Image recognition models based on machine learning can exhibit systematic error tendencies under specific conditions. For instance, if a model consistently fails to correctly recognize a truck in images under the condition of a ``white truck in broad daylight," this can be considered a systematic error. Such errors can lead to unexpected behavior for users, making it crucial to detect systematic errors in advance through testing. Existing research has proposed methods that efficiently detect systematic errors by assisting human adaptive exploration. However, for machine learning models that require continuous and extensive testing, these methods pose significant human cost issues. In this study, we propose a method called AdaSniper, which utilizes large language models (LLMs) to automatically perform adaptive exploration and efficiently detect systematic errors. The proposed method enables adaptive exploration by providing the LLM with information on misclassified target classes, indicating which class the model incorrectly recognized. Evaluation experiments demonstrated that AdaSniper could perform adaptive exploration and efficiently detect systematic errors compared to baseline methods.

During the process of software development, it is very important to monitor the status of the project and identify the potential risks. Researchers have proposed several approaches to disclose the risks of software projects, but they are either based on black-box models which are hard to explain or require lots of manual efforts. In this research, we adopt a machine-learning-based approach to predict the success/failure of a software project based on the previous development data. In contrast to those black-box approaches, our approach can output the importance of the features which explains the reason of the prediction. Firstly, we build a machine learning model and train this model with previous development data of software projects. Secondly, we feed the data of the project under development to this trained model and predict the success/- failure of this project. Finally, the reasons of the prediction are displayed to the users. We implemented this approach and evaluated it with a dataset of 11,954 real world software projects. The evaluation reached a recall of 80.1% and precision of 53.7%, which shows the feasibility of this machine-learning-based approach.

To measure software size, IFPUG method is the most major approach as function point (FP) analysis. IFPUG method needs longer time to measure software size. Therefore, simplified measurement methods such as Simple FP have been proposed. When such simplified methods are used, the measurement results are different from IFPUG method. This paper analyzed influence of the measurement methods to accuracy of effort estimation. Additionally, we clarified how to suppress the influence.

This study aims to improve the efficiency and accuracy of software development using LLMs by realizing the knowledge conversion of development materials through establishing traceability. In this presentation, we introduce the application of contradiction detection and correction technique between artifacts as one of the use cases utilizing traceability.

GraphQL is an API query language, and because it has a different structure from REST APIs, it is difficult to apply conventional automated testing methods, so a new approach is required. This research proposes an automated testing method for GraphQL APIs using reinforcement learning. In the proposed method, the test space is explored using Q-learning. A request is generated by selecting API fields and arguments based on the schema, and the Q-value is updated according to the response. By repeating this process and learning, efficient black-box testing is achieved. In the experiment, the effectiveness of the proposed method was verified using schema coverage and the rate of error responses as evaluation indicators for publicly available APIs. In the future, we plan to improve the Q-value initialization and reward design to avoid local optimum solutions, and further confirm the effectiveness through comparison with other methods.

As the number of open-source software (OSS) grows, users need better systems to quickly find tools that meet their needs. In this study, we propose a method to improve OSS search accuracy by expanding search keywords using large language models (LLMs). This method generates synonyms, related words, and translations for keywords extracted from user queries. These expanded keywords are then used for searching, providing more accurate results.

To help envisioning of program behavior, there is a material of computer science unplugged that combines a ball and pipes to express program. In this paper, we performed preliminary experiment to evaluate the effect of the expression for the program comprehension.

When users are accustomed to gamification, it is difficult to acquire effect of the gamification. The goal of the study is to obtain effect of the gamification sustainably. To achieve the goal, we preliminary analyzed used and usage interval of gamification.

In software requirements specifications, inconsistencies in the level of abstraction and expression may lead to a decrease in the accuracy of source code retrieval. This study proposes a method to improve the accuracy of source code retrieval using embedding model and vector similarity by employing large language model (LLM) as a preprocessing step to generate source code descriptions with a unified level of abstraction.

This document, based on existing deep neural network testing proposals, aims to more comprehensively test deep neural networks. It proposes a new neuron coverage criterion and a test case generation method based on this criterion.