The use of generative AI as a writing assistant or substitute is increasing, particularly in individual writing. As AI blurs the boundary between human and machine roles, writing remains an inherently human creative act that cannot be simply replaced. This study explores individual creativity in writing, emphasizing ”dialogue with one ’ s past self” as a key element in fostering creativity. Based on this perspective, we developed HALO-TextEditor, a prototype that enables interaction with past writing history. By dynamically visualizing a writer ’s past work in chronological order, it helps users reposition, interpret, and repurpose their previous creative outputs within their current context. Drawing from insights gained with HALO-TextEditor, this paper redefines the human role in the era of generative AI and explores the essence of creativity emerging from engagement with one ’s writing history.

With advancements in deep learning, image generation has been widely explored in fashion such as virtual try-on and attribute editing. However, understanding brand-specific clothing features through deep learning models remains underexplored. It is necessary to capture designers' creative intentions and brand-specific features. In this study, we explored image generation which reflects brand-specific features to understand clothing features for deep learning. With the support of a fashion brand specializing in Lolita fashion, we utilized Stable Diffusion to generate high-quality images. We trained LoRA models on clothing images with backgrounds and facial features removed, using prompts that reflect brand-specific attributes such as color, shape, and impression. Furthermore, we explored merging multiple LoRA models to create hybrid clothing designs that blend different elements. We developed a web application based on the proposed system and evaluated its effectiveness as a fashion design support tool by having actual design-related users test it.

In general research projects, the scope of application is explored in a forecasting manner based on the technologies and phenomena developed during the research project, and the research is then carried out. However, in cases where projects in multiple fields are being carried out simultaneously to create synergy, the backcasting method may be effective. In this study, we examined the benefits of science fiction prototyping by creating a new vision of the future through science fiction prototyping with science fiction writers in the Cybernetic Avatar Project of the Moonshot Research and Development Program.

With the rapid development of large-scale language models, automatic generation of short stories is already reaching a practical stage, when those are limited to specific genres or styles with high typicality. In addition, it is becoming possible to generate plot level general stories that cross genres or combine genres. On the other hand, although there have been various attempts to automatically generate stories with complex structures such as long works, it is difficult to say that this has been fully realized. In this study, existing works were first decomposed into short patterns ofshort plots, and quantified the characteristics of the complex structure of stories by converting the combination structures of these patterns, such as nesting, succession, and parallelism, into data. In addition, a hierarchical narrative structure generation system was created using the extracted structural features and plot patterns in existing works. After generating a base story, the system further subdivides each plot, lengthens the story hierarchically, and verbalizes it using a large-scale language model. The constructed system can also be used to lengthen existing story works based on their structures. In the future, an evaluation experiment of the validity of the output results and the controllability of diversity are plannned.

In addressing societal challenges through data utilization, a key perspective is the generation of narratives that effectively connect data analysis to practical actions. This paper proposes Multi-Dimensional Sugoroku (MDSugo6), a framework that visualizes and controls the data utilization process using the analogy of traditional games. The framework operates within a representation space that comprehensively addresses various narrative elements across multiple dimensions, ranging from abstract worldviews to concrete raw data expressions. Through case studies of both one-way and two-way movement patterns in this space, we demonstrate that the two-way pattern better captures the dynamic nature of narrative generation, as it encompasses both descriptive explanations and pattern-based understanding. As a result, MDSugo6 contributes to establishing a systematic method for effective data utilization in various practical contexts. Additionally, we specify requirements for creative support tools that can promote this dynamic process, particularly focusing on the visualization and operation of transitions within the representation space.

In recent years, advances in generative AI have enabled the streamlined creation of stories and illustrations. With the advent of large language models (LLMs), human-level natural language generation has become feasible, accelerating applications in education and entertainment, including children's content. This study proposes an automated picture-book creation system using LLMs for text generation and Stable Diffusion for illustrations. We first define a scenario structure and employ multiple LLMs to produce coherent text. In addition, LoRA-based training is applied to Stable Diffusion to ensure consistent character appearances throughout the book. Preliminary results indicate the successful generation of a four-part story alongside corresponding images.

This study aims to clarify the differences between popular and general works in the emotional characteristics of stories and how they change throughout story progression. Text data were collected from the "Shousetsuka ni Narou" website, with the top 300 works based on global points categorized as popular works and 300 randomly selected works as general works. BERT was then used to conduct sentence-level sentiment analysis, and the texts were divided into fixed-length segments, using the average sentiment score as a feature. Based on these features, a classification model was constructed using Random Forest to distinguish between popular and general works. Finally, SHAP analysis was conducted to reveal the emotional tendencies of the works. The analysis results indicated that the emotional balance of a story significantly influences readers’ emotional engagement and satisfaction. In particular, during the early stages, emotions such as sadness and anticipation related to the main character’s problem-solving play a crucial role in drawing readers into the story. Towards the end, popular works tend to maintain emotions related to anticipation and trust as the story progresses, while emotions such as surprise and fear gradually decrease.

