Saku Sugawara

LLM-jp, Akiko Aizawa, Eiji Aramaki et al.

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/.

Xin Sun, Yue Su, Yifan Mo et al.

Building trustworthy AI systems for mental health support is a shared priority across stakeholders from multiple disciplines. However, "trustworthy" remains loosely defined and inconsistently operationalized. AI research often focuses on technical criteria (e.g., robustness, explainability, and safety), while therapeutic practitioners emphasize therapeutic fidelity (e.g., appropriateness, empathy, and long-term user outcomes). To bridge the fragmented landscape, we propose a three-layer trust framework, covering human-oriented, AI-oriented, and interaction-oriented trust, integrating the viewpoints of key stakeholders (e.g., practitioners, researchers, regulators). Using this framework, we systematically review existing AI-driven research in mental health domain and examine evaluation practices for ``trustworthy'' ranging from automatic metrics to clinically validated approaches. We highlight critical gaps between what NLP currently measures and what real-world mental health contexts require, and outline a research agenda for building socio-technically aligned and genuinely trustworthy AI for mental health support.

Johannes Mario Meissner, Saku Sugawara, Akiko Aizawa

Debiasing language models from unwanted behaviors in Natural Language Understanding tasks is a topic with rapidly increasing interest in the NLP community. Spurious statistical correlations in the data allow models to perform shortcuts and avoid uncovering more advanced and desirable linguistic features. A multitude of effective debiasing approaches has been proposed, but flexibility remains a major issue. For the most part, models must be retrained to find a new set of weights with debiased behavior. We propose a new debiasing method in which we identify debiased pruning masks that can be applied to a finetuned model. This enables the selective and conditional application of debiasing behaviors. We assume that bias is caused by a certain subset of weights in the network; our method is, in essence, a mask search to identify and remove biased weights. Our masks show equivalent or superior performance to the standard counterparts, while offering important benefits. Pruning masks can be stored with high efficiency in memory, and it becomes possible to switch among several debiasing behaviors (or revert back to the original biased model) at inference time. Finally, it opens the doors to further research on how biases are acquired by studying the generated masks. For example, we observed that the early layers and attention heads were pruned more aggressively, possibly hinting towards the location in which biases may be encoded.

Kazutoshi Shinoda, Saku Sugawara, Akiko Aizawa

Question answering (QA) models for reading comprehension tend to learn shortcut solutions rather than the solutions intended by QA datasets. QA models that have learned shortcut solutions can achieve human-level performance in shortcut examples where shortcuts are valid, but these same behaviors degrade generalization potential on anti-shortcut examples where shortcuts are invalid. Various methods have been proposed to mitigate this problem, but they do not fully take the characteristics of shortcuts themselves into account. We assume that the learnability of shortcuts, i.e., how easy it is to learn a shortcut, is useful to mitigate the problem. Thus, we first examine the learnability of the representative shortcuts on extractive and multiple-choice QA datasets. Behavioral tests using biased training sets reveal that shortcuts that exploit answer positions and word-label correlations are preferentially learned for extractive and multiple-choice QA, respectively. We find that the more learnable a shortcut is, the flatter and deeper the loss landscape is around the shortcut solution in the parameter space. We also find that the availability of the preferred shortcuts tends to make the task easier to perform from an information-theoretic viewpoint. Lastly, we experimentally show that the learnability of shortcuts can be utilized to construct an effective QA training set; the more learnable a shortcut is, the smaller the proportion of anti-shortcut examples required to achieve comparable performance on shortcut and anti-shortcut examples. We claim that the learnability of shortcuts should be considered when designing mitigation methods.

Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara et al.

