We answer by granting LLMs 'instrumental knowledge': knowledge defined by a certain set of abilities. How is such knowledge related to the more ordinary, 'worldly' knowledge exhibited by humans? To address this, we turn to a core concept in cognitive science, world models, and explore the degree to which instrumental knowledge might incorporate such structured representations. We discuss how LLMs could recover degrees of worldly knowledge and suggest that such recovery will be governed by an implicit, resource-rational tradeoff between world models and task demands. In what sense does a large language model (LLM) have knowledge? We answer by granting LLMs 'instrumental knowledge': knowledge gained by using next-word generation as an instrument. We then ask how instrumental knowledge is related to the ordinary, 'worldly knowledge' exhibited by humans, and explore this question in terms of the degree to which instrumental knowledge can be said to incorporate the structured world models of cognitive science. We discuss ways LLMs could recover degrees of worldly knowledge and suggest that such recovery will be governed by an implicit, resource-rational tradeoff between world models and tasks. Our answer to this question extends beyond the capabilities of a particular AI system and challenges assumptions about the nature of knowledge and intelligence. [ABSTRACT FROM AUTHOR]

Business Process Model and Notation (BPMN) is a standard for formally modeling complex business processes. Manual creation of BPMN models can be time-consuming and error-prone, prompting a need for automation. Existing approaches, such as rule-based methods, machine learning, and machine translation, have progressed but face accuracy and real-world applicability challenges. In this research paper, we propose a novel method for automated extraction of BPMN models from textual descriptions using natural language processing (NLP) tools and deep learning models, including the spaCy library for text processing, a fine-tuned BERT model, and state-of-the-art large language models like GPT-3.5-Turbo and GPT-4. We utilize Graphviz, an open-source graph visualization software, to visualize the extracted processes. Our method supports representing tasks, exclusive gateways, parallel gateways, and start and end events in the generated BPMN models. The evaluation of 31 textual descriptions shows that our method generates process models with 96% accuracy using GPT-4 and 80% accuracy using GPT-3.5-Turbo large language models. Although subject to certain limitations, such as occasional inaccuracies in model outputs and reliance on well-formed input text, our approach offers a valuable contribution to the growing body of research on automating BPMN model generation. [ABSTRACT FROM AUTHOR]

The Common Vulnerabilities and Exposures (CVE) system is a widely used standard for identifying and tracking known vulnerabilities in software systems. The severity of these vulnerabilities must be determined in order to prioritize mitigation efforts. However, assigning severity to a vulnerability is a challenging task that requires careful analysis of its characteristics and potential impact. Considering the vast number of vulnerabilities identified every year, it is vital to automate the severity assignment, thereby reducing manual effort. This paper proposes a novel approach for predicting the severity of vulnerabilities based on their CVE description using GPT-2, a state-of-the-art language model. The CVSS severity values distribution imbalance is addressed using oversampling and contextual data augmentation techniques. This approach leverages the large-scale language modeling capabilities of GPT-2 to automatically extract relevant features from CVE descriptions and predict the severity level of the vulnerability. The model is evaluated on a test data set of 7,765 CVEs and achieves a high accuracy of 84.2% and an F 1 score of 0.82 in predicting the severity of the vulnerabilities on the test data. A comparative analysis of this approach was done against state-of-the-art methods, demonstrating the superior performance of the proposed approach. Based on the results, the proposed approach could be considered a valuable tool for quickly and accurately identifying high-severity vulnerabilities, facilitating more efficient and effective vulnerability management practices. Furthermore, this approach could be extended to other natural language processing tasks related to vulnerability analysis and management. [ABSTRACT FROM AUTHOR]

Artificial intelligence (AI) is advancing rapidly and continually evolving in various fields. Recently, the release of ChatGPT has sparked significant public interest. In this study, we revisit the '100 Important Questions Facing Plant Science' by leveraging ChatGPT as a valuable tool for generating thought-provoking questions relevant to plant science. These questions primarily revolve around the utilization of plants in product development, understanding plant mechanisms, plant–environment interactions, and enhancing plant traits, with an emphasis on sustainable product development. While ChatGPT may not capture certain crucial aspects highlighted by scientists, it offers valuable insights into the questions generated by experts. Our analysis demonstrates that ChatGPT can be cautiously employed as a supportive tool to facilitate, streamline, and expedite specific tasks in plant science. Large language models (LLMs) were recently made widely available to the public. ChatGPT is an LLM released in late 2022, which has received massive public attention and interest. A need has emerged to evaluate how ChatGPT and other such artificial intelligence (AI) tools can benefit plant science. Proper use of such AI tools can improve and expedite the scientific process. [ABSTRACT FROM AUTHOR]

As the influence of Incremental Sheet Forming (ISF) grows in manufacturing sectors, so does the demand for precise and updated knowledge construction in this domain. In this research, we evaluate the capability of Large Language Models (LLMs) to capture domain-specific knowledge, using ISF as a case study. Recognizing common LLMs' limitations such as potential inaccuracies and outdated information reliance, we propose a comprehensive approach involving automated and adaptive knowledge extraction, enrichment, and integration into an ISF-specific dataset. We then fine-tune the LLMs for ISF-related text classification and prompt response tasks. Our results reveal a significant enhancement in LLMs' performance within the ISF domain, with a domain knowledge acquisition rate exceeding that of GPT-3.5 by 10.4%, achieved by the fine-tuned Alpaca-33B model. Additionally, we introduce a novel conversational prototype designed to refine the accuracy and relevance of LLMs in the ISF domain. Our findings will guide future efforts in downstream tasks such as ISF-domain knowledge graph construction and quality prediction. [ABSTRACT FROM AUTHOR]

As new AI technologies such as Large Language Models (LLM) quickly evolve, the need for enhancing general-purpose LLMs with physical knowledge to better serve the manufacturing community has been increasingly recognized. This paper presents a method that tailors GPT-3.5 with domain-specific knowledge for intelligent aircraft maintenance. Specifically, aircraft ontology is investigated to curate maintenance logs with encoded component hierarchical structure to fine-tune GPT-3.5. Experimental results demonstrate the effectiveness of the developed method in accurately identifying defective components and providing consistent maintenance action recommendations, outperforming general-purpose GPT-3.5 and GPT-4.0. The method can be adapted to other domains in manufacturing and beyond. [ABSTRACT FROM AUTHOR]

Large language models (LLMs) like Generative Pre-trained Transformer 4 (GPT-4) provide an opportunity to utilize an artificial intelligence (AI) framework to extract and interpret data contained within unstructured clinical text to provide an interactive and context-aware understanding of the patient's history. HopeLLM is an AI framework specializing in processing oncology clinical notes and providing tools tailored for cancer care, including a conversational interface to query patient medical records. This tool has the potential to accelerate clinical and research workflows by helping researchers rapidly review medical records to abstract key clinical endpoints. We present results on the use of HopeLLM for data abstraction of prostate-specific antigen (PSA) values present in unstructured clinical notes, a first step for determining the biochemical recurrence of prostate cancer. HopeLLM is built around a custom embedding model trained on a corpus of billions of tokens of text drawn from City of Hope clinical notes. We tested the performance of HopeLLM in abstracting PSA values and their corresponding dates from unstructured clinical notes of patients who later developed biochemical recurrence and compared these results to data manually abstracted by physicians. We report on the concordance of PSA values and dates returned from HLM with manual abstraction performed by a clinician and explore the context of non-concordant data. An initial dataset of PSA values was manually abstracted from six patients. 32 unique records of PSA measurements were aligned across manual and HopeLLM abstraction of electronic medical records. The mean absolute percent error (MAPE) of the abstracted PSA values was 2.9%, and the mean difference of PSA measurement dates was 0.23 months, or less than one week. In addition, HopeLLM surfaced an identified 7 PSA measurements that appeared in the clinical notes but did not appear in the manual abstraction. Overall, manual abstraction is a time-consuming and error-prone process, and we find HopeLLM to quickly and accurately abstract PSA values and dates from clinical text. LLM can streamline data collection and warrants validation on larger datasets and at other institutions. [ABSTRACT FROM AUTHOR]

