A conventional approach to solving combinatorial problems can be divided into three steps: formulating the problem, implementing the codes and improving algorithms to get solutions smarter. Using large language models (LLM), we might solve the problems without the formulation and the implementation. In this paper, we investigated solving performance when using LLM as a solver for the traveling salesman problem (TSP). The method for providing prompts is based on OPRO [Yang 2023], where the LLM generates new solutions from the prompt which contains previously generated solutions, iteratively. Not only natural language representations of the problem, but also directed graphs representation are utilized as prompts. Approximation ratio, which is the ratio of a minimal distance of the TSP to distance of obtained solution from LLM is investigated as a solution performance. We found that LLM, Gemini-1.5-flash, can generates solutions with approximation ratio 1.55 of TSP called “att48”.

In this study, we propose a novel image generation technique called “Visual Concept Blending”, which extracts features that are either shared or distinct features from multiple reference images and transfers them to a base image. By using multiple reference images, our method provides fine control over which features are blended into the base image. At the same time, it leverages the CLIP embedding space to facilitate the transfer of higher-level concepts such as shape transformation and motion. The proposed approach directly utilizes existing pre-trained models without requiring additional training, making it both simple and robust. This technique is expected to find wide-ranging applications in fields like art and design.

Reinforced concrete, characterized by the placement of steel reinforcement inside the concrete, is one of the most widely used structural forms in building construction. The necessity of maintaining old buildings has risen and hence, studies have explored potential of Sub-Terahertz waves, a type of electromagnetic waves that fall between the radio wave and the light wave frequencies for non-destructive testing. This paper proposed a method that applies Deep Learning in estimation of the cover thickness, defined as the distance from the embedded steel plate to the concrete surface, using Sub-Terahertz. It also focused on expanding the measurable range of the Sub-Terahertz waves. The results showed a recall of over 80% on average with cover thickness of 10mm to 40mm. Furthermore, an analysis using SHAP revealed that the interpretation of the results varied depending on the frequency and cover thickness used for the measurements.

In clinical laboratories, accurate and timely identification of microorganisms is essential. To support the traditional microbial identification, we developed an automated identification system using colony images based on machine learning. For the dataset, we cultured 418 strains of 10 microbial species on agar media. Using image processing techniques, we automatically extracted 10,048 colony images from 418 plate images. The dataset was created to perform deep learning using three models of the ResNet with 18, 50, and 101 layers. The deep learning models were evaluated using validation datasets comprising 75 strains. The accuracy of microbial classification reached 92.0% (69/75 strains) with the ResNet-50 (50-layer) model. This result indicates that this model has potential to support the microorganism identification in clinical laboratories.

Deep learning has achieved significant success in modeling dynamical systems with unknown governing equations. However, existing models tend to treat systems as monolithic and indivisible entities, making it challenging to accurately model coupled systems. Furthermore, they are often inapplicable to domains outside mechanical systems, such as electrical circuits and hydraulic systems. To address these limitations, we propose Poisson-Dirac Neural Networks (PoDiNNs), which are based on the Dirac structure that unifies the port-Hamiltonian and Poisson formulations. The proposed approach enables a unified representation of various systems spanning multiple domains, as well as the interactions and degeneracies arising from the components that constitute these systems. Experimental results demonstrate that the proposed method effectively learns the interactions between components and achieves superior long-term prediction performance compared to existing methods.

In real-world deployment, machine learning models often experience concept drift, resulting in degraded predictive performance. Existing dynamic ensembles use a single set of weights, limiting their ability to address localized drifts in the feature space. We propose a novel Two-Layer Conditional Dynamic Ensemble (TL-CDE) that partitions the feature space into subregions defined by multiple conditions via a two-stage process, assigning distinct weights to each subregion. By leveraging a single pretrained model and its modulated variants, TL-CDE maintains performance without retraining while also providing detailed drift insights through weight update logs. Experiments on synthetic and real-world datasets demonstrated that TL-CDE outperforms existing methods.

