While there is a growing interest in applying large language models to recommendations, the discussion of fairness is still in its infancy. This study focuses on the memorization of training data, which is pointed out as one of the issues of large language models, for analyzing generative recommendations. Specifically, we finetuned Llama3 on log data from the Japanese news media to predict the transition of viewed articles and quantified the memorization of training data. The results suggested that there is a bias that excessively recommends popular articles, and this can be interpreted from the view of memorizing training data. We also demonstrated that deduplication, a technique for mitigating memorization bias, can be used to reduce the popularity bias in generative recommendations.

Although large language models (LLMs) have been applied in a wide range of fields in society, ``jailbreak'' attacks that exploit their vulnerability have raised serious security concerns. Jailbreak employs strategic prompts to circumvent their safeguards and induces outputs that are not intended by developers. Jailbreak strategies have rapidly evolved and diversified, making it nearly impossible to comprehensively address them during the training phase. Therefore, growing attention has been paid to in-context defenses, which aim to prevent inappropriate outputs by adding ideal responses or expected behaviors to user inputs. However, these additional prompts degrade the quality of responses in existing methods, for example, by causing refusal of responses even to normal inputs, which presents a significant challenge for practical application. This paper proposes a novel method for in-context defenses called ``Role Specification.'' In experiments using Llama-2-7b-chat, our proposed method (1) demonstrated superior defense performance against jailbreaks without compromising response quality, and (2) accomplished a more favorable trade-off between response quality and defense performance in conjunction with existing methods.

This research proposes LLM-AdaMerge, a method that extends AdaMerging to efficiently merge multiple specialized Large Language Models (LLMs). While existing LLM merging methods typically avoid using training data due to computational costs, leading to suboptimal task interactions, our approach enables data-driven optimization with efficient computational overhead. We introduce a language modeling loss function that directly optimizes weight coefficients for combining task-specific parameter differences, requiring only 4 samples per task for effective training. Experiments with three specialized Mistral-7B based models (mathematics, code generation, and Japanese language) demonstrate that our method achieves up to 12.95 points improvement in average accuracy compared to baselines. The results show superior performance over both non-data-driven methods and Bayesian optimization approaches, while maintaining computational efficiency through weight-only updates. Our method provides a practical solution for combining multiple specialized LLMs, though scaling to larger numbers of tasks remains challenging.

Techniques for vectorizing sentences and documents have become indispensable for developing various natural language processing applications, such as information retrieval and document classification. However, previous studies have pointed out that the quality of sentence vectors deteriorates as sentence length increases. This paper demonstrates that this degradation is caused by changes in the likelihood of function and content words appearing as sentences become longer. First, we empirically and theoretically demonstrate that the proportion of content words decreases in longer texts. Next, we demonstrate, both theoretically and empirically, that this decrease in content word proportion reduces the distance between sentence vectors, even for sentences on different topics. Building on these two analyses, we discuss how sentence vector quality declines for longer sentences. Our findings highlight the necessity of techniques that dynamically enhance the influence of content words based on sentence length.

In recent years, large language models (LLM) have been advancing rapidly worldwide, emphasizing the growing importance of cultivating capabilities within Japan. This paper presents an LLM development project led by the Matsuo and Iwasawa Lab as part of the GENIAC project, whose primary goal is to foster domestic expertise and reinforce national development capacity. Volunteers from the public worked with the lab to create 8B and 8×8B models from scratch. When we began our research in April 2024, domestically developed models still faced certain challenges in dialogue and text generation. On the other hand, our approach focused on improving dialogue and composition through synthetic data. Evaluations using the widely recognized “Japanese MT-Bench” indicated that our 8B model surpassed existing 10B-class models, while our 8×8B model performed on par with GPT-3.5, placing it at the forefront among domestically developed LLMs. Both models and their training code have been released under the Apache License 2.0, contributing to academic research and industrial applications of Japanese LLMs.

This study examines the feasibility of applying semantic entropy-based evaluation models across domains, from recipe to mathematical texts. Initial experiments showed that estimating semantic entropy for mathematical solutions is challenging. To address this, we leveraged Semantic Entropy Probes (SEPs) and hypothesized that recipe texts, which share a multi-step structure akin to mathematical texts, could provide an indirect estimation method. We thus trained a logistic regression model on recipe texts and tested its applicability to mathematical texts. Results suggest this model at least partially estimates uncertainty in mathematical solutions.