To explain the predicted answers and evaluate the reasoning abilities of models, several studies have utilized underlying reasoning (UR) tasks in multi-hop question answering (QA) datasets. However, it remains an open question as to how effective UR tasks are for the QA task when training models on both tasks in an end-to-end manner. In this study, we address this question by analyzing the effectiveness of UR tasks (including both sentence-level and entity-level tasks) in three aspects: (1) QA performance, (2) reasoning shortcuts, and (3) robustness. While the previous models have not been explicitly trained on an entity-level reasoning prediction task, we build a multi-task model that performs three tasks together: sentence-level supporting facts prediction, entity-level reasoning prediction, and answer prediction. Experimental results on 2WikiMultiHopQA and HotpotQA-small datasets reveal that (1) UR tasks can improve QA performance. Using four debiased datasets that are newly created, we demonstrate that (2) UR tasks are helpful in preventing reasoning shortcuts in the multi-hop QA task. However, we find that (3) UR tasks do not contribute to improving the robustness of the model on adversarial questions, such as sub-questions and inverted questions. We encourage future studies to investigate the effectiveness of entity-level reasoning in the form of natural language questions (e.g., sub-question forms).

Kazutoshi Shinoda, Saku Sugawara, Akiko Aizawa

Extractive question answering (QA) models tend to exploit spurious correlations to make predictions when a training set has unintended biases. This tendency results in models not being generalizable to examples where the correlations do not hold. Determining the spurious correlations QA models can exploit is crucial in building generalizable QA models in real-world applications; moreover, a method needs to be developed that prevents these models from learning the spurious correlations even when a training set is biased. In this study, we discovered that the relative position of an answer, which is defined as the relative distance from an answer span to the closest question-context overlap word, can be exploited by QA models as superficial cues for making predictions. Specifically, we find that when the relative positions in a training set are biased, the performance on examples with relative positions unseen during training is significantly degraded. To mitigate the performance degradation for unseen relative positions, we propose an ensemble-based debiasing method that does not require prior knowledge about the distribution of relative positions. We demonstrate that the proposed method mitigates the models' reliance on relative positions using the biased and full SQuAD dataset. We hope that this study can help enhance the generalization ability of QA models in real-world applications.

Kazushi Kondo, Saku Sugawara, Akiko Aizawa

In this study, we create a CConS (Counter-commonsense Contextual Size comparison) dataset to investigate how physical commonsense affects the contextualized size comparison task; the proposed dataset consists of both contexts that fit physical commonsense and those that do not. This dataset tests the ability of language models to predict the size relationship between objects under various contexts generated from our curated noun list and templates. We measure the ability of several masked language models and generative models. The results show that while large language models can use prepositions such as ``in'' and ``into'' in the provided context to infer size relationships, they fail to use verbs and thus make incorrect judgments led by their prior physical commonsense.

Hongkuan Zhang, Saku Sugawara, Akiko Aizawa et al.

Image captioning models require the high-level generalization ability to describe the contents of various images in words. Most existing approaches treat the image-caption pairs equally in their training without considering the differences in their learning difficulties. Several image captioning approaches introduce curriculum learning methods that present training data with increasing levels of difficulty. However, their difficulty measurements are either based on domain-specific features or prior model training. In this paper, we propose a simple yet efficient difficulty measurement for image captioning using cross-modal similarity calculated by a pretrained vision-language model. Experiments on the COCO and Flickr30k datasets show that our proposed approach achieves superior performance and competitive convergence speed to baselines without requiring heuristics or incurring additional training costs. Moreover, the higher model performance on difficult examples and unseen data also demonstrates the generalization ability.

Xanh Ho, Johannes Mario Meissner, Saku Sugawara et al.

The issue of shortcut learning is widely known in NLP and has been an important research focus in recent years. Unintended correlations in the data enable models to easily solve tasks that were meant to exhibit advanced language understanding and reasoning capabilities. In this survey paper, we focus on the field of machine reading comprehension (MRC), an important task for showcasing high-level language understanding that also suffers from a range of shortcuts. We summarize the available techniques for measuring and mitigating shortcuts and conclude with suggestions for further progress in shortcut research. Importantly, we highlight two concerns for shortcut mitigation in MRC: (1) the lack of public challenge sets, a necessary component for effective and reusable evaluation, and (2) the lack of certain mitigation techniques that are prominent in other areas.