In recent years, many companies have recognized the importance of human resources and are investing in human capital to revitalize organizations and enhance internal communication to foster innovation. However, conventional quantification methods have mainly focused on readily measurable indicators without addressing the fundamental role of conversations in human capital. This study focuses on routine meetings and proposes methods to visualize human capital by examining speech behavior during these meetings. We use a conversation visualization technology we have developed—which operates effectively even under noisy conditions—to quantify speech. We then measure differences in speech volume by attributes such as gender and job title, changes in speech volume depending on whether certain participants are present, and correlations between speech volume and continuous attributes. To verify the effectiveness of our proposed method, we analyzed speech volume by gender and departmental affiliation during weekly meetings at a small to medium enterprise.

Technical analysis is widely used by many investors. One of them is chart pattern analysis, which uses geometric features of charts. While various explanations exist for the mechanism underlying pattern formation, this study focuses on the concept of self-fulfilling prophecy, which has been mentioned in several behavioral finance studies, and tests a new hypothesis that these patterns emerge from trading activities of many investors expecting pattern completion. Analysis of the head-and-shoulders pattern, which is one of the most recognized chart patterns, in S\&P500, BTCUSD, and EUR/USD markets reveals that (1) the frequency of pattern occurrence consistently exceeds that of random walks generated from similar price distributions, and (2) trading volume demonstrates an increasing trend prior to pattern formation across multiple time series. These findings support our hypothesis, and indicate that trading activities conducted before pattern completion may have influenced price movements in a way that facilitates pattern formation. However, this analysis alone does not establish a clear causal relationship between volume fluctuations and price movements. For future research, it is desirable to adopt an agent-based approach with artificial markets.

Similarity among companies is important information that forms the basis of analysis in various financial practices, such as corporate valuation, investment and loan decisions, portfolio risk management, partner selection for business promotion, and in-house investor relations activities. A useful tool for calculating the degree of similarity between companies is the embedded representation of companies, which can be obtained by using BERT and other methods on textual information. While this embedding representation based on text data is effective, in the economic and financial fields, there are many numerical data that are expected to be useful for measuring the degree of similarity between companies. It is expected that combining these numerical financial data with textual data will enable us to search for more useful “similar companies”. Therefore, this study proposes a method for searching similar companies utilizing both “textual information” and “numerical information.” Our methso utilizes not only textual information on stocks, but also numerical information such as sales by segment, stock price time-series data, and shareholder composition.

This study extends prior research on using large language models (LLMs) to uncover return-predictive sentiment in Japanese 10-K reports by focusing on highly liquid stocks. Building on a dataset of Tokyo Stock Exchange-listed firms (2014–2023) and previously established methodologies, we narrow our scope to the TOPIX 100 and TOPIX 500—indices composed of the largest Japanese companies by market capitalization. Despite expectations that these well-followed and actively traded stocks should incorporate public information more efficiently, LLM-derived sentiment still predicts future returns, with larger abnormal returns (alpha) than when all listed stocks are included. These findings highlight the robustness of LLM-based approaches in detecting subtle signals within corporate disclosures and challenge the notion that highly liquid markets fully reflect available information. By highlighting the predictive power of LLM-extracted sentiment in large-cap portfolios, this study offers practical insights into how advanced natural language processing can enhance investment strategies, even in supposedly efficient segments of the equity market.

In recent years, cross-disciplinary data collaboration has become increasingly prevalent, making data an essential driver of innovation. Consequently, the demand for data trading markets has grown. However, these markets remain in their early stages, with incomplete regulatory frameworks and unclear guidelines governing participant behavior. Additionally, many data trading markets are inundated with low-quality or difficult-to-use data, posing challenges to their effective utilization. This study focuses on data providers within these markets and investigates the types of providers and strategies that can maximize both their own profits and the overall market’s profitability and trading volume. Experimental results demonstrate that data providers who adopt a long-term perspective achieve higher individual profits while also enhancing the market’s overall profitability and trading activity.

Some investors say increasing investors with the same strategy decreasing their profits per an investor. On the other hand, some investors using technical analysis used to use same strategy and parameters with other investors, and say that it is better. Those argues are conflicted each other because one argues using with same strategy decreases profits but another argues it increase profits. However, those arguments have not been investigated yet. In this study, we added the model increasing fundamental or technical agents with exactly same parameters to the previous agent-based mode. We found that in the case with increasing fundamental agents, decreasing variation of market prices and their profits. In the case with increasing technical agents, increasing variation of market prices and their profits.

