arXiv subjects

Sadao Kurohashi

Publications and source records attributed to Sadao Kurohashi.

At least 19 recordsLinked to original sources

arXiv2026

ShapleyLaw: A Game-Theoretic Approach to Multilingual Scaling Laws

In multilingual pretraining, the test loss of a pretrained model is heavily influenced by the proportion of each language in the pretraining data, namely the \textit{language mixture ratios}. Multilingual scaling laws can predict the test loss under different language mixture ratios and can therefore be used to estimate the optimal ratios. However, the current approaches to multilingual scaling laws do not measure the \textit{cross-lingual transfer} effect, resulting in suboptimal mixture ratios. In this paper, we consider multilingual pretraining as a cooperative game in which each language acts as a player that jointly contributes to pretraining, gaining the resulting reduction in test loss as the payoff. Consequently, from the perspective of cooperative game theory, we quantify the cross-lingual transfer from each language by its contribution in the game, and propose a game-theoretic multilingual scaling law called \textit{ShapleyLaw}. Our experiments show that ShapleyLaw outperforms baseline methods in model performance prediction and language mixture optimization.

cs.CL↗

arXiv2026

Reasoning Depth and Environment Complexity: A Controlled Study of RLVR Data Allocation across Logical Reasoning Tasks

Reinforcement learning with verifiable rewards (RLVR) has become central to post-training reasoning models, yet a key limitation of existing studies is their narrow view of the reasoning space: difficulty is treated as reasoning depth alone, and reward is concentrated on forward deductive state tracking. We instead characterize the reasoning space along two dimensions. Difficulty. Beyond reasoning depth, we study environment complexity, where models must identify the correct path amid distractors and interacting structures. Rewarded reasoning form. We consider four abilities core to real-world reasoning: deductive state tracking, abductive recovery of hidden events or facts, inductive rule induction, and analogical transfer. To disentangle these factors, we construct a synthetic knowledge-graph environment with controlled pre- and post-training distributions, where each instance varies along depth, complexity, and task family. Three findings emerge: joint depth-complexity coverage outperforms single-axis recipes; reasoning families respond non-uniformly, with abductive reasoning degrading outside the RL-covered region and task correlations clustering into deductive-abductive and inductive-analogy pairs; and uniform mixing outperforms staged curricula under a fixed budget. We also find that recent off-the-shelf models exhibit the same deductive-over-abductive asymmetry, suggesting that this gap is not merely an artifact of our controlled setup.

cs.CL↗

arXiv2026

Mechanistic Diagnostics of Spatial Lexical Bias in Multimodal Large Language Model Spatial Reasoning

Multimodal large language models (MLLMs) remain unreliable on spatial multiple-choice questions, and their failures are often attributed to poorly attended visual information. We identify a complementary failure mode, spatial lexical bias: a spatial relation word added to the answer options can act as a lexical-semantic distractor that draws the model's decision toward that option. Using nine open-weight MLLMs, we show that this phenomenon is widespread. We then isolate diagnostic cases in which a model answers a binary spatial question correctly yet consistently chooses a newly added third spatial option, which we call binary-stable but ternary-fragile cases. Leveraging mechanistic interpretability tools on these cases, we find that the failure arises on the language side rather than the visual side: visual attention analyses and residual-stream probes show the correct spatial relation remains internally available, while irrelevant-option controls, activation patching, and sparse component interventions trace the bias to specific LLM-side channels and neurons. Accordingly, we show that a lightweight LLM-only DPO update on tiny single-object-pair synthetic data mitigates the bias, lifting four-way robust accuracy by up to 100 points on synthetic data, and by 68.0, 32.6, and 20.1 points on broader evaluation datasets WhatsUp, SpatialMQA-Direct, and VSR.

cs.CL↗

arXiv2026

Memorization, Emergence, and Explaining Reversal Failures: A Controlled Study of Relational Semantics in LLMs

Autoregressive LLMs perform well on relational tasks that require linking entities via relational words (e.g., father/son, friend), but it is unclear whether they learn the logical semantics of such relations (e.g., symmetry and inversion logic) and, if so, whether reversal-type failures arise from missing relational semantics or left-to-right order bias. We propose a controlled Knowledge Graph-based synthetic framework that generates text from symmetric/inverse triples, train GPT-style autoregressive models from scratch, and evaluate memorization, logical inference, and in-context generalization to unseen entities to address these questions. We find a sharp phase transition in which relational semantics emerge with sufficient logic-bearing supervision, even in shallow (2-3 layer) models, and that successful generalization aligns with stable intermediate-layer signals. Finally, order-matched forward/reverse tests and a diffusion baseline indicate that reversal failures are primarily driven by autoregressive order bias rather than deficient inversion semantics.

cs.CL↗

arXiv2026

BIS Reasoning 1.0: The First Large-Scale Japanese Benchmark for Belief-Inconsistent Syllogistic Reasoning