In buzzer quizzes, retriever-reader models are commonly used, where a retriever extracts relevant information and a reader derives the answer. However, traditional retriever training uses full questions, while buzzer quiz players only hear partial questions, causing a mismatch. To address this problem, we limit question length during training and apply curriculum learning. This approach successfully creates a model that performs well at various points in the question. Additionally, our analysis reveals that the retriever’s performance varies based on question format and training length compatibility.

With the advancement of natural language processing technologies, dialogue systems that handle continuous speech are becoming increasingly prevalent. In particular, the responses of dialogue systems that provide backchanneling can disrupt natural conversation due to delays in response speed and interruptions during speech. However, evaluating these systems is challenging because it is difficult to separate backchanneling from the main dialogue. In this study, we focus on turn-taking to achieve natural interactions that include backchanneling, and we have developed a dialogue system utilizing Voice Activity Projection (VAP). This system predicts the start and end times of conversations, allowing for the distinction between backchanneling and interruptive speech. Experiments have confirmed improvements in naturalness, indicating its effectiveness for future dialogue system development.

In dialogue systems utilizing large language models (LLMs), generating high-quality interjections is essential to improve responsiveness and empathy. In this study, to support the development of dialogue systems, we constructed a system capable of quantitatively and automatically evaluating the quality of interjections. To examine quality evaluation criteria for interjections, we created dialogue scripts and conducted subjective evaluations using a pairwise comparison method for the interjections mentioned in them. Variations of interjections were prepared based on elements such as coherence with context, consistency of tone, and length. Additionally, we used ChatGPT as an automatic evaluation environment to build a system that scores the quality of interjections. As a result of subjective evaluations through pairwise comparisons, data on variations of interjections were obtained, revealing human evaluation scales for coherence with context, tone consistency, and length. Furthermore, we investigated the correlation between the results of the automatic evaluation system and subjective evaluations to assess the effectiveness of this automatic evaluation system.

Wani is an automated theorem prover for Dependent Type Theory (DTT), which is part of the pipeline recently proposed as a "linguistic" approach to natural language inference. However, the execution speed of wani is one of the bottlenecks in the pipeline, since proof-search in DTT is undecidable. We have therefore set up a project called Neural Wani to develop a neural network model that will increase the speed of wani by implementing a neural classifier that selects the next inference rule to apply. In this study, as part of the Neural Wani project, we implemented a neural classifier for type checking in DTT, which is a one step less difficult problem than proof search in DTT. We also compared several different strategies for embedding DTT judgments and evaluated their efficiency.

Multimodal emotion recognition is a technology that integrates multiple modalities—such as audio, text, and images—to more comprehensively and accurately identify and analyze human emotions. In the field of AI-driven dialogue systems, it has become an indispensable technology for facilitating smooth interactions. By fusing data from different modalities, such as audio and text, it is possible to account for inter-modal interactions and correlations that are not captured in single-modal emotion analysis, thereby improving both the generalizability and accuracy of emotion recognition.In this study, we constructed a multimodal emotion analysis model based on the Transformer architecture, which takes audio and text as inputs. By concatenating the outputs of the respective Transformer encoders for audio and text and then applying a Self-Attention mechanism to the concatenated representation, our model can fuse these modalities while preserving their Cross-modal relationships. In this paper, we conduct comparative evaluation experiments against multiple existing methods on CMU-MOSEI, a standard dataset for emotion recognition tasks, to validate the performance of the proposed model and confirm the advantages of multimodal fusion for emotion recognition.

Recently, Large Language Models (LLMs) have been integrated with other systems by producing output that follows a specific format, such as JSON or a programming language. There are several methods to control the output of LLMs so that they always follow a specified grammar. However, existing methods do not take into account the limit of the number of output tokens required for practical use, which leads to ungrammatical output because the generation can be terminated in the middle. We utilize LL(1) parsers and propose a novel method of forcing LLMs to generate grammatically correct output considering the maximum number of tokens. In a benchmark of generating JSON with strict constraints on the maximum number of tokens, our method improves the accuracy by 20 points compared to an existing method, and almost all outputs follow the JSON grammar.

