In recent years, as the proportion of children with developmental disabilities in regular classrooms has increased,the demand for clinical psychologists capable of conducting psychological assessments has also grown.Training methods for psychological assessments typically include classroom-based instruction and practical exercisesor internships with real participants. However, due to difficulties in securing internship sites and assessmentsubjects, clinical psychologists often lack sufficient opportunities to receive adequate hands-on training in psychologicalassessments. To address this issue, we developed a child-like robot (hereinafter referred to as the proposed robot)to support clinical psychologists in practicing the administration of intelligence tests. The proposed robot is equippedwith speech content that mimics the responses of children undergoing intelligence testing. In this study, we developedtwo types of robots: (1) a conversation-oriented robot that allows trainees to practice intelligence tests throughinteractive dialogue, and (2) a system-equipped robot with a built-in learning system that displays test procedures andmanuals. We then conducted a user study to investigate clinical psychologists ’impressions of training with the twotypes of proposed robots. The participants were clinical psychologists who had already studied the target intelligencetest. The experimental results revealed that the conversation-oriented robot received more favorable evaluations comparedto the system-equipped robot, suggesting that voice-based training may provide a more engaging experiencefor clinical psychologists.

In this study, we propose a novel method for constructing datasets for financial sentiment analysis．In recentyears, there has been growing interest in research aimed at supporting decision-making for investors and analysts．Among such approaches, sentiment analysis―which transforms text data into quantitative values to enable statisticalanalysis―has gained particular attention．Existing studies have primarily employed methods such as coarse multilabelclassification of text polarity or summing up scores assigned to individual words．However, these approachesoften fail to consider contextual information or to capture the subtle nuances specific to financial texts. The methodproposed in this study addresses these issues by leveraging two techniques: text comparison and a rating system．By interpreting the results of sentiment comparisons between texts as pseudo-matches, we can apply existing ratingsystems．The resulting sentiment scores are continuous values, which is expected to better reflect the strength ofpolarity in financial texts. Experimental results show that the sentiment data generated using the proposed methodexhibits trends similar to those of human-annotated data．Furthermore, training BERT on the generated training datademonstrated that the data quality is high enough to serve as effective supervision.

Deep reinforcement learning (RL) is a machine learning paradigm in which agents learn policies to maximizerewards through interactions with their environment. It has achieved remarkable success in domains such as roboticcontrol, gaming, and autonomous driving. Large language models (LLMs), trained on massive collections of textdata, have brought revolutionary progress to natural language processing. LLMs capture not only the grammar andsemantics of natural language but also internalize commonsense knowledge and reasoning abilities. LLMs demonstratestrong performance in diverse tasks such as text generation and program synthesis, and their influence extendsacross a wide range of applications. To address the limitations of RL, a growing body of research has explored incorporatingthe knowledge, reasoning capabilities, and program generation skills of LLMs into RL frameworks. LLMsprovide prior knowledge that facilitates policy learning from limited interactions, enable long-term prediction and efficientexploration, reduce the cost of reward design through natural language, and generate diverse environments thatsupport generalization. These properties improve sample efficiency, support reward specification, and enhance theadaptability of training environments in RL. This paper presents a comprehensive survey of such studies, organizingthem according to the roles that LLMs play within the RL process. Specifically, three major categories are examined:LLM-based agents, which enhance decision-making; LLM-based reward design, which automates or supports theconstruction of reward functions; and LLM-based training environments, which provide environments to facilitatelearning. Based on this taxonomy, the survey analyzes contributions, challenges, and future research directions at theintersection of LLMs and RL.