• Large language models (GPT-3.5, GPT-4, and Llama3) could predict human-generated ratings of coherence, content, and tangentiality in clinical speech samples with moderate accuracy. • GPT-4 and Llama3 predictions were more accurate to the human ratings than GPT-3.5. • Using default parameters, both GPT models showed high levels of inconsistency in their predicted ratings, while Llama3 was relatively consistent. • Inconsistency in large language model predictions could be mitigated by controlling the temperature parameter. • Ensemble methods which combined multiple sets of model-predicted ratings improved predictive accuracy. Natural Language Processing (NLP) methods have shown promise for the assessment of formal thought disorder, a hallmark feature of schizophrenia in which disturbances to the structure, organization, or coherence of thought can manifest as disordered or incoherent speech. We investigated the suitability of modern Large Language Models (LLMs - e.g., GPT-3.5, GPT-4, and Llama 3) to predict expert-generated ratings for three dimensions of thought disorder (coherence, content, and tangentiality) assigned to speech samples collected from both patients with a diagnosis of schizophrenia (n = 26) and healthy control participants (n = 25). In addition to (1) evaluating the accuracy of LLM-generated ratings relative to human experts, we also (2) investigated the degree to which the LLMs produced consistent ratings across multiple trials, and we (3) sought to understand the factors that impacted the consistency of LLM-generated output. We found that machine-generated ratings of the level of thought disorder in speech matched favorably those of expert humans, and we identified a tradeoff between accuracy and consistency in LLM ratings. Unlike traditional NLP methods, LLMs were not always consistent in their predictions, but these inconsistencies could be mitigated with careful parameter selection and ensemble methods. We discuss implications for NLP-based assessment of thought disorder and provide recommendations of best practices for integrating these methods in the field of psychiatry. [ABSTRACT FROM AUTHOR]

To underpin scientific evaluations of chemical risks, agencies such as the European Food Safety Authority (EFSA) heavily rely on the outcome of systematic reviews, which currently require extensive manual effort. One specific challenge constitutes the meaningful use of vast amounts of valuable data from new approach methodologies (NAMs) which are mostly reported in an unstructured way in the scientific literature. In the EFSA-initiated project 'AI4NAMS', the potential of large language models (LLMs) was explored. Models from the GPT family, where GPT refers to Generative Pre-trained Transformer, were used for searching, extracting, and integrating data from scientific publications for NAM-based risk assessment. A case study on bisphenol A (BPA), a substance of very high concern due to its adverse effects on human health, focused on the structured extraction of information on test systems measuring biologic activities of BPA. Fine-tuning of a GPT-3 model (Curie base model) for extraction tasks was tested and the performance of the fine-tuned model was compared to the performance of a ready-to-use model (text-davinci-002). To update findings from the AI4NAMS project and to check for technical progress, the fine-tuning exercise was repeated and a newer ready-to-use model (text-davinci-003) served as comparison. In both cases, the fine-tuned Curie model was found to be superior to the ready-to-use model. Performance improvement was also obvious between text-davinci-002 and the newer text-davinci-003. Our findings demonstrate how fine-tuning and the swift general technical development improve model performance and contribute to the growing number of investigations on the use of AI in scientific and regulatory tasks. [Display omitted] • Fine-tuning improves large language model performance for text-extraction tasks. • Swift technical development enhances ready-to-use model performance. • Task complexity is the main limiting factor for model performance. • Proper prompt engineering is essential for successful use of large language models. [ABSTRACT FROM AUTHOR]

Effective evidence synthesis is important for the integration of scientific research into decision-making. However, fully depicting the vast mosaic of concepts and applications in environmental sciences and ecology often entails a substantial workload. New Artificial Intelligence (AI) tools present an attractive option for addressing this challenge but require sufficient validation to match the vigorous standards of a systematic review. This article demonstrates the use of generative AI in the selection of relevant literature as part of a systematic review on indicators of ecosystem condition. We highlight, through the development of an optimal prompt to communicate inclusion and exclusion criteria, the need to describe ecosystem condition as a multidimensional concept whilst also maintaining clarity on what does not meet the criteria of comprehensiveness. We show that, although not completely infallible, the GPT-3.5 model significantly outperforms traditional literature screening processes in terms of speed and efficiency whilst correctly selecting 83 % of relevant literature for review. Our study highlights the importance of precision in prompt design and the setting of query parameters for the AI model and opens the perspective for future work using language models to contextualize complex concepts in the environmental sciences. Future development of this methodology in tandem with the continued evolution of the accessibility and capacity of AI tools presents a great potential to improve evidence synthesis through gains in efficiency and possible scope. • Developments in generative AI present a useful tool for evidence synthesis. • GPT-3.5 is an effective method for screening literature on Ecosystem Condition. • Clarity of prompt and repetition of key terms is crucial for accurate responses. • The iterative prompt development allows to sharpen key research terms. • AI approach mitigates against persistent subjectivity in the review process. [ABSTRACT FROM AUTHOR]

Addressing the challenge of limitations imposed by uni-modal tasks in healthcare AI, we introduce RadAIonc, a specialized GPT-4 based Large Language Model (LLM). RadAIonc is engineered to enhance the complex workflows in radiation oncology, offering a streamlined and multifaceted solution. RadAIonc leverages GPT-4's advanced computational capabilities through a novel prompt-driven methodology. This model exceeds the conventional boundaries of generating clinical reports by also proposing comprehensive radiation treatment plans. Our unique method eliminates the necessity for traditional model fine-tuning by utilizing complex, expertly crafted prompts that simulate the intricate decision-making processes of oncology specialists. This results in RadAIonc delivering precise, contextually relevant responses. We further developed a custom prompt-driven framework to improve the accuracy and efficiency of identifying optimal treatment options, covering aspects like equipment selection, dosing, and techniques. The model's outputs are thoroughly evaluated by clinical professionals against established benchmarks for accuracy and relevance. In trials focused on prostate and head and neck cancers, RadAIonc demonstrated superior performance in critical tasks within oncology, surpassing standard GPT-4 outputs. Clinical reports and treatment recommendations from RadAIonc were rated, on average, 0.5 points higher by expert reviewers than those from the base GPT-4 model. Moreover, the integration of an advanced segmentation framework markedly enhanced the model's efficiency in processing complex imaging data, facilitating more streamlined treatment planning processes. These improvements highlight significant strides in operational efficiency and accuracy, positioning RadAIonc as a leading innovator in radiation oncology patient care and treatment strategy. RadAIonc represents a powerful and flexible tool that mirrors the intricate clinical decision-making process in Radiation Oncology. Its pioneering application of GPT-4's prompt-driven features not only refine clinical workflows but also heralds substantial progress in enhancing the quality and precision of patient care. The model's proven efficacy across diverse tasks and contexts emphasizes its potential as a transformative resource in the field. [ABSTRACT FROM AUTHOR]

): Generative AI's rise prompts our exploration into AI-assisted scientific writing. By the 1990s, NLP models used statistical approaches to detect textual patterns. OpenAI's GPT-2 debuted in February 2019, and by 2023, ChatGPT 3.5 gained prominence. We aim to offer key insights for navigating the growing role of AI in scientific writing to foster productive discourse and a framework. Our hypothesis: increased generative AI accessibility correlates to a surge in AI-assisted content in academic writing. s: The zeroGPT was selected as the generative AI detection software based on its accuracy from published data. A validation study, including 21 journal articles from 1980 and 2000, was conducted to evaluate zeroGPT. A subsequent time-series analysis (2019-2024) was performed on 262 cancer-related, full-length, peer-reviewed research articles from 2 journals (J.1/ J.2) and 207 scientific conference abstracts using zeroGPT. Both journals' impact factors range between 5 and 8.5 (2019-2024). J.1 is paywalled, and J.2 is open access. Results were categorized using binary and percent classification methods. s: In the validation study, the average percentage of AI-assisted writing was 1.2% in 1980 and 12.3% in 2000. This serves as a baseline for the background benchmark. For J.1, average AI content varied from 18.2% (2022) to 22.2% (2021), with all-time maximum of 68.6% (2020). For J.2, average content varied from 28.8% (2019) to 52.2% (2024), with all-time maximum of 94.8% (2023). For the conference abstracts, average content varied from 3.6% (2019) to 12.2% (2021), with all-time maximum of 58.1% (2023). n: Journal-specific analysis reveals intriguing trends, notably J.1's relatively consistent average AI usage (18.2% to 22.2%) and J.2's fluctuation (28.8% to 52.2%). Maximum content as high as 94.8% (J.2, 2023) prompts questions about scientific merit. Inter-journal content differences may relate to the journal access model. Also, at some point, J.1 began mandating a statement on generative AI use; might this relate to the 2022 dip in J.1's AI content? Ultimately, while generative AI tools can lead to improved communication and a streamlined writing process, risks include erroneous conclusions about data and falsely authoritative tones. Looking ahead, proactive measures such as educational initiatives for effective AI integration should be considered. Our work does not aim to assign responsibility to specific journals or conferences but rather stands as a litmus test to predict the realistic relevance of AI generated content in published scientific articles from 2019-2024. [ABSTRACT FROM AUTHOR]