This paper newly constructs and analyzes the artificial market simulation using a large language model (LLM) agent.First, we constructed an LLM-based agent as a wise trader agent who consider fundamentals and trends, and a noise trader as an agent who trades on other exogenous factors.We then tested whether we could reproduce the stylized facts by changing the prompts and number of the LLM agents.The results showed that the prompts of the LLM agent are very important, and suggests that it is possible to reproduce the stylized facts by giving appropriate prompts as long as we tested.We also confirmed the current limitation of this approach.

Understanding the relationships between players through network structures plays an important role in analyzing modern social and economic systems. For example, in shareholding networks, it is essential to quantitatively evaluate how each player influences others to analyze power dynamics and influence structures. Although the Shapley-Shubik power index is a well-known measure used to evaluate each player’s influence in games where voting power is distributed among players, this index do not explicitly consider network structures, and they fail to fully capture the relationships and influence propagation between players. Previous studies have extended the Shapley-Shubik power index to networks, introducing the Network Power Index (NPI) to represent the power exerted by one company i over another company j in shareholding networks. In this study, we give a graph-theoretic interpretation and description of the idea and mechanism of NPI and discuss it from a purely mathematical approach.

We analyze the relationship between the disclosure of director skill matrices, corporate financial characteristics, and ESG scores in Japanese companies. Skill matrices have gained attention as a tool to enhance board functions. However, their creation is left to individual companies, leading to issues such as a lack of objectivity and standardization. Furthermore, we find that little empirical research has been conducted on the information value disclosed through these matrices. In this study, we classified director skill matrices by extracting skills from publicly available nomination statements for directors. We created a dataset, defined the skills, and formulated the classification as a multi-label problem. To address label imbalance, we developed a learning framework and trained a model to classify the skill matrices. We then analyze the relationship between these estimated skill matrices and corporate financial performance and ESG scores using data from TOPIX100 companies. Our analysis shows that directors with financial, business management, and international experience are positively associated with ROE and governance scores.

This paper proposes Utility Efficient Collaborative Filtering (UECF), a stock recommendation method based on utility-aware matrix factorization. Traditional stock recommendation methods, which rely on collaborative filtering and mean-variance optimization, aim to recommend stocks by considering investors' preferences and balancing risk and return. However, these methods face challenges in adequately reflecting investors' risk-return preferences in portfolio recommendations. Moreover, the performance of mean-variance optimization heavily depends on estimated parameters, such as expected returns, which can reduce reliability under uncertainty. Our UECF enables the recommendation of stocks that align with investors' preferences while providing high utility. By incorporating higher-order moments and asymmetric correlation structures, our approach more accurately captures investors' preferences in stock recommendations. Experiments using real-world data demonstrate that UECF achieves high recommendation performance while effectively considering Sharpe ratios and utility in its recommendations.

Deep learning-based multi-factor models have been employed to predict cross-sectional stock returns by incorporating numerous stock-level characteristics and capturing their nonlinear relationships. However, the inherent complexity of deep learning models often makes them difficult to interpret, posing challenges in practical applications where explainability is critical. To address these challenges, we focus on Neural Additive Models (NAM) – a recently proposed deep learning model designed with high interpretability – and investigate its applicability to cross-sectional stock return prediction. Although NAM’s characteristic subnetwork architecture is useful to obtain highly interpretable outputs analogous to factor returns and exposures in linear factor models, we argue that naive training procedure may lead to unstable predictions. To overcome this problem, we propose a modified NAM architecture that incorporates a novel regularization term, resulting in a framework well-suited for cross-sectional stock return prediction. Through numerical simulations, we demonstrate that the proposed method improves interpretability without sacrificing predictive accuracy compared to the conventional NAM.

The estimation of average treatment effects (ATEs), defined as the difference in expected outcomes between treatment and control groups, is a central topic in causal inference. This study develops semiparametric efficient methods for ATE estimation in a setting where only a treatment group and an unknown group, comprising units for which it is unclear whether they received treatment or control, are observable. This scenario is a variant of learning from positive and unlabeled data (PU learning) in weakly supervised learning and can be viewed as a special case of ATE estimation with missing data. For this setting, we derive semiparametric efficiency bounds, which are lower bounds of the asymptotic variance for regular estimators. We then propose semiparametric efficient ATE estimators that achieve these efficiency bounds in terms of their asymptotic variance. Our findings make significant contributions to causal inference with missing data and weakly supervised learning.