While biomedical data is highly accurate, the amount of data is limited, and there is a need to develop analytical methods that effectively utilize a small amount of data. In this study, we used data collected from approximately 2,500 individuals in the Fukushima vaccine cohort, Japan's largest and longest cohort for the COVID-19 vaccine. By applying an integrated approach of mathematical models and machine learning, we estimated IgG(S) antibody titer dynamics from 1 or 2 IgG(S) antibody titer data, age, and sex. This means that IgG(S) antibody titer data at any given time can be predicted from 1 or 2 blood samples. Furthermore, we researched the optimal timing of blood sampling to use this approach effectively. This approach can also be applied to speeding up clinical trials and fundamental research where data acquisition is difficult.

The analysis and prediction of Electronic Health Records (EHR) are essential components of modern healthcare systems, providing significant benefits to both providers and patients. With the increasing availability of vast amounts of patient data in EHR systems, these resources offer both challenges and opportunities for predictive modeling. Reinforcement Learning (RL) has gained popularity in EHR research, particularly due to its strength in handling sequential data, making it well-suited for time series analysis and prediction in healthcare. However, many existing approaches simplify disease progression by modeling it as a Markovian process, potentially overlooking the complex influence of a patient's medical history on disease development. In this study, we propose leveraging Large Language Models (LLMs) to encode patients' disease-treatment histories, enhancing the representation of historical data. We then employ reinforcement learning techniques to optimize treatment recommendations. Furthermore, we aim to develop an auxiliary LLM model independently trained to predict future disease occurrences, thereby improving the robustness and accuracy of our predictive framework.

In clinical drug development, complex information must be organized for decision-making. This process requires identifying key issues and developing effective strategies quickly. Defining endpoints (EPs) in clinical trial planning is crucial for assessing drug profiles. Knowledge graphs (KGs) integrate information based on relationships, enabling new inferences. KGs can improve comprehensiveness and efficiency in clinical development. Using generative AI as an interface cloud facilitate discussions among clinical researchers. This study explored EP search methods for clinical trials using KGs and generative AI. We integrated KEGG MEDICUS and ClinicalTrials.gov, implementing 1) 1-hop KG search, 2) 2-hop+ search, and 3) link prediction via KG completion algorithms. Results were summarized and reported using generative AI. A qualitative evaluation by clinical researchers confirmed that the proposed method provides comprehensive and efficient information extraction. Future challenges include expanding data sources and improving inference accuracy.

In neurodegenerative diseases such as Alzheimer’s and Parkinson’s, lesions begin in localized parts of the brain and gradually spread to adjacent regions. In order to gain a deeper understanding of this pathology and predict its progression, modeling and simulation based on partial differential equations have attracted attention. However, reliable predictions are hampered by lack of information to determine the model details such as spatial inhomogeneity of biochemical parameter values. To address this issue, we aim to derive a simulator based on human brain images, in particular spatial maps of lesions obtained by PET (positron emission tomography) technology. Specifically, we have developed a method for optimizing model parameters and initial conditions using the error back-propagation based on the discrepancy between simulated and observed lesion distributions. By applying this method to brain images from the ADNI project database, we have been able to infer the spatial distribution of biochemical parameters in the human brain, which cannot be directly observed. We will present the results of this work and discuss remaining issues of model selection and loss function design.

Many recent studies have reported high accuracy in classifying heart diseases and localizing affected regions using large-scale electrocardiogram (ECG) data with deep neural networks (DNNs). However, these methods often rely on residual networks (ResNet) and deep stacks of one-dimensional convolutions. Such architectures demand substantial computational resources. This paper employs tensor electrocardiogram (TCG) technology to extract ECG shape features. As a result, we achieve a lighter, whitebox model whose accuracy is comparable to blackbox DNN-based approaches. Experimental results show that we achieve a macro AUC of 0.933, which is close to ResNet's 0.937 and better than LSTM's 0.927.