Recent advances in artificial intelligence such as large language models (LLMs) offer a promising avenue for enhancing clinical documentation and monitoring patient-reported outcomes (PRO). This study aims to compare four leading open-source and proprietary LLMs, Mixtral-8x7B, Llama-2B, Qwen-1.5, and GPT-4, in generating summaries of patient-reported symptoms using the adapted Physician Documentation Quality Index (PDQI). A previously reported web-based application utilizing 35 items from the PRO-CTCAE scale was used to create an interactive form for breast cancer patients to report treatment-related symptoms. The four LLMs were used to generate natural language summaries for four hypothetical patients with non-identifiable patient data. Twelve resident physician raters evaluated the summaries using an abbreviated PDQI questionnaire, rating accuracy, usefulness, comprehensibility, succinctness scored on a 5-point Likert scale. IRB approval was not required in accordance with the NIH 2018 Revised Common Rule Requirements as the study used researcher-generated, non-identifiable data. Forty-seven physician ratings were collected. A repeated measures ANOVA showed significant differences in accuracy among the models (F(2.120, 97.52) = 15.30, p < 0.0001), with Mixtral-8x7b (M = 3.60), GPT-4 (M = 3.78), and Qwen-1.5 (M = 3.62) significantly surpassing Llama-2 (M = 2.77, p ≤ 0.001), with no significant difference between the three. Mixtral-8x7b (M = 3.47) outperformed Llama-2 in usefulness (M = 2.87, p<.05) and outperformed Llama-2 and Qwen in succinctness (p<.05). No significant differences were found in comprehensiveness. Reviewers noted 6 mistakes in Llama-2 summaries, and 1 mistake each in Mixtral-8x7b and Qwen-1.5 summaries. This study demonstrates that the latest open-source LLMs, such as Mixtral-8x7b and Qwen-1.5, can match the performance of their closed-source counterpart, GPT-4, in physician-rated measures of documentation quality and outperform their predecessor, Llama-2. This study highlights the narrowing gap between open-source and proprietary LLMs in medical documentation. These findings may help inform the strategic selection of cost-effective and data-safe LLMs for future clinical research and practice integration, potentially democratizing advanced AI tools for a broader healthcare audience. However, further validation in real-world clinical settings is necessary to assess the impact of these models on patient care efficiency and efficacy. [ABSTRACT FROM AUTHOR]

Language models (LM), like GPT-3.5 and GPT-4, have achieved expert-level performance on medical question-answer datasets. However, few studies have compared LMs to physicians in accomplishing real-world clinical tasks. Here we aimed to determine if LMs can extract features relevant to tumor staging from pathology reports (Task 1) and if LMs can "reason out" tumor staging from extracted features (Task 2). We randomly selected 40 breast cancer cases with pathological T and N stages (AJCC 6th edition) from the TCGA-Reports dataset. For task 1, we instructed the LMs (GPT-3.5 and GPT-4) and a physician (Radiation Oncologist) to extract features relevant to tumor staging from each pathology report. We then examined the percent agreement by an exact match of the extracted features between the LMs and the physician. For task 2, we either: (a) instructed the LMs and the physician to determine T and N stages from the extracted features given the AJCC staging criteria ("Reasoning Approach") or (b) inputted the extracted features into a deterministic Python function for staging ("Function Approach"). Finally, we compared the Mathews Correlation Coefficient (MCC) of each approach between the LMs and the physician. Pathological T and N stages could be determined for 33/40 cases by the physician. Analyzing this subset (Task 1), we found that the extracted features were an exact match between the LMs and the physician in a range of 76-100% with GPT-3.5 and 94-100% with GPT-4. Examining the few incorrectly extracted features by the LMs, we discovered that they made repeated, explainable mistakes. For example, LMs occasionally extracted the tumor size of the entire mass instead of the largest focus. In comparing MCCs for final tumor staging (Task 2), we discovered that (a) GPT-4 and the physician performed nearly as well irrespective of approach, while (b) GPT-3.5 had poorer performance overall but slightly improved results with the "Function Approach" (Table 1). In all, the cost and time per case for GPT-3.5 was $0.005 and 4 seconds, and for GPT-4 was $0.1 and 16 seconds. LMs can stage tumors similarly to a physician with minimal cost and time spent. In addition, we developed potential ways of improving LM performance, such as providing additional instructions for common fringe cases or utilizing deterministic Python functions to circumvent LM reasoning. Overall, this study lays the groundwork for a more extensive evaluation of other LMs on real-world clinical tasks involving larger datasets, generating clinical documents, and compiling oncologic histories. [ABSTRACT FROM AUTHOR]

Clinical trial screening is a critical yet resource-intensive step in the development of medical treatments. Additionally, Large Language Models (LLMs) have become increasingly powerful and popular tools in recent years. This study investigates the potential of Large Language Models (LLMs) as an alternative to traditional screening tools and techniques, measuring their efficiency and accuracy. We used two datasets: the n2c2 2018 cohort selection dataset, available publicly, and a dataset from patients screened for an institutional phase II clinical trial. Each trial had 13 and 14 criteria respectively. Criteria ground truth was readily available for the former while the latter required manual verification by clinical research staff. The LLM pipeline was created for the n2c2 dataset, requiring 20 patient records for initial LLM prompt fine-tuning. Twenty patients were used to further refine prompts in our institutional application. Patient electronic health records (EHRs) were processed using OpenAI's text-embedding-3-small model, creating embeddings to populate a vector database for efficient document section retrieval. For each screening criterion, we crafted physician-guided, LLM-enhanced prompts to extract the five most pertinent EHR sections. These were then analyzed by the GPT-3.5 Turbo/GPT-4 API on our HIPAA-compliant servers. The analysis focused on the accuracy, sensitivity, and specificity of LLMs for each criterion, complemented by a qualitative failure analysis to uncover the primary causes of any incorrect classifications. In the n2c2 dataset analysis, accuracy reached 87%, with sensitivity at 84% and specificity at 89% across 182 patients on 13 criteria. For 10 patients who were fully eligible across 14 criteria for the institutional phase II trial, accuracy and sensitivity were both high at 98%. For 10 ineligible patients, accuracy was 87%, sensitivity 87%, and specificity 86%. The LLM screening process took 1-5 minutes per patient. Failure analyses revealed that most incorrect classifications stemmed from challenges in document retrieval, temporal reasoning errors, or misinterpretations of criteria requirements. Our findings demonstrate that LLMs facilitate rapid and accurate clinical trial screening across both public and private datasets. This approach shows the ability to aid clinical research staff in the clinical trial screening process, improving enrollment, and reducing costs. Further investigation into more diverse environments and complex clinical trials are required to prove this approach's efficacy across disciplines. [ABSTRACT FROM AUTHOR]