Saku Sugawara, Nikita Nangia, Alex Warstadt et al.

For a natural language understanding benchmark to be useful in research, it has to consist of examples that are diverse and difficult enough to discriminate among current and near-future state-of-the-art systems. However, we do not yet know how best to select text sources to collect a variety of challenging examples. In this study, we crowdsource multiple-choice reading comprehension questions for passages taken from seven qualitatively distinct sources, analyzing what attributes of passages contribute to the difficulty and question types of the collected examples. To our surprise, we find that passage source, length, and readability measures do not significantly affect question difficulty. Through our manual annotation of seven reasoning types, we observe several trends between passage sources and reasoning types, e.g., logical reasoning is more often required in questions written for technical passages. These results suggest that when creating a new benchmark dataset, selecting a diverse set of passages can help ensure a diverse range of question types, but that passage difficulty need not be a priority.

Xanh Ho, Saku Sugawara, Akiko Aizawa

Several multi-hop reading comprehension datasets have been proposed to resolve the issue of reasoning shortcuts by which questions can be answered without performing multi-hop reasoning. However, the ability of multi-hop models to perform step-by-step reasoning when finding an answer to a comparison question remains unclear. It is also unclear how questions about the internal reasoning process are useful for training and evaluating question-answering (QA) systems. To evaluate the model precisely in a hierarchical manner, we first propose a dataset, \textit{HieraDate}, with three probing tasks in addition to the main question: extraction, reasoning, and robustness. Our dataset is created by enhancing two previous multi-hop datasets, HotpotQA and 2WikiMultiHopQA, focusing on multi-hop questions on date information that involve both comparison and numerical reasoning. We then evaluate the ability of existing models to understand date information. Our experimental results reveal that the multi-hop models do not have the ability to subtract two dates even when they perform well in date comparison and number subtraction tasks. Other results reveal that our probing questions can help to improve the performance of the models (e.g., by +10.3 F1) on the main QA task and our dataset can be used for data augmentation to improve the robustness of the models.

Kazutoshi Shinoda, Saku Sugawara, Akiko Aizawa

Question answering (QA) models are shown to be insensitive to large perturbations to inputs; that is, they make correct and confident predictions even when given largely perturbed inputs from which humans can not correctly derive answers. In addition, QA models fail to generalize to other domains and adversarial test sets, while humans maintain high accuracy. Based on these observations, we assume that QA models do not use intended features necessary for human reading but rely on spurious features, causing the lack of generalization ability. Therefore, we attempt to answer the question: If the overconfident predictions of QA models for various types of perturbations are penalized, will the out-of-distribution (OOD) generalization be improved? To prevent models from making confident predictions on perturbed inputs, we first follow existing studies and maximize the entropy of the output probability for perturbed inputs. However, we find that QA models trained to be sensitive to a certain perturbation type are often insensitive to unseen types of perturbations. Thus, we simultaneously maximize the entropy for the four perturbation types (i.e., word- and sentence-level shuffling and deletion) to further close the gap between models and humans. Contrary to our expectations, although models become sensitive to the four types of perturbations, we find that the OOD generalization is not improved. Moreover, the OOD generalization is sometimes degraded after entropy maximization. Making unconfident predictions on largely perturbed inputs per se may be beneficial to gaining human trust. However, our negative results suggest that researchers should pay attention to the side effect of entropy maximization.

Mana Ashida, Saku Sugawara

The possible consequences for the same context may vary depending on the situation we refer to. However, current studies in natural language processing do not focus on situated commonsense reasoning under multiple possible scenarios. This study frames this task by asking multiple questions with the same set of possible endings as candidate answers, given a short story text. Our resulting dataset, Possible Stories, consists of more than 4.5K questions over 1.3K story texts in English. We discover that even current strong pretrained language models struggle to answer the questions consistently, highlighting that the highest accuracy in an unsupervised setting (60.2%) is far behind human accuracy (92.5%). Through a comparison with existing datasets, we observe that the questions in our dataset contain minimal annotation artifacts in the answer options. In addition, our dataset includes examples that require counterfactual reasoning, as well as those requiring readers' reactions and fictional information, suggesting that our dataset can serve as a challenging testbed for future studies on situated commonsense reasoning.