We present BIS Reasoning 1.0, the first large-scale Japanese dataset of syllogistic reasoning problems explicitly designed to evaluate belief-inconsistent reasoning in large language models (LLMs). Unlike prior resources such as NeuBAROCO and JFLD, which emphasize general or belief-aligned logic, BIS Reasoning 1.0 systematically introduces logically valid yet belief-inconsistent syllogisms to expose belief bias, the tendency to accept believable conclusions irrespective of validity. We benchmark a representative suite of cutting-edge models, including OpenAI GPT-5 variants, GPT-4o, Qwen, and prominent Japanese LLMs, under a uniform, zero-shot protocol. Reasoning-centric models achieve near-perfect accuracy on BIS Reasoning 1.0 (e.g., Qwen3-32B $\approx$99% and GPT-5-mini up to $\approx$99.7%), while GPT-4o attains around 80%. Earlier Japanese-specialized models underperform, often well below 60%, whereas the latest llm-jp-3.1-13b-instruct4 markedly improves to the mid-80% range. These results indicate that robustness to belief-inconsistent inputs is driven more by explicit reasoning optimization than by language specialization or scale alone. Our analysis further shows that even top-tier systems falter when logical validity conflicts with intuitive or factual beliefs, and that performance is sensitive to prompt design and inference-time reasoning effort. We discuss implications for safety-critical domains, including law, healthcare, and scientific literature, where strict logical fidelity must override intuitive belief to ensure reliability.

cs.CL↗

arXiv2026

A Joint Neural Baseline for Concept, Assertion, and Relation Extraction from Clinical Text

Clinical information extraction (e.g., 2010 i2b2/VA challenge) usually presents tasks of concept recognition, assertion classification, and relation extraction. Jointly modeling the multi-stage tasks in the clinical domain is an underexplored topic. The existing independent task setting (reference inputs given in each stage) makes the joint models not directly comparable to the existing pipeline work. To address these issues, we define a joint task setting and propose a novel end-to-end system to jointly optimize three-stage tasks. We empirically investigate the joint evaluation of our proposal and the pipeline baseline with various embedding techniques: word, contextual, and in-domain contextual embeddings. The proposed joint system substantially outperforms the pipeline baseline by +0.3, +1.4, +3.1 for the concept, assertion, and relation F1. This work bridges joint approaches and clinical information extraction. The proposed approach could serve as a strong joint baseline for future research. The code is publicly available.

cs.CL↗

arXiv2026

Language Lives in Sparse Dimensions: Toward Interpretable and Efficient Multilingual Control for Large Language Models

Large language models exhibit strong multilingual capabilities despite limited exposure to non-English data. Prior studies show that English-centric large language models map multilingual content into English-aligned representations at intermediate layers and then project them back into target-language token spaces in the final layer. From this observation, we hypothesize that this cross-lingual transition is governed by a small and sparse set of dimensions, which occur at consistent indices across the intermediate to final layers. Building on this insight, we introduce a simple, training-free method to identify and manipulate these dimensions, requiring only as few as 50 sentences of either parallel or monolingual data. Experiments on a multilingual generation control task reveal the interpretability of these dimensions, demonstrating that the interventions in these dimensions can switch the output language while preserving semantic content, and that it surpasses the performance of prior neuron-based approaches at a substantially lower cost.

cs.CL↗

arXiv2025

SpeechIQ: Speech-Agentic Intelligence Quotient Across Cognitive Levels in Voice Understanding by Large Language Models

We introduce Speech-based Intelligence Quotient (SIQ) as a new form of human cognition-inspired evaluation pipeline for voice understanding large language models, LLM Voice, designed to assess their voice understanding ability. Moving beyond popular voice understanding metrics such as word error rate (WER), SIQ examines LLM Voice across three cognitive levels motivated by Bloom's Taxonomy: (1) Remembering (i.e., WER for verbatim accuracy); (2) Understanding (i.e., similarity of LLM's interpretations); and (3) Application (i.e., QA accuracy for simulating downstream tasks). We demonstrate that SIQ not only quantifies voice understanding abilities but also provides unified comparisons between cascaded methods (e.g., ASR LLM) and end-to-end models, identifies annotation errors in existing benchmarks, and detects hallucinations in LLM Voice. Our framework represents a first-of-its-kind intelligence examination that bridges cognitive principles with voice-oriented benchmarks, while exposing overlooked challenges in multi-modal training. Our code and data will be open source to encourage future studies.