Understanding informed consent forms, which outline the risks, costs, and procedures of clinical trials, presents a significant challenge for patients due to their complex language and length. Recognizing that such complexity can impede patient comprehension and decision-making, this study proposes the use of Large Language Models (LLMs) to distill these forms into concise, easy-to-understand cover pages. We hypothesize that we can significantly improve the readability of these forms using LLMs. Five approved institutional clinical trial consent forms were assessed for their readability using SMOG and the Flesch-Kincaid Ease of Reading score using Python libraries. The documents were segmented and catalogued in a vector database to facilitate similarity searches. OpenAI's GPT-4 API was prompted to extract information about costs, payments, contact information, eligibility criteria, treatments, duration, and requirements based on relevant information retrieved from the vector database. Furthermore, each prompt included instructions to respond at a 7th to 8th-grade reading level. The answers were collected, and their readability evaluated using the aforementioned metrics. Finally, the SMOG and Flesch-Kincaid readability scores were compared using a Wilcoxon signed-rank test. The original informed consent forms exhibited an average SMOG score of 16.74 (± 0.86), corresponding to a graduate reading level, and an average Flesch-Kincaid score of 12.93 (± 0.70), corresponding to an undergraduate reading level. Conversely, the LLM-generated information sheets exhibited an average SMOG score of 11.25 (± 0.50), corresponding to a high school reading level, and an average Flesch-Kincaid score of 8.15 (± 0.68), corresponding to an 8-9th grade reading level, demonstrating a significant reduction in reading level (p<0.05 for both SMOG and Flesh-Kincaid scores). Some LLM-generated information sheets were successful in adhering to a 7th to 8th-grade reading level, as denoted by a Flesch-Kincaid score < 8. Our approach successfully simplified clinical trial consent forms using LLMs, reducing reading levels and making information more accessible. This approach could significantly enhance the transparency and comprehensibility of the clinical trial consent process, fostering a more patient-centric approach. Feedback from patients on LLM-generated information sheets could provide invaluable insights into the practicality and usefulness of this approach. [ABSTRACT FROM AUTHOR]

Large language models (LLMs) have demonstrated potential as a tool in medical practice and education. However, their performance in radiation oncology requires further study. The present study aimed to evaluate the knowledge level and ability to interpret questions of three popular LLMs: OpenAI's ChatGPT-3.5 and GPT- 4, and Meta's Llama-2 in the field of cancer and radiation biology through a multiple-choice question (MCQ) examination format. The 2023 ASTRO Radiation and Cancer Biology Exam Study Guide was used to evaluate the performance of three LLMs: ChatGPT-3.5, GPT- 4, and Llama-2. The practice examination is comprised of 337 questions. Two questions containing graphical content were excluded. The exam questions were classified into the ten major topics and whether they required the use of math. The multiple-choice questions were individually entered verbatim into each LLM's chat interface, and answers were recorded and graded as either correct or incorrect. Performance of the LLMs was assessed in relation to overall scores, topic, and use of math. The word count of the response and level of detail was recorded. Reliability of the LLMs was also assessed by inputting each question into the LLMs five separate times and recording the number of unique answers. Statistical analysis was performed using one-way analysis of variance (ANOVA) followed by a post-hoc analysis using the Bonferroni correction to adjust for multiple comparisons. Biology practice exam was 221 of 335 (62%), 271 of 335 (81%), and 175 of 335 (51%), respectively. Thus, GPT-4 significantly outperformed GPT-3.5 (p < 0.001) and Llama-2 (p < 0.001). Overall, all three LLMs were strongest in domains related to molecular and cellular biology (molecular and cellular damage, tumor biology and microenvironment, and cancer biology). They performed poorly in domains related to the effects and delivery of radiation (interaction of radiation with matter, dose delivery, and late effects and radiation protection). All three LLMs performed significantly worse on questions requiring the use of math, with GPT-3.5, GPT-4, and Llama-2 scoring 26% (6/23), 57% (13/23), and 34% (8/23), respectively. The average word count of the response explanations provided by GPT-3.5, GPT-4, and Llama-2 were 134 words, 183 words, 162 words, respectively. When each question was inputted into each LLM five times, GPT3.5 outputted >1 distinct response for 39% of questions (132/335), GPT4 outputted >1 distinct response for 20% of questions (70/335), and Llama-2 outputted >1 distinct response for 19% of questions (62/335). Thus, despite having the lowest overall exam score, Llama-2 was the most consistent with its responses (p < 0.01). GPT-4 showed remarkable performance despite the lack of radiation and cancer biology specific training. While LLMs exhibit potential, these findings underscore the need for rigorous oversight and validation of LLM-generated information related to radiation and cancer biology. [ABSTRACT FROM AUTHOR]

Utilizing OpenAI-based chatbot's ability to retrieve relevant patient information from physician consultation notes. Ten unique questions were assessed on 10 diverse patients notes covering various cancer treatment sites. Sample questions included asking the AI chatbot a patient's name, age, gender, date of birth, presenting symptoms, diagnosis, date of diagnosis, previous medical and surgical history, familial oncologic history, previous oncologic treatment, KPS score, treatment recommendations, and if the patient smokes and is continuing to smoke. The questions were designed to represent realistic relevant information for Radiation Oncology chart rounds, tumor boards, and physician-handoffs. A chatbot system was developed using OpenAI's GPT-4 model to generate responses. A custom rubric system with a rating from 1 to 5 (1 = strongly disagree, 5 = strongly agree) was developed, which included several criteria: accuracy of content, completeness, context awareness, transparency of limitations, and lack of bias. Descriptive statistics were performed to evaluate the overall performance of the system. The overall performance of the chatbot achieved an average rubric score of 4.5 (± 0.9) out of 5 across all criteria. The results of descriptive statistics are as follows: the mean score (standard deviation) for accuracy of content is 4.4 (± 1.1), completeness is 4.4 (± 1.1), context awareness is 4.4 (± 1.0), transparency of limitations is 4.5 (± 0.9), and lack of bias 4.7 (± 0.5). The 'Lack of Bias' category received the highest average score and the lowest standard deviation, indicating a strong consensus on the system's ability to provide unbiased responses. In contrast, 'Accuracy of Content' and 'Completeness' had the highest standard deviations, suggesting more variability in the responses from OpenAI and identifying these areas as potential focal points for enhancement. Our research indicates that OpenAI's chatbot is a useful tool, specifically for retrieving patient information from consultations for tumor boards, chart rounds, and physician-handoffs. However, despite the excitement around using AI in clinical settings, it is crucial to conduct a comprehensive evaluation before its implementation to fully understand the limitations of such systems. A rubric similar to ours can provide valuable insights into the strengths and weaknesses of these systems. [ABSTRACT FROM AUTHOR]

Safety is an abstract concept that cannot be directly observed, and the Radiation Oncology Incident Learning System (RO-ILS) is designed as a surrogate to enable risk-informed safety assessment and management (RISAM) for enhancing patient safety measures in radiotherapy. Establishing a robust patient safety learning system necessitates assimilating all available, relevant, and adequately evaluated data (ARAED), including incidents from multi-institutional sources. However, HIPAA mandates that any shared medical information must not contain Protected Health Information (PHI). This work aims to develop an artificial intelligence framework leveraging state-of-the-art Large Language Models (LLMs) to effectively and efficiently de-identify PHI, potentially facilitating data-driven solutions for RISAM. LLM features interpretability, allowing us to utilize prompts to query information with semantic reasoning. We integrated a generative pre-trained transformer (GPT)-based open source LLM (LLaMa2-70B) with HIPAA identifiers into the proposed RISAM framework for de-identifying electronic PHI. The LLaMa2 with 70 billion pre-trained parameters based on 2 trillion tokens underwent validation by the institutional safety database with 1222 incidents happened between 2015 and 2023. The evaluation encompassed two objectives: 1) assessing model flexibility for potential multi-institutional collaboration, and 2) evaluating model robustness for semantic reasoning across data from different hospitals. A zero-shot testing was conducted using unstructured text data. The de-identified results were comprehensively reviewed by the institutional quality committee to ensure accuracy in fulfilling HIPAA guidelines. The evaluated GPT-based LLM achieved the zero-shot accuracy of 95.2% in de-identification. After the review of the internal quality committee, most mis-redacted content included unstructured dates without years (MM/DD) and ages followed by acronym of years (YO). Major HIPAA identifiers were redacted to protect patients' privacy, including but not limited to names, dates, and medical record numbers. LLaMa2-70B demonstrated highly accurate de-identification capabilities for unstructured text-based incident reports. This validation is crucial for the continued development of the RISAM framework, which aims to effectively analyze trends and root causes of medical incidents, utilizing the findings as medical education resources. A robust de-identification method can facilitate the development of a trustworthy LLM-based RISAM framework by assimilating ARAED. Future investigations will focus on deploying a radiation oncology-specific LLM-based safety framework as a virtual safety assistant in clinical settings, enabling medical staff to query solutions during emergent safety incidents. [ABSTRACT FROM AUTHOR]