Miyu Oba, Saku Sugawara

Recent work has examined language models from a linguistic perspective to better understand how they acquire language. Most existing benchmarks focus on judging grammatical acceptability, whereas the ability to interpret meanings conveyed by grammatical forms has received much less attention. We introduce the Linguistic Minimal-Pair Benchmark for Evaluating Constructional Understanding in Language Models (CxMP), a benchmark grounded in Construction Grammar that treats form-meaning pairings, or constructions, as fundamental linguistic units. CxMP evaluates whether models can interpret the semantic relations implied by constructions, using a controlled minimal-pair design across nine construction types, including the let-alone, caused motion, and ditransitive constructions. Our results show that while syntactic competence emerges early, constructional understanding develops more gradually and remains limited even in large language models (LLMs). CxMP thus reveals persistent gaps in how language models integrate form and meaning, providing a framework for studying constructional understanding and learning trajectories in language models.

Akira Kawabata, Saku Sugawara

To precisely evaluate a language model's capability for logical reading comprehension, we present a dataset for testing the understanding of the rationale behind critical reasoning. For questions taken from an existing multiplechoice logical reading comprehension dataset, we crowdsource rationale texts that explain why we should select or eliminate answer options, resulting in 3,003 multiple-choice subquestions that are associated with 943 main questions. Experiments on our dataset show that recent large language models (e.g., InstructGPT) struggle to answer the subquestions even if they are able to answer the main questions correctly. We find that the models perform particularly poorly in answering subquestions written for the incorrect options of the main questions, implying that the models have a limited capability for explaining why incorrect alternatives should be eliminated. These results suggest that our dataset encourages further investigation into the critical reasoning ability of language models while focusing on the elimination process of relevant alternatives.

Xin Sun, Shu Wei, Ting Pan et al.

Modeling users' cognitive states (e.g., cognitive load and decision confidence) is essential for building adaptive AI in high-stakes decision-making. While eye tracking provides non-invasive behavioral signals correlated with cognitive effort, prior work has not systematically examined how AI assistance contexts, specifically varying advice reliability and user heterogeneity, can alter the mapping between gaze signals and cognitive states. We conducted a within-subject lab eye-tracking study (N=54) on factual verification tasks under three conditions: No-AI, Correct-AI advice, and Incorrect-AI advice. We analyze condition-dependent changes in self-reports and eye-tracking patterns and evaluate the robustness of eye-tracking-based user modeling. Results show that AI advice increases decision confidence compared to No-AI, while Correct-AI is associated with lower perceived cognitive load and more efficient gaze behavior. Crucially, predictive modeling is context-sensitive: the relationship between eye-tracking signals and cognitive states shifts across AI conditions. Finally, fusing eye-tracking features with user priors (demographics, AI literacy/experience, and propensity to trust technology) improves cross-participant generalization. These findings support condition-aware and personalized user modeling for cognitively aligned adaptive AI systems.

Akira Kawabata, Saku Sugawara

Rubric-augmented verification guides reward models with explicit evaluation criteria, yielding more reliable judgments than single-model verification. However, most existing methods require costly rubric annotations, limiting scalability. Moreover, we find that rubric generation is vulnerable to a failure of cooperation; low-quality rubrics actively mislead reward models rather than help. Inspired by the principle of cooperative communication, we propose Cooperative yet Critical reward modeling (C2), a framework that significantly improves reward model judgments by having the reward model critically collaborate with a rubric generator trained solely from binary preferences. In C2, we synthesize helpful and misleading rubric pairs by measuring how each rubric shifts the reward model toward or away from the correct preference. Using these contrastive pairs, we train a cooperative rubric generator to propose helpful rubrics, and a critical verifier to assess rubric validity before making its judgment, following only rubrics it deems helpful at inference time. C2 outperforms reasoning reward models trained on the same binary preferences, with gains of up to 6.5 points on RM-Bench and 6.0 points length-controlled win rate on AlpacaEval 2.0. Without external rubric annotations, C2 enables an 8B reward model to match performance achieved with rubrics from a 4$\times$ larger model. Overall, our work demonstrates that eliciting deliberate cooperation in rubric-augmented verification makes reward models more trustworthy in a scalable way.