Endometriosis affects approximately 10% of women, causing infertility and chronic pain. Definitive diagnosis requires invasive laparoscopy, leading to delayed treatment intervention. While comprehensive diagnostic methods including non-invasive imaging are gaining attention for early intervention, the global shortage of expertized radiologists remains challenging. This study aimed to develop a machine learning-based diagnostic support system to improve diagnostic accuracy and address the shortage of interpreting radiologists. We designed an integrated model system called AI-based MR imaging Support Program (AMP). This system enables the detection of deep endometriosis (DE) nodular lesions in the posterior uterus and ovarian endometriomas using nnU-Net, as well as the prediction of pelvic inter-organ adhesions using radiomics features. AMP was established through collaboration between expertized radiologists and AI experts, based on the limited number of medical images. MRI is suitable for evaluating DE, but there is a need for advanced analytical expertise and reduction of interpretation burden. AMP addresses these challenges, supporting early diagnosis of DE and appropriate treatment strategy planning through preoperative assessment. This presentation introduces the development process and data collection for establishing AMP using medical images.

The large language model (LLM) has demonstrated their powerful performance in a variety of fields. To further improve the performance of a LLM in a specific field, fine-tuning is a common method. LLM in the medical field is often fine-tuned using general medical knowledge to improve performance, but when the model is faced with a specific disease, the model responses are not completely accurate and can sometimes be completely irrelevant. In this work, we focus on a specific disease (epilepsy), fine-tuning the pre-trained model using data from the epilepsy field. The epilepsy data includes the basic knowledge of the disease, conventional treatment plans, and commonly used drugs, as well as precautions in daily life, etc. In the experiment, a variety of evaluation methods are used to compare the fine-tuned model with the pre-trained model. From the results, the performance of the fine-tuned model has been greatly improved.

In protein engineering, introducing mutations into natural proteins is a common strategy to enhance desired properties. However, comprehensive evaluation of all potential mutations through large-scale screening is prohibitively expensive. This study proposes a method that leverages small-scale mutant data obtained during initial experiments to predict the performance of unknown mutations and efficiently identify beneficial mutation sites. The method is based on the concept that protein performance fundamentally depends on its three-dimensional structure. Specifically, the proposed approach involves: (1) predicting the 3D structures of mutant proteins using AlphaFold2, followed by energy minimization; (2) applying molecular field mapping to the optimized structures to extract fixed-length vectors that retain physicochemical features; (3) constructing performance prediction models using Partial Least Squares (PLS) regression; and (4) analyzing regression coefficients to visualize key regions contributing to performance. When applied to publicly available channelrhodopsin data used in optogenetics, the method achieved high prediction accuracy and successfully identified mutation candidates that improve performance.

With the popularization of the model-informed drug development (MIDD) approach, most newly approved drugs in recent years have included mathematical models that describe the relationship between blood drug concentration/exposure and efficacy/safety in their regulatory submission documents. These documents include not only summaries of the data used to build the models but also information on the covariates influencing the outcomes. Additionally, because these models are defined as stochastic, the likelihood of efficacy or safety is indicated as an estimated probability, such as "the expected response rate following administration of x mg of the drug is y%." Most of these probabilities fall within the range of 30% to 70%. However, there is potential for more efficient treatment choices if these probabilities could be significantly increased (e.g. to over 80%).Recently, “precision medicine” has attracted significant attention for its potential to enhance the accuracy of therapeutic outcome predictions by integrating AI technology with a diverse array of data, including genetic information and life logs. In this presentation, we will show a case study that attempts to realize precision medicine by combining genetic data obtained from genome-wide association studies (GWAS) with deep learning models.

The use of direct oral anticoagulants (DOACs) in thrombosis treatment involves risks of recurrence when efficacy is insufficient and bleeding when excessive. An optimal index for treatment modulation remains undefined. This study aims to explore predictive indicators for identifying patients with low DOACs plasma levels (<50 ng/mL) using time-series changes in blood light transmission. A prediction model is constructed using exhaustive features extracted from time-series changes in transmitted light using a method developed based on tsfresh. Results are compared with those using conventional indicators from the perspective of eXplainable Artificial Intlligence. The tsfresh-based model outperforms conventional models, demonstrating higher AUCs for both ROC and PR curves. Notably, some highly predictive features exhibited low correlations with traditional indicators, suggesting that this approach introduces a novel evaluation perspective distinct from existing methods. The finding has the potential to develop new clinical indicators for accurately estimating low DOACs levels.