The paradigm in computer vision has shifted from Convolutional Neural Networks (CNNs) to Transformers,yet recent work has revitalized CNNs by showing that modernized architectures can achieve state-of-the-art results.This study embraces this trend by improving the High-Resolution Network (HRNet), a prominent architecture fordense prediction tasks, focusing on practical performance. Previous HRNet enhancements often relied on the invertedbottleneck structure to reduce theoretical model parameters and FLOPs. However, this design is frequentlyinefficient on parallel hardware like GPUs. Its core component, depth-wise convolution, has low arithmetic intensity,which can slow down execution and increase memory consumption, creating a gap between theoretical and practicalefficiency. Our work aims to improve the performance of HRNet while maintaining its practical efficiency. We deliberatelyavoid the inverted bottleneck and instead integrate four established, computationally inexpensive techniquesinto HRNet. Our proposed modifications include: (1) the multi-path down-sampling from ResNet-D to preserve information;(2) the optimized block component layout from ResNetV2; (3) the Cross Stage Partial Network (CSPNet)for richer gradient flow; and (4) selective integration of attention mechanisms to enhance representational power efficiently.Our models were evaluated on image classification and human pose estimation. The results demonstrated thatthe optimal enhancements are task-specific: a combination of all four techniques was best for classification. Furthermore,practical efficiency evaluations on an NVIDIA A100 GPU revealed that while the inverted bottleneck structurecaused a severe drop in inference throughput and out-of-memory (OOM) errors during training, our proposed modelsuccessfully maintained high execution efficiency. Meanwhile, only the down-sampling improvement excelled at humanpose estimation. This research provides a concrete method for improving HRNet’s baseline performance withoutsacrificing its practical efficiency. It also offers valuable insights for future CNN design by presenting a successfulimprovement strategy that is not dependent on the inverted bottleneck structure.

Large Language Models exhibit impressive linguistic competence, yet this has given rise to a foundationaldebate as to whether their capacity for generalization stems from human-like systematicity or mere stochastic parroting.Central to this is the principle of compositionality. This survey clarifies the mathematical requirements forcompositionality by examining major paradigms—from n-grams to deep neural architectures—through the lens ofhomomorphisms between syntactic and semantic algebras. We demonstrate that the limitations of these modelsconverge to the curse of dimensionality, rendering the learning of compositionality an inherently ill-posed inverseproblem. To address this, we propose formalizing learning as an inverse problem in representation theory, where linguisticsymmetries act as essential regularization to constrain the hypothesis space. We further suggest that geometricmachine learning, leveraging these symmetries as inductive biases, offers a novel mechanism, potentially utilizingmathematical formulations like Clifford algebra, where the establishment of homomorphisms serves as a critical indicatorof compositional generalization. This framework redirects the research focus toward elucidating the algebraicstructure of language itself.

One of the fundamental problems in motor control theory is ill-posedness, and the principle that the agent ororganism selects one among many possible solutions to their motor task. Two theoretical approaches have been proposedto explain the mechanism that resolves this ill-posedness: One is based on optimal control theory, and the otherdynamical system theory. However, since these two approaches consider it under their own theoretical assumptions,either approach alone may not be sufficient for a broad range of realistic motor tasks that usually have characteristicsof both approaches. The present study analyzed a throwing task that allows us to clearly define the task performance(error), as required by optimal control approach. By analyzing the motor task from the developmental perspective—emphasizing the natural fact that the constraints on task, environment, and body can be dynamic and changing—,we showed mathematically and computationally that a type of ill-posedness in this motor task can be manipulated andthen resolved. We discussed potential applications of these findings to curriculum learning in artificial intelligenceresearch.

Modern AI systems, such as deep learning models, remain opaque and lack unified theoretical foundations.Category theory offers a mathematical language to integrate diverse paradigms of AI: gradient-based, probabilistic,and geometric learning, but a gap persists between abstract theory and concrete computation.We introduce Typed Petri Nets (TPNs) as a formal and operational framework bridging this gap. TPNs modelneural architectures compositionally and unify deterministic and probabilistic data flows. By linking TPNs with sheafand cosheaf theory, we formalize the relation between local computation and global consistency, and by derivingcausal nets from TPN executions, we analyze causal dependencies within AI processes. This framework advances theinterpretability and verifiability of AI systems.