Momoka Furuhashi, Kouta Nakayama, Takashi Kodama et al.

Automatic evaluation of generative tasks using large language models faces challenges due to ambiguous criteria. Although automatic checklist generation is a potentially promising approach, its usefulness remains underexplored. We investigate whether checklists should be used for all questions or selectively, generate them using six methods, evaluate their effectiveness across eight model sizes, and identify checklist items that correlate with human evaluations. Through experiments on pairwise comparison and direct scoring tasks, we find that selective checklist use tends to improve evaluation performance in pairwise settings, while its benefits are less consistent in direct scoring. Our analysis also shows that even checklist items with low correlation to human scores often reflect human-written criteria, indicating potential inconsistencies in human evaluation. These findings highlight the need to more clearly define objective evaluation criteria to guide both human and automatic evaluations. \footnote{Our code is available at~https://github.com/momo0817/checklist-effectiveness-study

Yuchen Guo, Zhicheng Dou, Huy H. Nguyen et al.

Content creation has dramatically progressed with the rapid advancement of large language models like ChatGPT and Claude. While this progress has greatly enhanced various aspects of life and work, it has also negatively affected certain areas of society. A recent survey revealed that nearly 30% of college students use generative AI to help write academic papers and reports. Most countermeasures treat the detection of AI-generated text as a binary classification task and thus lack robustness. This approach overlooks human involvement in the generation of content even though human-machine collaboration is becoming mainstream. Besides generating entire texts, people may use machines to complete or revise texts. Such human involvement varies case by case, which makes binary classification a less than satisfactory approach. We refer to this situation as participation detection obfuscation. We propose using BERTScore as a metric to measure human involvement in the generation process and a multi-task RoBERTa-based regressor trained on a token classification task to address this problem. To evaluate the effectiveness of this approach, we simulated academic-based scenarios and created a continuous dataset reflecting various levels of human involvement. All of the existing detectors we examined failed to detect the level of human involvement on this dataset. Our method, however, succeeded (F1 score of 0.9423 and a regressor mean squared error of 0.004). Moreover, it demonstrated some generalizability across generative models. Our code is available at https://github.com/gyc-nii/CAS-CS-and-dual-head-detector

Julian Schnitzler, Xanh Ho, Jiahao Huang et al.

Most existing multi-hop datasets are extractive answer datasets, where the answers to the questions can be extracted directly from the provided context. This often leads models to use heuristics or shortcuts instead of performing true multi-hop reasoning. In this paper, we propose a new multi-hop dataset, MoreHopQA, which shifts from extractive to generative answers. Our dataset is created by utilizing three existing multi-hop datasets: HotpotQA, 2WikiMultihopQA, and MuSiQue. Instead of relying solely on factual reasoning, we enhance the existing multi-hop questions by adding another layer of questioning that involves one, two, or all three of the following types of reasoning: commonsense, arithmetic, and symbolic. Our dataset is created through a semi-automated process, resulting in a dataset with 1,118 samples that have undergone human verification. We then use our dataset to evaluate five different large language models: Mistral 7B, Gemma 7B, Llama 3 (8B and 70B), and GPT-4. We also design various cases to analyze the reasoning steps in the question-answering process. Our results show that models perform well on initial multi-hop questions but struggle with our extended questions, indicating that our dataset is more challenging than previous ones. Our analysis of question decomposition reveals that although models can correctly answer questions, only a portion - 38.7% for GPT-4 and 33.4% for Llama3-70B - achieve perfect reasoning, where all corresponding sub-questions are answered correctly. Evaluation code and data are available at https://github.com/Alab-NII/morehopqa