Comprehensive identification of target genes and accurate prediction of their therapeutic effects are of paramount importance in personalized cancer medicine. However, conventional approaches using animal models or cell lines are costly, time-consuming, and limited in their ability to achieve thorough analyses. Here we show that a new Graph Neural Network (GNN)-based learning scheme, which aligns experimental data from cancer cell lines with Virtual Knockdown manipulations in cancer patients, enables accurate prognosis prediction after gene suppression through transfer learning. Focusing on the HER2-positive subtype of breast cancer, we conducted a Virtual Knockdown of ERBB2 (Her2). Our findings not only confirmed improvements in survival rates but also suggested better hazard ratios compared to clinical data, demonstrating the biological consistency and utility of our approach for treatment efficacy prediction. This research is expected to serve as a valuable foundation for enhancing the reliability of molecular-targeted drug design, discovering new therapeutic targets, and advancing clinical applications. Moving forward, the application of this method to a broader range of cancer types and further integration with clinical data will enable the development of more versatile and precise models for predicting treatment outcomes.

Pancreatic Cancer is one of intractable cancer. The important factor in the judgment of progress in metastases and genetic mutation. The metastasis is considered how to find the sign. In this model, variational autencoder and decision tree, principal component analysis are researched using public cancer database. The causes of metastasis from the perspective of gene mutations is investigated. Pancreatic cancer metastasis is often caused by gene mutations to the liver accompanied by KRAS and TP53, and the analysis results show that the incidence rate is approximately 60% and 55% in cases with OS of less than 1 year. In addition to KRAS, the expression of TP53 and CDKN2A is involved in predicting metastasis. In addition, to extend survival time, it's suggested that the number of metastatic sites is 1 or less, or even if it is 1 or more, the age of metastasis is less than 60 years old.

In disease research, patient stratification through clustering in high-dimensional spaces represented by multidimensional clinical test data has gained attention as a novel data-driven analytical approach. This method aims to characterize the disease groups represented by each cluster. When applying nonlinear dimensionality reduction techniques, such as UMAP or t-SNE, to disease-related high-dimensional data, the resulting low-dimensional spaces can exhibit complex geometries, necessitating tailored clustering methods. One example is spectral clustering, which leverages spectral graph theory. However, when conducting significance testing on features associated with data-driven clusters, a challenge known as "selective inference" arises. Unlike classical hypothesis testing, where clustering and inference are performed independently, this scenario involves using the same dataset for both. Consequently, to accurately control the type I error rate, the influence of clustering must be incorporated into the inference process. In this presentation, we propose a method for addressing selective inference in the context of patient stratification via spectral clustering. Our approach adapts techniques developed for selective inference in hierarchical clustering. We also demonstrate practical applications of this method.

In recent years, advancements in bioinformatics have significantly improved epitope prediction. Notably, the IEDB database enables the prediction of both T-cell epitopes and linear B-cell epitopes with a certain level of accuracy using only antigen sequence data. Furthermore, tools for evaluating the toxicity and allergenicity of these epitopes have also been well-developed. Leveraging these tools, multi-epitope vaccines, which align multiple epitopes in sequence, have been developed for various antigens and have demonstrated effectiveness. In the case of influenza, numerous in silico vaccine designs have been published, and some studies have shown superior efficacy compared to the quadrivalent influenza vaccine (QIV) in mouse infection experiments. This suggests that multi-epitope design is a viable approach for effective antigen design in influenza. In this study, we focused on H5N1 influenza and performed antigen design using the in silico tools.

In the aircraft collision accident that occurred at Haneda Airport in January 2024, all passengers were safely rescued; however, pets in the cargo hold lost their lives. This incident sparked active online discussions about the handling of pets as baggage on aircraft. This study examines people’s perceptions of pets—often regarded as family members—and their owners by analyzing online comments posted on YAHOO! News articles about the incident. The study uses two analytical approaches. First, topic modeling and word co-occurrence network analysis identify themes in the comments, suggesting a focus on the inconvenience caused by owners’ actions during emergencies and the nuisance posed by dog and cat allergies in normal circumstances. Second, from the perspective of “inconvenience,” the study applies a moral foundation dictionary based on moral foundations theory in psychology, revealing how particular moral views are embedded in the comments. The findings indicate that discussions of inconvenience are closely tied to notions of rules, authority, and related concepts.