Conversational AI has improved significantly with the rise of Large Language Model(LLM). As research on character design grows, conversation with story creator's character offers new insights into a story’s deeper layers and cultural background, while fostering emotional involvement. However, current conversational systems still face issues: for example, a Meiji-era author might provide details on modern devices or give contradictory statements about the story world, reducing engagement. I propose design guidelines—supported by conversational log examples—that ensure consistent representation of characters and settings while enhancing user experience. These guidelines can be applied in areas such as entertainment, education, and creative support. Based on the guidlines, I developed a web application and the user survey showed a high level of satisfaction. By a narrative-focused view, I aim to elevate AI beyond a simple tool, making it a source of new storytelling experiences and cultural value.

In asset management, asset allocation and selection are often based on the relationships of asset returns. Therefore, analyzing and interpreting news and events that influence these relationships is crucial. This study proposes a method to detect and analyze significant news and events that impact asset relationships. It is said that when a major news event occurs that has a significant impact on an asset, the asset price may jump, leading to discontinuities. Consequently, we propose a method to detect price jumps in assets and to create and analyze a graph by connecting assets that experience jumps in the same direction with edges, and we conducted validation of this method.

We introduce Plug-FIN (Pseudo-Labeling Using Generative Modeling Approaches to Find Investment Strategies), a novel framework that leverages large language models to enhance quantitative investment strategies in three key ways. First, the framework assigns pseudo-labels to substantial volumes of text data. Second, it transforms these labels into interpretable long/short trading strategies. Finally, it incorporates a feedback system that systematically identifies top-performing strategies and integrates newly developed trading strategies as well as custom labeling functions. This iterative process ensures the models remain current as market conditions evolve. We conduct experiments with real stock market data and demonstrate that Plug-FIN not only improves predictive accuracy but also produces interpretable and profitable strategies.

In the financial market, information on social events and people's reactions are important factors that have a large impact on the market. In this study, we will use mass media and social media to examine a method for improving profits in option trading using the Nikkei 225 VI stock index.

We focus on the critical task of accurately estimating volatility in financial markets, which is essential for risk management and portfolio optimization. Since volatility cannot be directly observed, we typically rely on theoretical models such as GARCH and Stochastic Volatility (SV) models. However, SV models are known for their nonlinearity and high dimensionality, which require computationally intensive methods like MCMC or particle filters. These methods, however, often face challenges related to computational efficiency and convergence. In this study, we propose a new approach to volatility estimation using a SV model with a Stein Particle Filter (SPF). By leveraging interactions between particles, we address the limitations of traditional particle filters based on importance sampling. Specifically, we adopt a gradient-based update rule using a Radial Basis Function kernel, enabling particles to efficiently converge to the true posterior distribution. Through numerical experiments with a regime-switching SV model, we demonstrate that SPF outperforms conventional SIR filters in both accuracy and convergence speed.

The increasing adoption of Retrieval-Augmented Generation (RAG) for long-context question answering in organizations has heightened the demand for human verification of outputs from large language models (LLMs). To address this, researchers propose instructing LLMs to provide sentence-level quotations as evidence alongside their responses to facilitate the verification process. However, LLMs occasionally generate quotation errors, which complicate verification efforts. Our research introduces a novel method integrating automated correction processes with the self-correction capabilities of LLMs to address quotation errors. Our experiments demonstrate that the proposed method effectively corrects quotation errors and generates higher-quality quotations with fewer tokens than existing methods. This work highlights the utility of post-correction methods, like ours, for addressing sentence-level quotation errors generated by LLMs.

As large language models (LLMs) gain increasing attention, concerns have also been raised about stereotypical outputs and underlying social biases. While extensive research has been conducted on English-based LLMs, studies on Japanese models remain limited. This study examines the safety of Japanese LLMs in responding to stereotype-triggering prompts. We constructed 3,612 prompts by combining 301 social groups with 12 stereotype-inducing templates in Japanese and conducted three tasks using models trained on Japanese, English, and Chinese. Our findings show that LLM-jp had the lowest refusal rate and was more likely to generate toxic and negative responses compared to other models. Additionally, prompt format significantly influenced all models, and the generated responses included exaggerated reactions toward specific social groups, varying across models. These results highlight the need to improve safety mechanisms in Japanese LLMs and contribute to discussions on bias mitigation and their safe and responsible deployment.