cs.CL↗

arXiv2025

Assessing Agentic Large Language Models in Multilingual National Bias

Large Language Models have garnered significant attention for their capabilities in multilingual natural language processing, while studies on risks associated with cross biases are limited to immediate context preferences. Cross-language disparities in reasoning-based recommendations remain largely unexplored, with a lack of even descriptive analysis. This study is the first to address this gap. We test LLM's applicability and capability in providing personalized advice across three key scenarios: university applications, travel, and relocation. We investigate multilingual bias in state-of-the-art LLMs by analyzing their responses to decision-making tasks across multiple languages. We quantify bias in model-generated scores and assess the impact of demographic factors and reasoning strategies (e.g., Chain-of-Thought prompting) on bias patterns. Our findings reveal that local language bias is prevalent across different tasks, with GPT-4 and Sonnet reducing bias for English-speaking countries compared to GPT-3.5 but failing to achieve robust multilingual alignment, highlighting broader implications for multilingual AI agents and applications such as education. \footnote{Code available at: https://github.com/yiyunya/assess_agentic_national_bias

cs.CL↗

arXiv2025

Causal Tree Extraction from Medical Case Reports: A Novel Task for Experts-like Text Comprehension

Extracting causal relationships from a medical case report is essential for comprehending the case, particularly its diagnostic process. Since the diagnostic process is regarded as a bottom-up inference, causal relationships in cases naturally form a multi-layered tree structure. The existing tasks, such as medical relation extraction, are insufficient for capturing the causal relationships of an entire case, as they treat all relations equally without considering the hierarchical structure inherent in the diagnostic process. Thus, we propose a novel task, Causal Tree Extraction (CTE), which receives a case report and generates a causal tree with the primary disease as the root, providing an intuitive understanding of a case's diagnostic process. Subsequently, we construct a Japanese case report CTE dataset, J-Casemap, propose a generation-based CTE method that outperforms the baseline by 20.2 points in the human evaluation, and introduce evaluation metrics that reflect clinician preferences. Further experiments also show that J-Casemap enhances the performance of solving other medical tasks, such as question answering.

cs.CL↗

arXiv2024

LLM-jp: A Cross-organizational Project for the Research and Development of Fully Open Japanese LLMs

This paper introduces LLM-jp, a cross-organizational project for the research and development of Japanese large language models (LLMs). LLM-jp aims to develop open-source and strong Japanese LLMs, and as of this writing, more than 1,500 participants from academia and industry are working together for this purpose. This paper presents the background of the establishment of LLM-jp, summaries of its activities, and technical reports on the LLMs developed by LLM-jp. For the latest activities, visit https://llm-jp.nii.ac.jp/en/.

cs.CL↗

arXiv2024

Reformulating Domain Adaptation of Large Language Models as Adapt-Retrieve-Revise: A Case Study on Chinese Legal Domain

While large language models (LLMs) like GPT-4 have recently demonstrated astonishing zero-shot capabilities in general domain tasks, they often generate content with hallucinations in specific domains such as Chinese law, hindering their application in these areas. This is typically due to the absence of training data that encompasses such a specific domain, preventing GPT-4 from acquiring in-domain knowledge. A pressing challenge is that it's not plausible to continue training LLMs of such scale on in-domain data. This paper introduces a simple and effective domain adaptation framework for GPT-4 by reformulating generation as an \textbf{adapt-retrieve-revise} process. The initial step is to \textbf{adapt} an affordable 7B LLM to the target domain by continuing learning on in-domain data. When solving a task, we leverage the adapted LLM to generate a draft answer given a task query. Then, the draft answer will be used to \textbf{retrieve} supporting evidence candidates from an external in-domain knowledge base. Finally, the draft answer and retrieved evidence are concatenated into a whole prompt to let GPT-4 assess the evidence and \textbf{revise} the draft answer to generate the final answer. Our proposal combines the advantages of the efficiency of adapting a smaller 7B model with the evidence-assessing capability of GPT-4 and effectively prevents GPT-4 from generating hallucinatory content. In the zero-shot setting of four Chinese legal tasks, our method improves accuracy by 33.3\% compared to the direct generation by GPT-4. When compared to two stronger retrieval-based baselines, our method outperforms them by 15.4\% and 23.9\%. Our code will be released

cs.CL↗

arXiv2024

Beyond English-Centric LLMs: What Language Do Multilingual Language Models Think in?

In this study, we investigate whether non-English-centric LLMs, despite their strong performance, `think' in their respective dominant language: more precisely, `think' refers to how the representations of intermediate layers, when un-embedded into the vocabulary space, exhibit higher probabilities for certain dominant languages during generation. We term such languages as internal $\textbf{latent languages}$. We examine the latent language of three typical categories of models for Japanese processing: Llama2, an English-centric model; Swallow, an English-centric model with continued pre-training in Japanese; and LLM-jp, a model pre-trained on balanced English and Japanese corpora. Our empirical findings reveal that, unlike Llama2 which relies exclusively on English as the internal latent language, Japanese-specific Swallow and LLM-jp employ both Japanese and English, exhibiting dual internal latent languages. For any given target language, the model preferentially activates the latent language most closely related to it. In addition, we explore how intermediate layers respond to questions involving cultural conflicts between latent internal and target output languages. We further explore how the language identity shifts across layers while keeping consistent semantic meaning reflected in the intermediate layer representations. This study deepens the understanding of non-English-centric large language models, highlighting the intricate dynamics of language representation within their intermediate layers.