Kazutoshi Shinoda, Saku Sugawara, Akiko Aizawa

Question answering (QA) models for reading comprehension have been demonstrated to exploit unintended dataset biases such as question-context lexical overlap. This hinders QA models from generalizing to under-represented samples such as questions with low lexical overlap. Question generation (QG), a method for augmenting QA datasets, can be a solution for such performance degradation if QG can properly debias QA datasets. However, we discover that recent neural QG models are biased towards generating questions with high lexical overlap, which can amplify the dataset bias. Moreover, our analysis reveals that data augmentation with these QG models frequently impairs the performance on questions with low lexical overlap, while improving that on questions with high lexical overlap. To address this problem, we use a synonym replacement-based approach to augment questions with low lexical overlap. We demonstrate that the proposed data augmentation approach is simple yet effective to mitigate the degradation problem with only 70k synthetic examples. Our data is publicly available at https://github.com/KazutoshiShinoda/Synonym-Replacement.

Kazutoshi Shinoda, Saku Sugawara, Akiko Aizawa

Question answering (QA) models for reading comprehension have achieved human-level accuracy on in-distribution test sets. However, they have been demonstrated to lack robustness to challenge sets, whose distribution is different from that of training sets. Existing data augmentation methods mitigate this problem by simply augmenting training sets with synthetic examples sampled from the same distribution as the challenge sets. However, these methods assume that the distribution of a challenge set is known a priori, making them less applicable to unseen challenge sets. In this study, we focus on question-answer pair generation (QAG) to mitigate this problem. While most existing QAG methods aim to improve the quality of synthetic examples, we conjecture that diversity-promoting QAG can mitigate the sparsity of training sets and lead to better robustness. We present a variational QAG model that generates multiple diverse QA pairs from a paragraph. Our experiments show that our method can improve the accuracy of 12 challenge sets, as well as the in-distribution accuracy. Our code and data are available at https://github.com/KazutoshiShinoda/VQAG.

Xin Sun, Di Wu, Sijing Qin et al.

Large language models (LLMs) are increasingly used as automated evaluators (LLM-as-a-Judge). This work challenges its reliability by showing that trust judgments by LLMs are biased by disclosed source labels. Using a counterfactual design, we find that both humans and LLM judges assign higher trust to information labeled as human-authored than to the same content labeled as AI-generated. Eye-tracking data reveal that humans rely heavily on source labels as heuristic cues for judgments. We analyze LLM internal states during judgment. Across label conditions, models allocate denser attention to the label region than the content region, and this label dominance is stronger under Human labels than AI labels, consistent with the human gaze patterns. Besides, decision uncertainty measured by logits is higher under AI labels than Human labels. These results indicate that the source label is a salient heuristic cue for both humans and LLMs. It raises validity concerns for label-sensitive LLM-as-a-Judge evaluation, and we cautiously raise that aligning models with human preferences may propagate human heuristic reliance into models, motivating debiased evaluation and alignment.

Momoka Furuhashi, Kouta Nakayama, Noboru Kawai et al.

Large language models (LLMs) show promise for automatically generating feedback in education settings. However, it remains unclear how specific feedback elements, such as tone and information coverage, contribute to learning outcomes and learner acceptance, particularly across learners with different personality traits. In this study, we define six feedback elements and generate feedback for multiple-choice biology questions using GPT-5. We conduct a learning experiment with 321 first-year high school students and evaluate feedback effectiveness using two learning outcomes measures and subjective evaluations across six criteria. We further analyze differences in how feedback acceptance varies across learners based on Big Five personality traits. Our results show that effective feedback elements share common patterns supporting learning outcomes, while learners' subjective preferences differ across personality-based clusters. These findings highlight the importance of selecting and adapting feedback elements according to learners' personality traits when we design LLM-generated feedback, and provide practical implications for personalized feedback design in education.