This study introduces an innovative methodology for rationale-based distillation in textual entailment. Central to our methodology is the use of diverse and deep rationale types generated by large language models, eliminating the need for explicit feature engineering between text-hypothesis pairs. Through extensive experimentation, we demonstrate the effectiveness of our rationale-enhanced distillation process, underpinned by a comprehensive study of rationales. Remarkably, our model, which utilizes the most impactful rationales and operates with 795 times fewer parameters, exhibits competitive performance, especially in contexts limited by resource availability. Specifically, our model significantly surpasses RoBERTa, FLAN-T5 Large, and GPT-3.5 in performance. Our findings underscore the potential of rationale-based approaches, which improve textual entailment modeling and pave the way for future research. These techniques could be applied to other areas of NLP and beyond. This study's contributions are poised to benefit researchers and practitioners seeking to leverage the power of large language models in augmenting reasoning capabilities with rationales. The novelty of our approach lies in its ability to deliver high performance with significantly fewer computational resources, making it a valuable advancement in the field. [ABSTRACT FROM AUTHOR]

Power system optimization problems which include the nonlinear AC power flow equations require powerful and robust numerical solution algorithms. Within this sub-field of nonlinear optimization, interior point methods have come to dominate the solver landscape. Over the last decade, however, a number of efficient numerical optimizers have emerged from the field of Machine Learning (ML). One algorithm in particular, Adam, has become the optimizer-of-choice for a massive percentage of ML training problems (including, e.g., the training of GPT-3), solving some of the largest unconstrained optimization problems ever conceived of. Inspired by such progress, this paper designs a parallelized Adam-based numerical solver to overcome one of the most challenging power system optimization problems: security and reserve constrained AC Unit Commitment. The resulting solver, termed QuasiGrad, recently competed in the third ARPA-E Grid Optimization (GO3) competition. In the day-ahead market clearing category (with systems ranging from 3 to 23,643 buses over 48 time periods), QuasiGrad's aggregated market surplus scores were within 5% of the winningest market surplus scores. The QuasiGrad solver is now released as an open-source Julia package: QuasiGrad.jl. The internal gradient-based solver (Adam) can easily be substituted for other ML-inspired solvers (e.g., AdaGrad, AdaDelta, RMSProp, etc.). Test results from large experiments are provided. • An Adam-based solver, called QuasiGrad, is introduced to solve AC unit commitment. • QuasiGrad exploits parallelized projection methodologies to accelerate convergence. • QuasiGrad stochastically backpropagates through parallelized security constraints. • QuasiGrad is released as an open-source Julia package. [ABSTRACT FROM AUTHOR]

Prompt learning has demonstrated impressive efficacy in the fine-tuning of multimodal large models to a wide range of downstream tasks. Nonetheless, applying existing prompt learning methods for the diagnosis of neurological disorder still suffers from two issues: (i) existing methods typically treat all patches equally, despite the fact that only a small number of patches in neuroimaging are relevant to the disease, and (ii) they ignore the structural information inherent in the brain connection network which is crucial for understanding and diagnosing neurological disorders. To tackle these issues, we introduce a novel prompt learning model by learning graph prompts during the fine-tuning process of multimodal models for diagnosing neurological disorders. Specifically, we first leverage GPT-4 to obtain relevant disease concepts and compute semantic similarity between these concepts and all patches. Secondly, we reduce the weight of irrelevant patches according to the semantic similarity between each patch and disease-related concepts. Moreover, we construct a graph among tokens based on these concepts and employ a graph convolutional network layer to extract the structural information of the graph, which is used to prompt the pre-trained multimodal models for diagnosing neurological disorders. Extensive experiments demonstrate that our method achieves superior performance for neurological disorder diagnosis compared with state-of-the-art methods and validated by clinicians. • We propose a prompt-driven multimodal model for diagnosis. • We design a concept learning method to tackle irrelevant patches. • We design a graph prompt incorporating the pathogenesis of neurological disorder. [ABSTRACT FROM AUTHOR]

Over the last few years, graph-based methods have manifested a significant enhancement in document mining applications such as spam detection, news recommendation, and legal document classification. However, existing graph-based methods have a limited ability to utilize word position and multi-element information within the documents, limiting their effectiveness in practical application. To mitigate this limitation, we propose a novel multi-element hypergraph gated attention network that can capture word position and multi-element information for accurate document classification. Specifically, a new multi-element hypergraph is first proposed to describe the word position, sentence, and full content within the document. Then, a new multi-element homogenization module is applied to mitigate heterogeneity of constructed hypergraph. Meantime, a new hypergraph gated attention module is proposed to filter noise in the constructed hypergraph and derive various element representations that incorporate word position information. Finally, a new block-wise read-out module is designed to fuse learned element representations into comprehensive document representations for classification. Extensive experiments conducted on several real-world datasets demonstrate that the proposed method not only outperforms related state-of-the-art methods but is also faster, making it suitable for a wide range of practical applications. For instance, our method achieved an accuracy improvement of 1.1 % over the best comparative method on some datasets while also operating at a faster speed. Additionally, it demonstrated an impressive 14 % improvement in accuracy over the well-known Generative Pre-trained Transformer 3.5 (GPT-3.5) on one dataset. [ABSTRACT FROM AUTHOR]

In the age of information overload, the ability to distill essential content from extensive texts is invaluable. DeepExtract introduces an advanced framework for extractive summarization, utilizing the groundbreaking capabilities of GPT-4 along with innovative hierarchical positional encoding to redefine information extraction. This manuscript details the development of DeepExtract, which integrates semantic-driven techniques to analyze and summarize complex documents effectively. The framework is structured around a novel hierarchical tree construction that categorizes sentences and sections not just by their physical placement within a text, but by their contextual and thematic significance, leveraging dynamic embeddings generated by GPT-4. We introduce a multi-faceted scoring system that evaluates sentences based on coherence, relevance, and novelty, ensuring that summaries are not only concise but rich with essential content. Further, DeepExtract employs optimized semantic clustering to group thematic elements, which enhances the representativeness of the summaries. This paper demonstrates through comprehensive evaluations that DeepExtract significantly outperforms existing extractive summarization models in terms of accuracy and efficiency, making it a potent tool for academic, professional, and general use. We conclude with a discussion on the practical applications of DeepExtract in various domains, highlighting its adaptability and potential in navigating the vast expanses of digital text. [ABSTRACT FROM AUTHOR]

• GPT-4 accurately detected postoperative complications from renal surgeries. • GPT-4 performed worse for Clavien-Dindo classification and institutional categories. • GPT-4 may struggle with more complex reasoning tasks. • GPT-4 can be used with human oversight to help identify postoperative complications. Large language models, a subset of artificial intelligence, have immense potential to support human tasks. The role of these models in science and medicine is unclear, requiring strong critical thinking and analysis skills. The objective of our study was to evaluate GPT-4's abilities to assess postoperative complications after renal surgeries. Discharge summaries were compiled, and patient information was deidentified in a Python-based program. Prompts were engineered in GPT-4 to assess for the presence of postoperative complications. GPT-4 was further asked to interpret each complication's Clavien-Dindo classification and institutional-specific category. GPT-4's database was compared to a human-curated database. Discrepancies were manually reviewed to calculate match and accuracy rates. Approximately 944 renal surgeries were conducted from August 2005 to March 2022. There was a 79.6% match rate between GPT-4 and human-curated data in detecting postoperative complications. Accuracy rates were 86.7% for GPT-4 and 92.9% for human-curated. A subgroup of 139 patients had a complication detected by both GPT-4 and human with available Clavien-Dindo classification and category information. There was a 37.4% overall match rate for Clavien-Dindo grade and 55.4% match rate for category. GPT-4 was able to accurately detect if there were any postoperative complications. It struggled with the complex task of further analyzing complications, especially with Clavien-Dindo classification, which requires more critical thinking and interpretation. While GPT-4 is not yet ready for advanced postoperative complication analysis, it can still be used to support clinicians in this endeavor. [ABSTRACT FROM AUTHOR]