In recent years, large language models (LLMs) have become increasingly pervasive in society. Specifically, interactions have begun to emerge in which LLMs induce changes in human behavior, and the data produced by these transformed behaviors are subsequently used to further train the LLMs. In such circumstances, the impact of LLMs’ social behaviors on society is becoming impossible to ignore. To investigate these social behaviors, previous research conducted donation games with multiple LLM agents and suggested that cooperative behavior evolved in Claude, while it did not in Gemini and GPT-4o. In this study, we examine the process by which strategic diversity and cooperative behavior become fixed in donation games. Specifically, we evaluated strategy diversity by measuring the similarity between strategy texts and assessed the temporal evolution of strategic actions. The results indicate that while Claude ’s strategic repertoire converged toward cooperative strategies and exhibited reduced diversity over successive generations, Gemini ’s strategy group tended to maintain its diversity. These findings are expected to inform our understanding of the formation of cooperation and the design of rules in a hybrid society of humans and LLM agents.

Learning in zero-sum games studies a situation where multiple agents competitively learn their strategy. In such multi-agent learning, we often see that the strategies cycle around their optimum, i.e., Nash equilibrium. When a game periodically varies (called a "periodic" game), however, the Nash equilibrium moves generically. How learning dynamics behave in such periodic games is of interest but still unclear. Interestingly, we discover that the behavior is highly dependent on the relationship between the two speeds at which the game changes and at which players learn. We observe that when these two speeds synchronize, the learning dynamics diverge, and their time-average does not converge. Otherwise, the learning dynamics draw complicated cycles, but their time-average converges. Under some assumptions introduced for the dynamical systems analysis, we prove that this behavior occurs. Furthermore, our experiments observe this behavior even if these assumptions are removed. This study discovers a novel phenomenon, i.e., synchronization, and gains insight widely applicable to learning in periodic games.

Large Language Models (LLMs) are playing an increasingly influential role in human decision-making. While previous research has primarily considered LLMs as tools that assist human decision-making, less attention has been given to their role in shaping a co-creative decision-making process, in which humans and LLMs iteratively update their distributions through interactions. This study presents a theory and model of LLM-human interaction in co-creative decision-making, formulated within the framework of distributed Bayesian inference, where the iterative process can be interpreted as a sampling-importance-resampling (SIR) algorithm. To examine the validity of the model, we conduct two experiments: (1) a cooperative card-guessing task, analyzing how variations in agent interaction dynamics affect decision convergence, and (2) an iterative brainstorming task, exploring its applicability to broader decision-making contexts. This study aims to establish a theoretical foundation for adaptive and democratic LLM-human decision-making.

This study aims to clarify how the "Reiwa Rice Riot" spread and subsided through X and news media, focusing on changes in emotions. To analyze the characteristics of information diffusion and resolution, we used three indicators: the transition of the number of posts, emotions, and polarity strength. X exhibited highly immediate reactions, with fear and expectation intersecting at the earliest stage. Additionally, news articles showed a temporary surge in polarity before reaching their peak in number, followed by a decline. These results suggest that different media have distinct timing in information transmission and emotional expression.

Recommendation systems play an important role in a variety of fields by suggesting items based on user preferences. In recent years, multi-objective recommendations that balance accuracy with diversity and novelty have attracted attention, but they are particularly difficult to apply with cold-start users. This study proposes a foundational approach leveraging the preferences of similar existing users to provide multi-objective recommendations to cold-start users. We evaluated popular recommendation models SVD, LightGCN, and NCF, by examining their capability to capture diversity and novelty. Results indicate that SVD and LightGCN generally capture diversity and novelty effectively. In contrast, NCF consistently struggles to capture diversity and novelty at the embedding layer. These findings support our hypothesis that users with similar embeddings in the latent space also have similar diversity and novelty values, offering a pathway to improve cold-start recommendations through multi-objective frameworks.

We optimize inventory by leveraging machine learning techniques to the planning of sales, production, and inventory management. Our primary focus is on assessing the uncertainty associated with sales predictions, which directly impacts safety stock decisions across various inventory strategies. Conventional methods for uncertainty estimation often rely on state space models, widely used in time-series forecasting; however, these models have limitations regarding symmetric distribution assumptions and reduced data efficiency. In contrast, Sequential Predictive Conformal Inference (SPCI) addresses these challenges by non-parametrically estimating residual. We experimentally confirm that SPCI effectively lowers stock levels while minimizing the risk of stockouts.