The phenomenon of the rapid spread of bad news is widely recognized in human society, but it is unclear whether the same phenomenon occurs in a community of LLM agents. In this study, we examined whether bad news spreads more easily among LLM agents, and confirmed that bad news tends to propagate less easily than good news. We also proposed a human-mimetic agent to resolve the difference in information propagation characteristics between humans and LLM agents, and showed results that contribute to improving the accuracy of simulation studies.

Identifying which actions performed by an agent are feasible or infeasible for an AI is valuable knowledge when considering or designing guidelines for determining which aspects of an agent's tasks can be delegated to an AI agent. This paper primarily draws on the Third Restatement of Agency in the United States to analyze various actions performed by agents, examining whether they can be carried out by a tool AI that operates strictly according to the principal's instructions, or by an autonomous AI that acts independently while adhering to the principal's directives.

Large Language Models (LLMs) continue to improve in their conversational performance, and their ability to replicate human behavior is advancing. With the increase in the number of tokens available for in-context learning in LLMs, it has become easier to configure LLMs to act as experts. However, clear procedures for replicating the intellectual activities of highly skilled experts using LLMs have not yet been established. This study aims to establish procedures for configuring expert LLMs by constructing an LLM that replicates the behavior of a law professor. System prompts were created based on the professor's own written works and feedback from acquaintances, and the behavior of the LLM incorporating these prompts was compared to the professor's actual responses. The LLM reproduced the professor’s answers from the General Social Survey with approximately 60% accuracy, while the replication of personality traits remained at a lower level, showing little correlation with the traits reported by the law professor. On the other hand, the LLM’s personality showed a higher correlation with those obtained from the professor’s acquaintances' ratings. Considering the gap between self-assessment and external evaluation of personality, as well as the biases present in the LLM's responses, we discuss methods for future performance improvements.

In recent years, with the rapid development of science and technology, the skills required of engineers have become increasingly diverse and complex. In this context, skill network analysis is attracting attention as a method for systematically analyzing the skills of workers. In this study, we constructed a skill network using career data for Japanese engineers and analyzed the patterns of core skills and distinctive skills possessed by engineers. As a result, it was found that while there are general skills that are commonly used by a wide range of engineers, specialized skills play an important role as core skills in specific fields such as data science and security.

Democracy is connected to property rights as fundamental human rights. A key aspect of democracy is ensuring that taxpayers understand how their taxes are utilized and reflecting to governance. In Japan, documents evaluating government projects conducted by administrative bodies are publicly available. At the national, these evaluations are known as administrative reviews, while at the local, they are referred to as administrative evaluations. These documents cover a wide range of government projects and include various indicators and numerical data, making them difficult to interpret without expert knowledge. This study focuses on ChatGPT, a type of AI chatbot, to examine the capability to summarize and analyze such documents, aiming to improve residents' understanding of government projects. Specifically, rather than having AI evaluate the projects directly, this study sought to assist residents in making informed judgments by organizing quantitative indicators and summarizing the content to clarify the objectives of local government evaluations.

Machine learning and data mining from graph structured data are studied intensively. Using interval representations for allocation of resource and time is useful in problem solving of their overlap and shortage. An interval graph is a graph that represents an interval representation. We propose a method for acquiring characteristic interval graph patterns from positive and negative interval graph data, by using evolutionary learning.

Japan Society for Intellectual Production (JSIP) has evolved over more than 20 years and has made extensive contributions to the field. While existing work has relied mainly on quantitative statistical methods for context analysis, recent advances in large language models (LLMs) have made more possible methods such as context learning and analysis. In this study, a retrieval system is constructed, which aims to perform similarity retrieval analysis on JSIP papers according to predefined categories. It uses LLMs for text vectorization and improves the retrieval results through similarity filtering, while effectively analyzing the topic trends. The experimental results verify the effectiveness of the proposed approach, in which "Industry-Academia-Government Collaboration Policy" appears with the highest frequency and " collaboration between different fields" with the lowest frequency, which will help to expand the content analysis of industry-academia collaboration activities in the future.

Under the "2024 Problem," stricter overtime regulations for drivers have exacerbated labor shortages in the logistics industry. While recruiting novice drivers is a critical countermeasure, their lack of knowledge about optimal parking locations poses efficiency challenges. In contrast, experienced drivers possess valuable knowledge in selecting optimal parking locations, which can be leveraged to support novices and improve logistics efficiency. This study proposes a parking recommendation algorithm based on veteran driver knowledge. The algorithm utilizes a database combining delivery destination addresses and parking locations and comprises three main components: (1) selecting reliable parking locations from past parking data with a focus on proximity; (2) identifying nearby parking locations for unregistered buildings; and (3) If multiple delivery destinations can be consolidated into a single parking location, clustering-based aggregation processing is applied. The proposed method was validated through subjective evaluations by drivers and proof-of-concept experiments, demonstrating its high practicality and accuracy in effectively recommending parking locations to novice drivers.