The vast volume of currently available unstructured text data, such as research papers, news, and technical report data, shows great potential for ecological research. However, manual processing of such data is labour-intensive, posing a significant challenge. In this study, we aimed to assess the application of three state-of-the-art prompt-based large language models (LLMs), GPT-3.5, GPT-4, and LLaMA-2-70B, to automate the identification, interpretation, extraction, and structuring of relevant ecological information from unstructured textual sources. We focused on species distribution data from two sources: news outlets and research papers. We assessed the LLMs for four key tasks: classification of documents with species distribution data, identification of regions where species are recorded, generation of geographical coordinates for these regions, and supply of results in a structured format. GPT-4 consistently outperformed the other models, demonstrating a high capacity to interpret textual data and extract relevant information, with the percentage of correct outputs often exceeding 90% (average accuracy across tasks: 87–100%). Its performance also depended on the data source type and task, with better results achieved with news reports, in the identification of regions with species reports and presentation of structured output. Its predecessor, GPT-3.5, exhibited slightly lower accuracy across all tasks and data sources (average accuracy across tasks: 81–97%), whereas LLaMA-2-70B showed the worst performance (37–73%). These results demonstrate the potential benefit of integrating prompt-based LLMs into ecological data assimilation workflows as essential tools to efficiently process large volumes of textual data. • Large Language Models (LLMs) are tested for extracting biodiversity data from text. • Tasks include identification of relevant sources, geocoding, and data structuring. • LLMs, especially GPT-4, show over 90% accuracy on multiple tasks. • LLMs promise effective data assimilation from unstructured text in ecology. [ABSTRACT FROM AUTHOR]

This paper presents a novel method for utilizing fine-tuned Large Language Models (LLMs) to minimize data requirements in load profile analysis, demonstrated through the restoration of missing data in power system load profiles. A two-stage fine-tuning strategy is proposed to adapt a pre-trained LLMs, i.e., GPT-3.5, for missing data restoration tasks. Through empirical evaluation, we demonstrate the effectiveness of the fine-tuned model in accurately restoring missing data, achieving comparable performance to state-of-the-art specifically designed models such as BERT-PIN. Key findings include the importance of prompt engineering and the optimal utilization of fine-tuning samples, highlighting the efficiency of few-shot learning in transferring knowledge from general user cases to specific target users. Furthermore, the proposed approach demonstrates notable cost-effectiveness and time efficiency compared to training models from scratch, making it a practical solution for scenarios with limited data availability and computing resources. Additionally, we applied fine-tuned LLM to load forecasting and showed its significant potential for application to other power system load profile analysis tasks. Consequently, it advances the use of LLMs in power system analytics, offering promising implications for enhancing the resilience and efficiency of power distribution systems. • We introduce a two-stage fine-tuning strategy which significantly reduces the data needs for load profile analysis. • The fine-tuned model achieves acceptable performance compared with state-of-the-art specifically designed models. • We evaluate the model with different fine-tuning strategies. This offers valuable insights for other data analysis tasks. • Fine-tuned LLMs offer the flexibility to adapt to various downstream tasks without extensive reprogramming. [ABSTRACT FROM AUTHOR]

State-of-the-art image captioning often relies on supervised models trained on domain-specific image–text pairs, which can be expensive and time-consuming to annotate. Additionally, these models struggle with performance drops when applied to data from different distributions. To overcome these challenges, zero-shot image captioning has emerged, utilizing models trained on unlabeled image and text data. Pre-trained vision-language and generative language models have shown promise in this area, but existing methods do not fully leverage the cross-modal representation capabilities of large-scale models and tend to overlook smaller objects in images. We propose EntroCap, a novel zero-shot image captioning method that incorporates an Entropy-based Retrieval Strategy. During training, our Hierarchical Projector optimizes CLIP's joint embeddings to capture comprehensive global signals with multi-level representations. During inference, the Entropy-based Retrieval Strategy adjusts prediction logits to better include smaller, less frequent objects. Additionally, our Balancing Gate mechanism guides the language model to produce more detailed and contextually relevant captions by balancing local and global signals. Extensive experiments on MSCOCO, Flickr30k, and NoCaps demonstrate that EntroCap outperforms previous zero-shot approaches in both in-domain and cross-domain settings. We also conduct experiments to quantitatively validate EntroCap's effectiveness in describing small objects. • EntroCap retrieves small targets and captions unseen images in zero-shot settings. • A Hierarchical Projector learns global signals from CLIP for better discrimination. • An Entropy-based Strategy adjusts logits, capturing small targets in local signals. • A Balancing Gate refines GPT-2 for more accurate description generation. • EntroCap achieves state-of-the-art transferability in cross-domain captioning. [ABSTRACT FROM AUTHOR]

Large language models (LLMs) have exhibited great potential in fault diagnosis of heating, ventilation, and air conditioning systems. However, the fault diagnosis accuracy of LLMs is still unsatisfactory, due to the lack of effective diagnosis accuracy enhancement methods for LLMs. To fill this gap, this study proposes a LLM fine-tuning method supervised by data with fault and fault-free labels to enhance the fault diagnosis accuracy of LLMs. This method designs a LLM self-correction strategy to automatically generate a fine-tuning dataset based on the labeled data. The generated fine-tuning dataset is applied to fine-tune a LLM. Moreover, a data augmentation-based approach is put forward to adaptively update the fine-tuning dataset for iteratively developing a high-performance fine-tuned LLM. The proposed method is utilized to fine-tune the GPT-3.5 model using the air handling unit (AHU) fault dataset from the RP-1312 project. The results show that the diagnosis accuracy of the GPT-3.5 model is increased from 29.5 % to 100.0 % after model fine-tuning. Compared with the GPT-4 model, the fine-tuned GPT-3.5 model achieves a 31.1 % higher average diagnosis accuracy. The fine-tuned GPT-3.5 model is also applied to diagnose faults in two AHUs from another open-source dataset to verify the generalization ability of this model. The two AHUs have different system structures and sensor configurations compared to the AHU in the RP-1312 dataset, and this dataset is not utilized to fine-tune the GPT-3.5 model. The average diagnosis accuracy of the GPT-3.5 model is increased from 46.0 % to 99.1 % and from 38.8 % to 98.9 % for the faults in the two AHUs, respectively, after model fine-tuning. Furthermore, the proposed method is verified using two fault datasets from a variable air volume box and a chiller plant system. After fine-tuning the GPT-3.5 model using the two datasets, the average diagnosis accuracy of this model is increased from 33.0 % to 98.3 % for variable air volume box faults and from 36.0 % to 99.1 % for chiller plant system faults. This study provides an effective solution to the development of domain-specific LLMs for this domain. • A labeled-data-supervised fine-tuning method is proposed for large language models. • A self-correction strategy is proposed to automatically generate fine-tuning datasets. • A data augmentation approach is designed to adaptively update fine-tuning datasets. • Fine-tuned models show nearly 100 % diagnosis accuracy for air handling unit faults. • Fine-tuned models have strong generalization abilities and are dimension-agnostic. [ABSTRACT FROM AUTHOR]

Accurate real-time object detection is vital across numerous industrial applications, from safety monitoring to quality control. Traditional approaches, however, are hindered by arduous manual annotation and data collection, struggling to adapt to ever-changing environments and novel target objects. To address these limitations, this paper presents DART, an innovative automated end-to-end pipeline that revolutionizes object detection workflows from data collection to model evaluation. It eliminates the need for laborious human labeling and extensive data collection while achieving outstanding accuracy across diverse scenarios. DART encompasses four key stages: (1) Data D iversification using subject-driven image generation (DreamBooth with SDXL), (2) A nnotation via open-vocabulary object detection (Grounding DINO) to generate bounding box and class labels, (3) R eview of generated images and pseudo-labels by large multimodal models (InternVL-1.5 and GPT-4o) to guarantee credibility, and (4) T raining of real-time object detectors (YOLOv8 and YOLOv10) using the verified data. We apply DART to a self-collected dataset of construction machines named Liebherr Product, which contains over 15K high-quality images across 23 categories. The current instantiation of DART significantly increases average precision (AP) from 0.064 to 0.832. Its modular design ensures easy exchangeability and extensibility, allowing for future algorithm upgrades, seamless integration of new object categories, and adaptability to customized environments without manual labeling and additional data collection. The code and dataset are released at https://github.com/chen-xin-94/DART. • DART, an automated end-to-end object detection pipeline without manual labeling. • DART streamlines data diversification, annotation, review, and training stages. • DART boosts AP from 0.064 to 0.832 on a collected dataset of construction machines. • DART's modular design ensures flexibility, robustness, and customization. • Code and dataset are made publicly available for further research. [ABSTRACT FROM AUTHOR]