Artifacts that perform a function, such as automobiles and watches, will fail as they are used. Therefore, designers of artifacts need to infer the effects that a fault of a target part will have on other parts and on the entire artifact. The authors have been studying fault inference, which infers possible faults of an artifact from a model of its function and faults. In a previous study, we constructed a model of the automobile's powertrain and confirmed the effectiveness of fault inference. In this study, we apply fault inference to automotive lighting devices and proceed with modeling and formulation of inference methods. Specifically, we create a knowledge model using functional decomposition trees and ontologies for an automotive lighting system as an example, and compare it with previous study to analyze how to define the concepts necessary for fault inference. In this paper, we show that we have created and analyzed a knowledge model using ontology and obtained knowledge about how to define concepts necessary for constructing an ontology on fault inference and about how to define concepts related to possible faults in each part.

This article focuses on the importance and challenges of extracting, sharing, and utilizing explicit, latent, and tacit field knowledge (called “Gen-Ba knowledge”) in industries such as manufacturing, healthcare, and agriculture. While generative AI has improved access to explicit knowledge, handling tacit and latent knowledge remains challenging. To address this, we have proposed the concept of a “Digital Knowledge Twin,” in which Gen-Ba knowledge is extracted as knowledge fragments including voices, images, and sensor data and it is shared among members through internalization workshops. This article discusses this concept and key technologies needed for its implementation.

Secondhand trading apps promote reuse by enabling individuals to sell unwanted items and purchase usable or collectible goods. Unlike conventional e-commerce sites dominated by professional sellers, these apps connect individual buyers and sellers, making platform support crucial. This paper examines advancements in AI, including large language models that offer contextual understanding, natural language processing, and emotion analysis. Such technologies can automatically generate product descriptions, propose negotiation strategies, and facilitate effective communication, reducing transaction-related anxiety. We present case studies from the “Value Exchange Engineering” project—a collaboration between Mercari and The University of Tokyo—focusing on two aspects: presenting desired communication styles for negotiation support and developing “EmoBalloon,” a system that visually conveys emotional arousal in text chats using speech balloons. Our findings highlight AI’s potential to empower user interactions in online marketplaces.

This study aims to explore the application of a human-centered AI ecosystem in the production of medical information content. Although medical communication involves diverse stakeholders, considerable time is dedicated to regulatory adherence and quality assurance, thereby reducing opportunities for medical writers to concentrate on value creation. To address this issue, we developed a document-checking and suggestion system employing rule-based artificial intelligence and large-scale language models (LLMs). Consequently, the system organized documents in PDF and Word formats and efficiently checked each meaningful unit. Furthermore, the system includes functionality to facilitate risk assessment by incorporating custom rules that integrate the tacit knowledge of medical writers. This study examines the importance of implementing a human-centered AI ecosystem by posing the question: ‘How can AI extend beyond efficiency to foster the development of advanced competencies among medical writers?’

Many golfers face challenges in improving their swing skills and often choose self-study instead of taking golf school lessons. However, self-study can lead to the adoption of incorrect techniques or swing forms, which may eventually develop into bad habits. To address this issue, we have developed a golf swing diagnosis system that is more reliable than self-study and more convenient than lessons. This system combines the expertise of golf professionals with the latest computer vision technology. It was released as a smartphone application in November 2024, and we are currently considering further improvements to enhance the user experience based on feedback from users.

Large Language Models (LLMs) possess the ability to perform well on unknown tasks and flexibly alter their behavior according to prompts. Leveraging this characteristic, there are attempts to assign virtual personas or personalities to LLMs and make them behave accordingly. If we could intentionally limit LLM performance, the constructed virtual personas would likely become more realistic (e.g., making a kindergartener unable to solve integral calculus). This paper addresses such intentional performance degradation of LLMs. Using multiple Japanese benchmark tasks, we report that it is difficult to degrade LLM performance in downstream tasks through prompts alone. We also examine the benchmarks necessary for measuring performance degradation.