Contract review is a significant burden for legal professionals, and various AI tools are being developed to address this. Recent studies have reported that large language models (LLMs) demonstrate higher accuracy than junior lawyers and can significantly reduce costs. However, when using AI to assist humans, the interpretability of outputs generated by LLMs becomes a challenging issue.This paper proposes a pattern matching method aimed at improving interpretability by predefining risk-related expressions using an ontology and matching them with problematic parts of input sentences to determine risks. Experimental results show that while the proposed method falls short by about 10 points in risk determination accuracy compared to a prompt-based classification method, it is capable of making risk determinations with a certain level of accuracy.In conclusion, we demonstrate that a hybrid approach is promising, where the initial risk determination is achieved with high accuracy using the prompt-based classification method, and the interpretability is supported by the proposed method through verification against predefined criteria for the determined risks.

The design of individual houses typically involves the responsible designer examining appropriate design elements based on the residents' family composition, lifestyle, and sensibilities, and materializing these elements through dialogue with the residents. In particular, it is crucial to identify the residents' sensibilities and incorporate them into the design. However, such design processes require advanced communication skills, a deep understanding of the residents, and the ability to translate these insights into concrete designs, which can be challenging for less experienced designers.To address this, we developed a housing design support system that proposes specific design elements based on information about the residents and facilitates dialogue with them. This system incorporates innovations such as process decomposition suitable for individual house design and allows for revisions to the design content through chat-based interactions. Additionally, identifying the residents' sensibilities and reflecting them in the design are key features of this system.Evaluation by practitioners confirmed the potential of this system to support dialogue and design with residents. This paper outlines the overall structure of the system, the reasoning algorithms for sensibilities and design elements, and the evaluation results obtained from the prototype system.

In the development of thermal management products for electric vehicles, a significant amount of effort is dedicated to thermal fluid analysis for temperature control performance evaluation. To reduce this effort, high-speed inference using surrogate models that train from past design shapes is useful. To improve the inference accuracy of surrogate models, a wide-ranging and dense dataset is required. However, components such as cooling pipelines have complex layout, which makes it difficult to prepare various shape data within realistic effort constraints. Therefore, in this study, we developed an automatic generation technology for piping layout diagrams based on path-planning methods. By training on the generated shape data, we obtained a surrogate model with applicable range and inference accuracy that contributes to practical design utilization.

Efficient route control in automated material handling systems, particularly Overhead Hoist Transport (OHT) systems, is a critical challenge in semiconductor manufacturing facilities. As production line conditions constantly change and unexpected congestion occurs, conventional shortest path algorithms based on mathematical optimization struggle to effectively avoid traffic congestion. To address this challenge, this research proposes a route control method called "Alpha-Q-Routing." Unlike conventional Q-learning-based methods that only utilized edge-level partial simulation results for Q-value updates, this method combines partial simulation results for multiple candidate routes with estimated travel times from Q-tables, enabling route selection based on more accurate arrival time predictions. Numerical experiments in an environment with 50 vehicles demonstrated that the proposed method achieved the highest number of completed deliveries compared to conventional methods, particularly proving its superiority in environments with high vehicle density. This achievement contributes to realizing efficient logistics management in dynamically changing manufacturing environments.

This paper first introduces Multi-Agent Pickup and Delivery with immediate tasks as MAPD-I. While MAPD for the automated warehouses suppose that all delivery tasks are given at planning phases, there are several situations where the system needs to accept immediate and emergent tasks in real-time in practical applications. Subsequently, we propose an efficiency improvement method using deep reinforcement learning for the acceptance of immediate tasks in automated warehouses. Typical interrupt handling methods prioritize immediate tasks for rapid processing but do not consider the delays caused to regular tasks. In this study, we aim to improve overall efficiency by moderately procrast the handling of interrupt tasks. We propose a deep reinforcement learning model, ProcrastiNet, which determines the appropriate degree of procrastination, and compare its performance with typical interrupt handling methods and rule-based approach.