cs.CL↗

arXiv2024

EMS: Efficient and Effective Massively Multilingual Sentence Embedding Learning

Massively multilingual sentence representation models, e.g., LASER, SBERT-distill, and LaBSE, help significantly improve cross-lingual downstream tasks. However, the use of a large amount of data or inefficient model architectures results in heavy computation to train a new model according to our preferred languages and domains. To resolve this issue, we introduce efficient and effective massively multilingual sentence embedding (EMS), using cross-lingual token-level reconstruction (XTR) and sentence-level contrastive learning as training objectives. Compared with related studies, the proposed model can be efficiently trained using significantly fewer parallel sentences and GPU computation resources. Empirical results showed that the proposed model significantly yields better or comparable results with regard to cross-lingual sentence retrieval, zero-shot cross-lingual genre classification, and sentiment classification. Ablative analyses demonstrated the efficiency and effectiveness of each component of the proposed model. We release the codes for model training and the EMS pre-trained sentence embedding model, which supports 62 languages ( https://github.com/Mao-KU/EMS ).

cs.CL↗

arXiv2024

MELD-ST: An Emotion-aware Speech Translation Dataset

Emotion plays a crucial role in human conversation. This paper underscores the significance of considering emotion in speech translation. We present the MELD-ST dataset for the emotion-aware speech translation task, comprising English-to-Japanese and English-to-German language pairs. Each language pair includes about 10,000 utterances annotated with emotion labels from the MELD dataset. Baseline experiments using the SeamlessM4T model on the dataset indicate that fine-tuning with emotion labels can enhance translation performance in some settings, highlighting the need for further research in emotion-aware speech translation systems.

cs.CL↗

arXiv2024

J-CRe3: A Japanese Conversation Dataset for Real-world Reference Resolution

Understanding expressions that refer to the physical world is crucial for such human-assisting systems in the real world, as robots that must perform actions that are expected by users. In real-world reference resolution, a system must ground the verbal information that appears in user interactions to the visual information observed in egocentric views. To this end, we propose a multimodal reference resolution task and construct a Japanese Conversation dataset for Real-world Reference Resolution (J-CRe3). Our dataset contains egocentric video and dialogue audio of real-world conversations between two people acting as a master and an assistant robot at home. The dataset is annotated with crossmodal tags between phrases in the utterances and the object bounding boxes in the video frames. These tags include indirect reference relations, such as predicate-argument structures and bridging references as well as direct reference relations. We also constructed an experimental model and clarified the challenges in multimodal reference resolution tasks.

cs.CL↗

arXiv2024

AcTED: Automatic Acquisition of Typical Event Duration for Semi-supervised Temporal Commonsense QA

We propose a voting-driven semi-supervised approach to automatically acquire the typical duration of an event and use it as pseudo-labeled data. The human evaluation demonstrates that our pseudo labels exhibit surprisingly high accuracy and balanced coverage. In the temporal commonsense QA task, experimental results show that using only pseudo examples of 400 events, we achieve performance comparable to the existing BERT-based weakly supervised approaches that require a significant amount of training examples. When compared to the RoBERTa baselines, our best approach establishes state-of-the-art performance with a 7% improvement in Exact Match.

cs.CL↗

arXiv2024

Rapidly Developing High-quality Instruction Data and Evaluation Benchmark for Large Language Models with Minimal Human Effort: A Case Study on Japanese

The creation of instruction data and evaluation benchmarks for serving Large language models often involves enormous human annotation. This issue becomes particularly pronounced when rapidly developing such resources for a non-English language like Japanese. Instead of following the popular practice of directly translating existing English resources into Japanese (e.g., Japanese-Alpaca), we propose an efficient self-instruct method based on GPT-4. We first translate a small amount of English instructions into Japanese and post-edit them to obtain native-level quality. GPT-4 then utilizes them as demonstrations to automatically generate Japanese instruction data. We also construct an evaluation benchmark containing 80 questions across 8 categories, using GPT-4 to automatically assess the response quality of LLMs without human references. The empirical results suggest that the models fine-tuned on our GPT-4 self-instruct data significantly outperformed the Japanese-Alpaca across all three base pre-trained models. Our GPT-4 self-instruct data allowed the LLaMA 13B model to defeat GPT-3.5 (Davinci-003) with a 54.37\% win-rate. The human evaluation exhibits the consistency between GPT-4's assessments and human preference. Our high-quality instruction data and evaluation benchmark have been released here.

cs.CL↗