In AI-driven decision support systems, explanations are essential for users to assess the system's suggestions. With the advent of large language models, it has become significantly easier to tailor the way explanations are expressed. However, the impact of these expressions on human decision-making remains largely unexplored. This study investigates how explanation tones, such as formal or humorous, influence decision-making, focusing on the roles of AI and user attributes. We conducted user experiments using three scenarios based on distinct AI roles: assistant, second-opinion provider, and expert. The results revealed that in the second-opinion scenario, explanation tone had a significant impact on decision-making regardless of user attributes. In contrast, in the assistant and expert scenarios, the influence of tone varied depending on user attributes. Older users were found to be more susceptible to tone influences, while highly extroverted users tended to exhibit discrepancies between their perceptions and decisions. This study offers valuable insights for designing effective explanation styles in AI systems.

In order to achieve the basic commercial objective of selling products, it is considered important that customers feel comfortable shopping at a store.Providing comfort is expected to increase customers' willingness to buy, which in turn will lead to higher sales. Higher ceilings are expected to have a positive impact on users, but it is difficult to actually change the ceiling height. Therefore, we verified that using MR to simulate the perception of ceiling height has the same effect as actually changing the height of the ceiling.In this study, an HMD (head-mounted display) is used.Preliminary experiments compared three types of ceilings: "normal ceiling," "monochromatic: blue ceiling," and "blue sky ceiling," and suggested that the "blue sky ceiling" enhanced the feeling of openness and comfort the most.On the other hand, "monochromatic: blue ceiling" tended to give a feeling of discomfort and oppression.An additional experiment at this university co-op confirmed the possibility that "blue ceiling" not only increases attention to the ceiling, but also positively influences purchasing behavior.Conversely, "monochromatic: blue ceiling" decreased appetite, suggesting that color may influence purchasing behavior.

The rise of financial technologies has transformed access to financial services, introducing diverse products such as microloans, personal loans, and insurance. While this progress enhances financial inclusion, it also presents a significant challenge: accurately assessing credit risk for individuals with limited credit histories and businesses managing complex data. To address this issue, we propose a novel transfer learning framework for credit risk prediction using minimal data. Our approach involves converting sparse tabular customer data into image format and fine-tuning pre-trained image classification models, including VGG-16, ResNet-50, and GoogleNet. Additionally, we explore training these models from scratch on source domain data, followed by fine-tuning with target domain data. Rigorous testing on public and specialized datasets demonstrates the robustness of our method in handling data scarcity. Our framework consistently outperforms benchmark transfer learning algorithms in prediction accuracy, showcasing its potential to bridge the gap in financial risk assessment and promote broader financial inclusion. This study highlights the innovative application of transfer learning in addressing critical challenges in the financial sector, offering a scalable and effective solution for credit risk evaluation in data-constrained environments.

We collected data from Video Research Ltd.'s "Creative Carte", a large-scale CM survey dataset (3,387 CMs, ~1.8 million responses), and conducted path analysis to quantitatively analyze the impact of TV commercial (CM) favorability on consumer purchase behavior using a purchase funnel. The analysis examined the influence of CM favorability and persuasiveness on each stage of the purchase funnel (awareness, interest, and purchase intention). Results showed that CM favorability had a remarkably high overall effect (0.701) on purchase intention. Text mining was performed on viewers' free-form responses, utilizing a large language model and principal component analysis, suggested the importance of factors contributing to "intuitive appeal," such as visual elements, in effective CMs. Conversely, the analysis indicated that factors such as insufficient or unclear expressions could potentially detract from CM effectiveness. This research provides empirical evidence of CM favorability's significant impact on consumer purchase behavior, particularly through factors such as intuitive appeal and expression quality. These findings offer practical insights for advertisers aiming to enhance CM performance.

Predicting the next shopping basket of a customer is a crusial task for retailers. Among many algorithms developed, personalized large language models (LLMs) attract much attention because of their zero-shot ability. In this paper, we apply LLM-based next-purchase prediction to grocery shopping data. Compared to datasets used in existing works, repeated purchases ocuur frequently and item names are explanatory in grocery shopping data. Numerical experiments show that our LLM-based method can take advantage of these properties; It outputs higher prediction scores for more frequently perchased items and can recognize items similar to those previously purchased using item names. On the other hand, it suffers from the "Lost in the Middle" phenomena in the purchase history and cannot capture recency. It also fails to utilize item popularities. These lead to the lower accuracy than conventional machine learning approaches.