The number of social media users has increased dramatically in recent years, and it is important to clarify the impact of social media use on the users' mental health. While some studies have shown a negative impact of social media use on users' well-being, others have denied the existence of such an impact, and no consistent conclusions have been reached yet. One of the reasons for this is the lack of consideration to the purpose of social media use (social interaction, information gathering, entertaining themselves, etc.), behavior (commenting, looking at posts, posting, etc.), and factors unrelated to social media use (e.g., achieving good grades on tests). In this study, we used an experience sampling method to analyze the effects of social media use on well-being in detail. Specifically, we collected data not only on the presence of social media use, but also on the purpose and behavior of social media use. Through data collection and analysis, we aimed to clarify the conditions under which social media usage influenced well-being. The results suggested that each type of social media, purpose of use, and behavior had different effects on well-being.

In recent years, the widespread use of the internet and the rapid development of generative AI technologies have significantly transformed the information space. In particular, the digital advertising market has been experiencing rapid growth, and marketing strategies utilizing social media are increasingly being prioritized. Amid these changes, the phenomenon of echo chambers has garnered attention. An echo chamber refers to a state where individuals are exposed only to information that aligns with their existing beliefs, resulting in the amplification of such information, which can lead to social division and the reinforcement of confirmation bias. This study aims to empirically elucidate the relationship between influencers and the echo chamber phenomenon using data from the social media platform X (formerly Twitter). Specifically, the research will analyze the impact of influencer activities on the formation of echo chambers, identify the characteristics of users who are susceptible to echo chamber formation, and explore the traits of influencers that strongly influence the creation of echo chambers. Through this analysis, the study aims to provide insights into the effective utilization of influencers in digital marketing and contribute to the construction of a healthy information space.

This study aims to understand collective emotions by demographics in Japan using social media data and to identify the differences among them. The research employs the "Japanese Atmosphere Index," which is based on an emotion dictionary constructed using the Profile of Mood States 2 (POMS 2). The index is applied to data from X (formerly Twitter) provided by Hottolink Inc., investigating changes influenced by societal events. Focusing on attributes such as gender and age group, the study highlights findings like a significant difference in the Fatigue between employed and non-employed individuals during the post-Golden Week period, suggesting that societal events have varying emotional impacts across demographics. Furthermore, an analysis of emotional fluctuations by day of the week and month using the Prophet model revealed demographic-specific characteristics, such as differing trends in the Anger between men and women. By applying the methods and results of this demographic-based collective emotion analysis, the study is expected to contribute to enabling businesses and governments to better understand consumer emotional tendencies and needs, thereby facilitating the development of effective marketing strategies and policies.

As the use of social media has rapidly grown, more and more people are relying on social media even for political decision-making. It is widely known that the echo chamber effect, the phenomenon that people with similar opinions tend to connect more deeply, appears in social media, as prior research showed that two political clusters, regarded as left and right, were formed on Twitter. In this study, we illustrate that current political discourse on Twitter has become more fragmented than before, same as the real political situation in Japan. We also show when and how such new clusters were formed by creating temporal networks and tracking affiliation in the political echo chambers, suggesting that the political discourse on Twitter closely reflects the real political situation.

To address tasks where complete AI automation is unsuitable, a collaborative decision-making process that integrates human and AI expertise is needed. However, conventional approaches in which humans approve or reject AI suggestions risk overreliance on AI and reduction of accuracy. An alternative has humans make an initial prediction, then review the AI’s recommendation before finalizing decisions, yet complex real-world tasks may cause people to rely heavily on AI without sufficient initial analysis. This paper investigates the effects of presenting hypothetical decision criteria to AI prior to forming predictions, comparing changes in AI dependence and accuracy. Specifically, humans select key features they prioritize, and AI references a Rashomon set with similar predictive performance but distinct rationales, offering predictions aligned with those criteria. User study on student grade prediction tasks reveals that examining presenting decision criteria fosters more balanced AI reliance.

This study investigates perception gaps regarding generative AI between university researchers and students (N=337), a crucial issue given the increasing integration of this technology in academia. Using a modified Generative AI Attitude Scale, we compared researcher and student attitudes and usage. Results revealed three key discrepancies: a trust gap (researchers < students), a gap in willingness-to-use and perceived usefulness (researchers > students), and a perceived risk of ‘lack of originality’ gap (researchers < students). These findings suggest that researchers, while demonstrating a lack of trust in generative AI, exhibit a strong intention to utilize it, and, furthermore, assess the risk of ‘lack of originality’ as lower than students. This research contributes to the discourse on building trust in generative AI and addressing algorithm aversion, offering implications for technology policy in higher education.