Large language models(LLMs) have shown promising performance for various downstream tasks, especially machine translation. However, LLMs and Specialized Translation Models (STMs) are designed to handle general translation needs, they are not well-suited for domains with specialized terms and writing styles, such as e-commerce, legal, and medicine. In the e-commerce domain, the text often contains many domain-specific terms and keyword-stacked structures, leading to poor translation quality with existing NMT methods. To tackle these problems, we have collected two resources specifically for the e-commerce domain, including aligned Chinese-English bilingual terms and parallel corpus from real e-commerce scenarios for model fine-tuning. We propose an LLMs-based E-commerce machine translation approach(LEMT) which includes LLMs utilization, e-commerce resources collection, and tokenizer optimization. We conduct two-stage fine-tuning and self-contrastive enhancement based on general LLMs to enable the model to learn translation features in the e-commerce domain. Through comprehensive evaluations on real e-commerce titles, our LEMT methodology demonstrates superior translation quality and robustness, outperforming leading NMT models such as NLLB, LLaMA, and even GPT-4. • Existing universal translation models perform poorly in the e-commerce field. • Translation models based on LLMs improve translation capabilities more efficiently. • Resources based on e-commerce texts are conducive to e-commerce translation. • Two-step fine-tuning gradually enhance the translation ability of specialized domain. [ABSTRACT FROM AUTHOR]

Mandarin Chinese, a globally prevalent language, boasts an abundance of regularly refreshed short news texts accessible online. Consequently, devising concise summaries of these texts has emerged as a pivotal challenge for enhancing information dissemination and comprehension efficiency. To tackle this issue, we introduce DMSeqNet-mBART, an innovative model grounded in the mBART framework, positioning it as a state-of-the-art solution for Chinese short news summarization. DMSeqNet-mBART incorporates the Adaptive-DropMessage technique, an innovative methodology that intelligently discards or retains information contingent upon the attention mechanism's output. Furthermore, the model integrates several enhanced technologies, such as dynamic convolutional layers, gated residual connections, customized feed-forward networks enhanced with batch normalization, self-attention, and cross-attention, all aimed at bolstering the performance and robustness of Chinese short news summarization. Rigorous comparative experiments, conducted across six recognized Chinese short news summary datasets, demonstrate that DMSeqNet-mBART significantly surpasses industry-leading models like T5, MLC, PLCC, and GPT-4 in terms of fluency, completeness, robustness, and accuracy. These results, as validated by benchmarks including BERTScore, BLEU, and ROUGE metrics, underscore the model's superiority across diverse evaluation standards. • Introduces DMSeqNet-mBART for Chinese text summarization. • Integrates Adaptive-DropMessage, enhancing key info extraction. • Achieves SOTA in six key Chinese news summarization datasets. • Sheds light on complex Chinese text processing challenges. [ABSTRACT FROM AUTHOR]

Most current audio–visual datasets are class-relevant, where audio and visual modalities are annotated. Thus, current audio–visual recognition methods apply cross-modality attention or modality fusion. However, leveraging the audio modality effectively in vision-specific videos for human activity recognition is of particular challenge. We address this challenge by proposing a novel audio–visual recognition framework that effectively leverages audio modality in any vision-specific annotated dataset. The proposed framework employs language models (e.g., GPT-3, CPT-text, BERT) for building a semantic audio–video label dictionary (SAVLD) that serves as a bridge between audio and video datasets by mapping each video label to its most K-relevant audio labels. Then, SAVLD along with a pre-trained audio multi-label model are used to estimate the audio–visual modality relevance. Accordingly, we propose a novel learnable irrelevant modality dropout (IMD) to completely drop the irrelevant audio modality and fuse only the relevant modalities. Finally, for the efficiency of the proposed multimodal framework, we present an efficient two-stream video Transformer to process the visual modalities (i.e., RGB frames and optical flow). The final predictions are re-ranked with GPT-3 recommendations of the human activity classes. GPT-3 provides high-level recommendations using the labels of the detected visual objects and the audio predictions of the input video. Our framework demonstrated a remarkable performance on the vision-specific annotated datasets Kinetics400 and UCF-101 by outperforming most relevant human activity recognition methods. • Introduced a novel framework for improved human activity recognition using audio–visual data. • Employed pre-trained language models to bridge audio and video datasets. • Developed a learnable mechanism to selectively ignore irrelevant audio modalities. • Proposed an efficient video Transformer that processes visual data with fewer parameters. • Achieved superior performance on benchmark datasets, outperforming existing methods. [ABSTRACT FROM AUTHOR]

Large language models (LLMs) exhibit expert-level performance in tasks across a wide range of different domains. Ethical issues raised by LLMs and the need to align future versions makes it important to know how state of the art models reason about moral and legal issues. In this paper, we employ the methods of experimental psychology to probe into this question. We replicate eight studies from the experimental literature with instances of Google's Gemini Pro, Anthropic's Claude 2.1, OpenAI's GPT-4, and Meta's Llama 2 Chat 70b. We find that alignment with human responses shifts from one experiment to another, and that models differ amongst themselves as to their overall alignment, with GPT-4 taking a clear lead over all other models we tested. Nonetheless, even when LLM-generated responses are highly correlated to human responses, there are still systematic differences, with a tendency for models to exaggerate effects that are present among humans, in part by reducing variance. This recommends caution with regards to proposals of replacing human participants with current state-of-the-art LLMs in psychological research and highlights the need for further research about the distinctive aspects of machine psychology. [ABSTRACT FROM AUTHOR]

The computational logics of large language models (LLMs) or generative AI – from the early models of CLIP and BERT to the explosion of text and image generation via ChatGPT and DALL-E − are increasingly penetrating the social and political world. Not merely in the direct sense that generative AI models are being deployed to govern difficult problems, whether decisions on the battlefield or responses to pandemic, but also because generative AI is shaping and delimiting the political parameters of what can be known and actioned in the world. Contra the promise of a generalizable "world model" in computer science, the article addresses how and why generative AI gives rise to a model of the world , and with it a set of political logics and governing rationalities that have profound and enduring effects on how we live today. The article traces the genealogies of generative AI models, how they have come into being, and why some concepts and techniques that animate these models become durable forms of knowledge that actively shape the world, even long after a specific material commercial GPT model has moved on to a new iteration. Though generative AI retains significant traces of former scientific and computational regimes – in statistical practices, probabilistic knowledge, and so on – it is also dislocating epistemological arrangements and opening them to novel ways of perceiving, characterising, classifying, and knowing the world. Four defining aspects of the political logic of generative AI are elaborated: i) generativity as something more than the capacity to generate image or text outputs, so that a generative logic acts upon the world understood as estimates of "underlying distributions" in data; ii) latency as a political logic of compression in which (by contrast with claims to reduction or distortion) the thing that is hidden, unknown or latent becomes surfaced and amenable to being governed; iii) broken and parallelized sequences as the ordering device of the political logic of generative AI, where attention frameworks radically change the possibilities for governing non-linear problems; iv) pre-training and fine-tuning as a computational logic of generative AI that simultaneously shapes a "zero shot politics" oriented towards unencountered data and new tasks. Across each of the four aspects, the article maps the emerging contemporary political logic of generative AI. [ABSTRACT FROM AUTHOR]