In recent years, the expansion of the e-commerce market has driven up delivery demand, while stricter labor regulations and other factors have led to a serious shortage of drivers, with predictions that transport capacity may fall short by about 34% by 2030. To address this challenge, this study proposes a new delivery planning method that incorporates the “segment-based units” actually considered by drivers. The proposed method creates a two-stage plan—first at the segment level, where parcels are grouped, and then at the parking-lot level—and uses a genetic algorithm to optimize delivery sequences, including those subject to specific time constraints. Additionally, it accommodates a “docking” process, where parcels that cannot be loaded all at once are brought partway by another vehicle and then replenished en route, making the design closer to real-world operations. A case study using actual data showed that, compared to existing methods, the proposed approach reduces unnecessary movement between segments while shortening total delivery time by approximately 1%. Moreover, even when the number of time-constrained deliveries increases, the system can flexibly accommodate them without delays. Future work will focus on identifying specific parcels causing bottlenecks and further improving efficiency through collaboration with demand-side stakeholders.

Geographic information is important in various industries. However, in many cases, it is superficial such as land structures, population, and events which are tied to uniform units like points, sites, and administrative divisions. Therefore, it does not reflect actual human behaviors especially at detail scales. We propose a method for discovering important geographic areas in appropriate shapes where some human activity is frequently observed by defining weight for the activity. The weight is defined for each point of time-series coordinates in human flow data. The weights are aggregated to small areas to evaluate the importance of them, and adjacent and similar important areas are integrated. The weights can reflect various activities by using not only the coordinate points but also human attributes and other geographic information. In addition, by using these information, features of the areas are defined to evaluate the similarities between areas and to analyze themselves. Further knowledge discovery can be expected by repeating this process of discovery, analysis, and redefinition of the weights. Examples of discovered areas by using this method are also shown.

In recent years, various big data have been acquired, and their use has begun in many fields. However, because the amount of data is large, the number of people who can handle it is limited, and it requires a lot of effort, so it is sometimes difficult to extract enough information. There are various methods to extract information from big data, but generative AI, which has recently made remarkable progress, can also extract information from big data, and it is thought that many users will be able to extract information from big data by combining it with large-scale language models. We are conducting research on obtaining information on roads and roadside facilities from images taken with in-vehicle cameras etc., and this time, we used ChatGPT-4o to determine the number of floors of a building from images taken with a 360° camera, and used the latest LLM technology to improve the accuracy of readings. In addition, in order to understand the impact that differences in the read information have on road obstruction due to building collapses in an earthquake, evacuation simulations were conducted to understand the impact on evacuation times, etc.

In this study, we consider incorporating time-series forecasting into applications. We use a foundation model for time-series forecasting, and perform forecasting only by prior learning, without any fine tuning. In addition, this study assumes the existence of correlation and periodicity in the behavior of application users. It is difficult to identify whether the outliers that occur in a time-series forecast are specific values that depend on the date (referred to as "events") or user-specific outliers. We propose a method to identify event patterns and outliers without prior information on event dates and without modifying the model itself to maintain the generality of the foundation model.

We aim to realize smooth traffic flow to solve social issues in transportation, such as traffic congestion and accidents. One of the methods is traffic flow simulation. By simulating various traffic conditions and their countermeasures, we believe decision-making for traffic measures that ensure smooth traffic flow can be conducted more efficiently. Appropriate parameterization of the model is important to accurately reproduce traffic conditions in simulations, but appropriate parameters are rarely known in advance. Therefore, we are researching parameter estimation through data assimilation using traffic probe data. In this study, we employ a data assimilation method called the Ensemble Kalman Filter, performing sequential data assimilation along with the time evolution of traffic flow simulations. In data assimilation, system noise is considered to express the uncertainty of the modeled system, and we found that the variance of system noise significantly affects the accuracy and stability of parameter estimation. This study confirmed that varying the system noise variance with time evolution, which is generally kept constant, can improve the accuracy of parameter estimation.

In the introduction of autonomous driving in urban areas, control systems are required to prioritize human lives during emergencies, which differ from normal driving conditions. Control during "emergencies" is typically triggered based on quantified metrics like Time Headway (THW) or Time to Collision (TTC), which reflect the driver's perception of danger while driving. However, in environments where pedestrians, motorcycles, and other vehicles coexist, acquiring appropriate control rules that ensure safety for all entities is extremely challenging. Traditional control theory-based approaches and image-based control methods have inherent limitations in such complex scenarios. Furthermore, in addition to avoiding emergencies, it is necessary to consider the transition back to normal conditions to minimize secondary damage.This study introduces deep reinforcement learning (DRL) for emergencies defined by TTC. It proposes a method that dynamically switches between obstacle avoidance control and recovery control to return to the vehicle's original lane after avoidance, depending on the situation. Experiments conducted in a simulated urban environment demonstrate that the proposed method improves destination arrival accuracy while preventing traffic accidents and maintaining safety.