Rei Emura, Saku Sugawara

Language models (LMs) behave more like humans when their cognitive resources are restricted, particularly in predicting sentence processing costs such as reading times. However, it remains unclear whether such constraints similarly affect sentence comprehension strategies. Besides, existing methods do not directly target the balance between memory storage and sentence processing, which is central to human working memory. To address this issue, we propose a dual-task paradigm that combines an arithmetic computation task with a sentence comprehension task, such as "The 2 cocktail + blended 3 =..." Our experiments show that under dual-task conditions, GPT-4o, o3-mini, and o4-mini shift toward plausibility-based comprehension, mirroring humans' rational inference. Specifically, these models show a greater accuracy gap between plausible sentences (e.g., "The cocktail was blended by the bartender") and implausible sentences (e.g., "The bartender was blended by the cocktail") in the dual-task condition compared to the single-task conditions. These findings suggest that constraints on the balance between memory and processing resources promote rational inference in LMs. More broadly, they support the view that human-like sentence comprehension fundamentally arises from the allocation of limited cognitive resources.

Xin Sun, Shu Wei, Jos A Bosch et al.

Large Language Models (LLMs) increasingly show reasoning rationales alongside their answers, turning "reasoning" into a user-interface element. While step-by-step rationales are typically associated with model performance, how they influence users' trust and decision-making in factual verification tasks remains unclear. We ran an online study (N=68) manipulating three properties of LLM reasoning rationales: presentation format (instant vs. delayed vs. on-demand), correctness (correct vs. incorrect), and certainty framing (none vs. certain vs. uncertain). We found that correct rationales and certainty cues increased trust, decision confidence, and AI advice adoption, whereas uncertainty cues reduced them. Presentation format did not have a significant effect, suggesting users were less sensitive to how reasoning was revealed than to its reliability. Participants indicated they use rationales to primarily audit outputs and calibrate trust, where they expected rationales in stepwise, adaptive forms with certainty indicators. Our work shows that user-facing rationales, if poorly designed, can both support decision-making yet miscalibrate trust.

Momoka Furuhashi, Hiroaki Funayama, Yuya Iwase et al.

Short-reading comprehension questions help students understand text structure but lack effective feedback. Students struggle to identify and correct errors, while manual feedback creation is labor-intensive. This highlights the need for automated feedback linking responses to a scoring rubric for deeper comprehension. Despite advances in Natural Language Processing (NLP), research has focused on automatic grading, with limited work on feedback generation. To address this, we propose a system that generates feedback for student responses. Our contributions are twofold. First, we introduce the first system for feedback on short-answer reading comprehension. These answers are derived from the text, requiring structural understanding. We propose an "answer diagnosis graph," integrating the text's logical structure with feedback templates. Using this graph and NLP techniques, we estimate students' comprehension and generate targeted feedback. Second, we evaluate our feedback through an experiment with Japanese high school students (n=39). They answered two 70-80 word questions and were divided into two groups with minimal academic differences. One received a model answer, the other system-generated feedback. Both re-answered the questions, and we compared score changes. A questionnaire assessed perceptions and motivation. Results showed no significant score improvement between groups, but system-generated feedback helped students identify errors and key points in the text. It also significantly increased motivation. However, further refinement is needed to enhance text structure understanding.

Daiki Asami, Saku Sugawara

What makes a presupposition of an utterance -- information taken for granted by its speaker -- different from other pragmatic inferences such as an entailment is projectivity (e.g., the negative sentence the boy did not stop shedding tears presupposes the boy had shed tears before). The projectivity may vary depending on the combination of presupposition triggers and environments. However, prior natural language understanding studies fail to take it into account as they either use no human baseline or include only negation as an entailment-canceling environment to evaluate models' performance. The current study attempts to reconcile these issues. We introduce a new dataset, projectivity of presupposition (PROPRES, which includes 12k premise-hypothesis pairs crossing six triggers involving some lexical variety with five environments. Our human evaluation reveals that humans exhibit variable projectivity in some cases. However, the model evaluation shows that the best-performed model, DeBERTa, does not fully capture it. Our findings suggest that probing studies on pragmatic inferences should take extra care of the human judgment variability and the combination of linguistic items.

