Research about long text generation has been actively conducted with the advancement of large language models. However, generating long text that considers the unique philosophy within a small community, such as speeches used in graduation ceremonies or company inductions, remains challenging. The reason is that information and literature about prominent leaders in small communities are generally extremely limited compared to those about well-known prominent leaders. This study focuses on prominent leaders within small communities, such as university or company founders. It aims to generate speech scenarios that automatically share small communities’ unique philosophies. In this paper, we target Hachisaburo Hirao, the founder of our university, and extract sentences containing his philosophies from his diaries, lectures, and autobiographies to create a quotations database. We then propose a method to generate speech scenarios based on the user’s input theme using Retrieval-Augmented Generation (RAG) with the quotations database. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.

Recent advances in large language models (LLMs) have enabled the replication of speech patterns and philosophies of prominent historical figures. However, generating dialogue that reflects the philosophies of prominent leaders in small communities, such as founders of local universities or small businesses, remains a challenge due to the limited availability of public data. Nevertheless, the philosophies of such leaders often serve as important educational and behavioral foundations for members of these communities. In this study, we propose a dialogue generation method that enables the sharing of a prominent local leader’s philosophy through natural conversation. Specifically, we classify sentences left behind by the leader—such as those in books or diaries—into four types: statements, thoughts, actions, and facts. We then perform two-stage fine-tuning using the statements and thoughts to generate dialogues that faithfully reflect the leader’s values and philosophy. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.

We propose a novel white-box membership inference attack (MIA) for large language models (LLMs) that leverages internal parameter dynamics to determine whether a given text sample was included in a model’s pre-training data. Prior MIA approaches rely primarily on input-output behavior and struggle to distinguish memorized samples from semantically similar but unseen ones due to the probabilistic nature of LLM generation. To address this challenge, we introduce a method based on parameter drift—defined as the Euclidean distance between the entire set of model parameters before and after continual pre-training on a single input. Our hypothesis is that continual pre-trained inputs induce minimal parameter changes, while unseen inputs require greater updates to the model. Notably, we find that even semantically similar inputs yield distinct drift magnitudes, enabling more precise membership inference. We validate our approach on multiple LLMs, including Pythia and LLaMA-2, and show that it consistently outperforms existing MIA baselines such as Min-K% Prob and SaMIA*zlib under various evaluation settings. Furthermore, we demonstrate that focusing on parameters with high drift further improves inference accuracy, achieving state-of-the-art results on benchmark datasets. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.

During disasters, a large volume of messages are posted on social networking services (SNS). Some of these messages contain behavioral facilitation information, which either encourages or discourages specific actions. However, the interpretation of such information depends on the personality traits of the individuals affected. In this study, we hypothesize that victims’ personality traits influence their perception of behavioral facilitation information, and we analyze the characteristics of these differences. Focusing on typhoons, we propose a method for extracting behavioral facilitation information from posts on X (formerly Twitter) during typhoon-related disasters. The extracted information is then classified into four content-based categories: suggest, inhibition, encouragement, and wish. Furthermore, we categorize individual personality traits into five dimensions (the Big Five), and also take into account their age and sex. We then analyze how the perception of each type of behavioral facilitation information varies according to these traits. Our analysis reveals that, during disasters, the interpretation of behavioral facilitation information exhibits distinct and consistent patterns depending on the personality traits of the victims. © 2026, Slovene Society Informatika. All rights reserved.

On crowdsourcing platforms, the quality of collected data depends on the clarity of instructions, but requesters struggle to create instructions that capture their own implicit criteria. To address this issue, we propose a novel framework that uses two Large Language Models (LLMs) – a Creator and an Evaluator – to automatically explore the space of possible instructions. In this iterative process, the Creator LLM generates diverse instruction candidates, and the Evaluator LLM, acting as a proxy for human workers, assesses their performance on a task, providing a fitness score. Our experiments show that this exploratory approach is effective for discovering high-quality instructions, even if the process does not show monotonic improvement. Using the best-performing instruction created by our method with gemma3, we achieved 5.4% higher accuracy and 0.035 lower RMSE than when gemma used an instruction created by a requester. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.

During a disaster, a large amount of Behavioral Facilitation information (BF information) is posted on SNSs. We have proposed the method of extracting and categorizing BF information into four labels (behavioral axis). In this paper, we analyze the differences in how BF information is received before, during, and after a typhoon, focusing on the attributes of disaster victims. This research aims to clarify the appropriate information for the target victims in each typhoon period and analyze their relationship. The results have shown differences in the relationship between the victims’ attributes and their perception of BF information. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.

We find adequate additional layers of auxiliary tasks for BERT-based multi-task learning. Multi-task learning is a method for improving the accuracy of machine learning model by adding auxiliary tasks. Previous studies propose multi-task learning models with auxiliary tasks added to different layers. Our research question is which layer is effective for adding auxiliary tasks to improve accuracy because the answer is still unknown. The aim of this study is to find adequate additional layer of auxiliary tasks that maximizes model accuracy. We use a BERT-base model consisting of twelve layers of Transformer and experiment with seven datasets. Our experimental results show that changing the additional layer of auxiliary tasks improves macro-F1 by up to 5.1% (p-value=0.019). Moreover, our findings suggest that the insertion of auxiliary tasks into layers with the main task’s characteristics increases accuracy. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.