• Revolutionizing drug discovery with deep attention neural networks. • Exploring attention mechanism and extended architectures like GATs, transformers, BERT, GPTs and BART for complex data. • Uncovering a pivotal role in catalyzing de novo drug design, predicting molecular properties and deciphering elusive drug–target interactions. • Addressing challenges to deepen understanding for pharmaceutical breakthroughs. In the dynamic field of drug discovery, deep attention neural networks are revolutionizing our approach to complex data. This review explores the attention mechanism and its extended architectures, including graph attention networks (GATs), transformers, bidirectional encoder representations from transformers (BERT), generative pre-trained transformers (GPTs) and bidirectional and auto-regressive transformers (BART). Delving into their core principles and multifaceted applications, we uncover their pivotal roles in catalyzing de novo drug design, predicting intricate molecular properties and deciphering elusive drug–target interactions. Despite challenges, these attention-based architectures hold unparalleled promise to drive transformative breakthroughs and accelerate progress in pharmaceutical research. [ABSTRACT FROM AUTHOR]

Recently, GPT-4 with Vision (GPT-4V) has demonstrated remarkable visual capabilities across various tasks, but its performance in emotion recognition has not been fully evaluated. To bridge this gap, we present the quantitative evaluation results of GPT-4V on 21 benchmark datasets covering 6 tasks: visual sentiment analysis , tweet sentiment analysis , micro-expression recognition , facial emotion recognition , dynamic facial emotion recognition , and multimodal emotion recognition. This paper collectively refers to these tasks as "Generalized Emotion Recognition (GER)". Through experimental analysis, we observe that GPT-4V exhibits strong visual understanding capabilities in GER tasks. Meanwhile, GPT-4V shows the ability to integrate multimodal clues and exploit temporal information, which is also critical for emotion recognition. However, it is worth noting that GPT-4V is primarily designed for general domains and cannot recognize micro-expressions that require specialized knowledge. To the best of our knowledge, this paper provides the first quantitative assessment of GPT-4V for GER tasks. We have open-sourced the code and encourage subsequent researchers to broaden the evaluation scope by including more tasks and datasets. Our code and evaluation results are available at: https://github.com/zeroQiaoba/gpt4v-emotion. • This is the first work to quantitatively evaluate GPT-4V's performance in emotion recognition. • We conduct a comprehensive analysis of GPT-4V's multi-faceted capabilities. • This paper provides guidance for future directions following the emergence of GPT-4V. • This paper can serve as a zero-shot benchmark for subsequent research. • We have open-sourced the code and evaluation results. [ABSTRACT FROM AUTHOR]

Existing studies apply Large Language Model (LLM) to knowledge-based Visual Question Answering (VQA) with encouraging results. Due to the insufficient input information, the previous methods still have shortcomings in constructing the prompt for LLM, and cannot fully activate the capacity of LLM. In addition, previous works adopt GPT-3 for inference, which has expensive costs. In this paper, we propose PCPA: a framework that P rompts LLM with C ontext and P re- A nswer for VQA. Specifically, we adopt a vanilla VQA model to generate in-context examples and candidate answers, and add a pre-answer selection layer to generate pre-answers. We integrate in-context examples and pre-answers into the prompt to inspire the LLM. In addition, we choose LLaMA instead of GPT-3, which is an open and free model. We build a small dataset to fine-tune the LLM. Compared to existing baselines, the PCPA improves accuracy by more than 2.1 and 1.5 on OK-VQA and A-OKVQA, respectively. • We propose a novel framework that prompts LLM for knowledge-based VQA. • We add dynamic routing to vanilla VQA model to further inspire the LLM. • we add a pre-answer selection layer to generate more suitable pre-answers. • We build a small dataset for fine-tuning LLM. [ABSTRACT FROM AUTHOR]

Patients are increasingly using Generative Pre-trained Transformer 4 (GPT-4) to better understand their own radiology findings. To evaluate the performance of GPT-4 in transforming cardiovascular magnetic resonance (CMR) reports into text that is comprehensible to medical laypersons. ChatGPT with GPT-4 architecture was used to generate three different explained versions of 20 various CMR reports (n = 60) using the same prompt: "Explain the radiology report in a language understandable to a medical layperson". Two cardiovascular radiologists evaluated understandability, factual correctness, completeness of relevant findings, and lack of potential harm, while 13 medical laypersons evaluated the understandability of the original and the GPT-4 reports on a Likert scale (1 "strongly disagree", 5 "strongly agree"). Readability was measured using the Automated Readability Index (ARI). Linear mixed-effects models (values given as median [interquartile range]) and intraclass correlation coefficient (ICC) were used for statistical analysis. GPT-4 reports were generated on average in 52 s ± 13. GPT-4 reports achieved a lower ARI score (10 [9–12] vs 5 [4–6]; p < 0.001) and were subjectively easier to understand for laypersons than original reports (1 [1] vs 4 [4,5]; p < 0.001). Eighteen out of 20 (90%) standard CMR reports and 2/60 (3%) GPT-generated reports had an ARI score corresponding to the 8th grade level or higher. Radiologists' ratings of the GPT-4 reports reached high levels for correctness (5 [4, 5]), completeness (5 [5]), and lack of potential harm (5 [5]); with "strong agreement" for factual correctness in 94% (113/120) and completeness of relevant findings in 81% (97/120) of reports. Test-retest agreement for layperson understandability ratings between the three simplified reports generated from the same original report was substantial (ICC: 0.62; p < 0.001). Interrater agreement between radiologists was almost perfect for lack of potential harm (ICC: 0.93, p < 0.001) and moderate to substantial for completeness (ICC: 0.76, p < 0.001) and factual correctness (ICC: 0.55, p < 0.001). GPT-4 can reliably transform complex CMR reports into more understandable, layperson-friendly language while largely maintaining factual correctness and completeness, and can thus help convey patient-relevant radiology information in an easy-to-understand manner. [Display omitted] [ABSTRACT FROM AUTHOR]

Generative artificial intelligence (GAI) is a rapidly growing field with a wide range of applications. In this paper, a thorough examination of the research landscape in GAI is presented, encompassing a comprehensive overview of the prevailing themes and topics within the field. The study analyzes a corpus of 1319 records from Scopus spanning from 1985 to 2023 and comprises journal articles, books, book chapters, conference papers, and selected working papers. The analysis revealed seven distinct clusters of topics in GAI research: image processing and content analysis, content generation, emerging use cases, engineering, cognitive inference and planning, data privacy and security, and Generative Pre-Trained Transformer (GPT) academic applications. The paper discusses the findings of the analysis and identifies some of the key challenges and opportunities in GAI research. The paper concludes by calling for further research in GAI, particularly in the areas of explainability, robustness, cross-modal and multi-modal generation, and interactive co-creation. The paper also highlights the importance of addressing the challenges of data privacy and security in GAI and responsible use of GAI. [ABSTRACT FROM AUTHOR]

[Display omitted] • Farmers must be an allrounder for plant growth, plant protection, legislation and different economic fields. • AI service integration holds great potential for assistance, documentation, education, interpretation forecasts or data-driven predictions. • LLMs depict a fundamental step to reduce the gap between AI-driven data analysis and common user. • Reproduction of processed data would increase confidence in the model support farmer's interpretation and guide the LLM for more precise results. • Introducing these technologies requires not only new training but also significant effort in structural transformation. Since the launch of the "Generative Pre-trained Transformer 3.5", ChatGPT by Open, artificial intelligence (AI) has been a main discussion topic in public. Especially large language models (LLM), so called "intelligent" chatbots, and the possibility to automatically generate highly professional technical texts get high attention. Companies, as well as researchers, are evaluating possible applications and how such a powerful LLM can be integrated into daily work and bring benefits, improve their business or to make the research outcome more efficient. In general, underlying models are trained on large datasets, mainly on sources from websites, and online books and articles. In combination with information provided by the user, the model can give an impressively fast response. Even if the range of questions and answers look unrestricted, there are limits to the models. In this paper, possible use cases for agricultural tasks are elucidated. This includes the textual preparation of facts, consulting tasks, interpretation of decision support models in plant disease management, as well as guides for tutorials to integrate modern digital techniques into agricultural work. Opportunities and challenges are described, as well as limitations and insufficiencies. The authors describe a map of easy-to-reach topics in agriculture where the integration of LLMs seems to be very likely within the next few years. [ABSTRACT FROM AUTHOR]