Mechanistic Interpretability (MI), originating from traditional interpretability which tracks the reasons behindneural network behaviors, dedicates itself to reconstructing neural network models into human-understandablealgorithms, often expressible as pseudocode or referred to as circuits. MI typically attributes each step in the pseudocodeto specific model modules, thereby accomplishing a semantic decomposition of the neural network model.This methodology is particularly effective for large models and has thus risen in prominence with the emergenceof large language models, attracting growing attention from the perspectives of AI safety and reliability. However,this field’s rapid growth has led researchers to adopt disparate concepts and methods, resulting in a lack of a unifiedframework. Moreover, the precise meaning of “mechanistic” remains ambiguous, and the distinction between MI andexisting interpretability methods has yet to be clearly established. In this paper, we first survey the historical and culturalbackground of MI, clarify its differences from traditional interpretability approaches, and propose a conceptualframework that organizes key ideas in MI. Additionally, we discuss MI methods and their limitations, ranging fromobservational to interventional approaches. Finally, we explore current challenges in MI research and offer directionsfor future work to understand increasingly complex AI systems and ensure their safety.

In high-stakes domains such as healthcare, the reliability of Large Language Models (LLMs) is critical, particularlywhen generating clinical insights from incident reports. This study proposes a category-based few-shot exampleselection method for prompting LLMs to generate background/causal factors and preventive measures from detailsof the medical incidents. For our experiments, we use the Japanese Medical Incident Dataset (JMID), a structureddataset of 3,884 real-world medical accident and near-miss reports. These reports are variably annotated with a widerange of tags―some include descriptive information (e.g.,“ medications, ”“ blood transfusion therapy ”). We alsomake this dataset publicly available to support future research. We compare a zero-shot baseline and three few-shotexample selection strategies―random sampling, cosine similarity-based selection, and our proposed category-basedmethod―using GPT-4o and Llama 3.3. Results show that the category-based approach achieves the highest precisionand most stable generation behavior, while similarity-based selection often leads to unintended outputs and safety filteractivation. These findings suggest that selecting examples based on human-interpretable dataset tags can improvegeneration precision and stability in clinical LLM applications.

This paper provides a survey on memorization of training data of large language models (LLMs), summarizingaround 180 studies across natural language processing, security, and other research fields. With the rapid proliferationof LLMs, concerns about memorization of training data have become increasingly critical. The issues involved areextensive, encompassing security, copyright, and sound model evaluation. Recent studies have sought to defineand categorize memorization, to quantify its prevalence across different model scales, and to explore mitigationstrategies such as deduplication, unlearning, and knowledge editing. These efforts have revealed that memorizationis not merely a byproduct of overfitting but a nuanced phenomenon dependent on string duplication, model size, andcontext length. Looking forward, promising directions include (i) developing principled frameworks to balance utilityand safety by distinguishing harmful memorization from legitimate knowledge retention, (ii) establishing unifiedtaxonomies and benchmarks to evaluate memorization across domains and languages, and (iii) extending the analysisof memorization into multimodal models, where text, image, and audio data interact in complex ways. Through thissurvey, we synthesize the fragmented literature along the axes of training sets, models, and outputs and highlightemerging research avenues that will shape the responsible development of LLMs.

This paper surveys social simulation research using Large Language Models (LLMs), arguing that their contributionsshould be evaluated through two distinct objectives: (1) building a simulation as a faithful replication ofa specific real-world system, and (2) building a simulation as a theoretical model of mechanisms that recur acrossmultiple systems. Social simulation has historically pursued these aims separately: constructing detailed models ofparticular systems (e.g., organizations, cities) to test interventions, and developing abstract toy models to identifygeneral principles (e.g., polarization, segregation). The “human-like” capabilities of LLM-based agents hold promisefor both directions, yet their value and challenges differ fundamentally. For the purpose of faithfully replicating specificsystems, LLMs enable agents to embody complex individual attributes and engage in natural language-basedinteractions difficult to model traditionally. The central challenge is validation and calibration against real-world datato yield trustworthy insights for decision-makers. For the purpose of working as theoretical models, LLMs expandthe scope of theoretical modeling by internalizing linguistic interactions (e.g., negotiation, persuasion, rule modification)within agents, making phenomena such as opinion propagation, polarization, and norm formation amenableto formal analysis. A longer-term direction within this objective may be “high-dimensional science”, where detailedmulti-system simulations combined with large-scale predictive models and interpretability techniques may extractgeneral structures in a data-driven manner. Key challenges on the generality path include clarifying the theoreticalstatus of LLM agents and developing methods to extract interpretable principles from complex simulations. Overall,this survey argues that explicitly separating these two classical objectives enables the field to define appropriateevaluation criteria and research agendas for each.