Yoko Kayano, Saku Sugawara

In professional settings, translation is guided by communicative goals and client needs, often formalized as specifications. While existing evaluation frameworks acknowledge the importance of such specifications, these specifications are often treated only implicitly in machine translation (MT) research. Drawing on translation studies, we provide a theoretical rationale for why specifications matter in professional translation, as well as a practical guide to implementing specification-aware MT and evaluation. Building on this foundation, we apply our framework to the translation of investor relations texts from 33 publicly listed companies. In our experiment, we compare five translation types, including official human translations and prompt-based outputs from large language models (LLMs), using expert error analysis, user preference rankings, and an automatic metric. The results show that LLM translations guided by specifications consistently outperformed official human translations in human evaluations, highlighting a gap between perceived and expected quality. These findings demonstrate that integrating specifications into MT workflows, with human oversight, can improve translation quality in ways aligned with professional practice.

Yiwei Liu, Emma Jane Pretty, Jiahao Huang et al.

While recent studies explore Large Language Models' (LLMs) performance on Theory of Mind (ToM) reasoning tasks, research on ToM abilities that require more nuanced social context is limited, such as white lies. We introduce TactfulToM, a novel English benchmark designed to evaluate LLMs' ability to understand white lies within real-life conversations and reason about prosocial motivations behind them, particularly when they are used to spare others' feelings and maintain social harmony. Our benchmark is generated through a multi-stage human-in-the-loop pipeline where LLMs expand manually designed seed stories into conversations to maintain the information asymmetry between participants necessary for authentic white lies. We show that TactfulToM is challenging for state-of-the-art models, which perform substantially below humans, revealing shortcomings in their ability to fully comprehend the ToM reasoning that enables true understanding of white lies.

Futa Waseda, Saku Sugawara, Isao Echizen

Defending pre-trained vision-language models (VLMs), such as CLIP, against adversarial attacks is crucial, as these models are widely used in diverse zero-shot tasks, including image classification. However, existing adversarial training (AT) methods for robust fine-tuning largely overlook the role of language in enhancing visual robustness. Specifically, (1) supervised AT methods rely on short texts (e.g., class labels) to generate adversarial perturbations, leading to overfitting to object classes in the training data, and (2) unsupervised AT avoids this overfitting but remains suboptimal against practical text-guided adversarial attacks due to its lack of semantic guidance. To address these limitations, we propose Quality Text-guided Adversarial Fine-Tuning (QT-AFT), which leverages high-quality captions during training to guide adversarial examples away from diverse semantics present in images. This enables the visual encoder to robustly recognize a broader range of image features even under adversarial noise, thereby enhancing robustness across diverse downstream tasks. QT-AFT overcomes the key weaknesses of prior methods -- overfitting in supervised AT and lack of semantic awareness in unsupervised AT -- achieving state-of-the-art zero-shot adversarial robustness and clean accuracy, evaluated across 16 zero-shot datasets. Furthermore, our comprehensive study uncovers several key insights into the role of language in enhancing vision robustness; for example, describing object properties in addition to object names further enhances zero-shot robustness. Our findings point to an urgent direction for future work -- centering high-quality linguistic supervision in robust visual representation learning.

Daiki Asami, Saku Sugawara

Psycholinguistic research suggests that humans may build a representation of linguistic input that is 'good-enough' for the task at hand. This study examines what architectural features make language models learn human-like good-enough language processing. We focus on the number of layers and self-attention heads in Transformers. We create a good-enough language processing (GELP) evaluation dataset (7,680 examples), which is designed to test the effects of two plausibility types, eight construction types, and three degrees of memory cost on language processing. To annotate GELP, we first conduct a crowdsourcing experiment whose design follows prior psycholinguistic studies. Our model evaluation against the annotated GELP then reveals that the full model as well as models with fewer layers and/or self-attention heads exhibit a good-enough performance. This result suggests that models with shallower depth and fewer heads can learn good-enough language processing.

Saku Sugawara, Shun Tsugita

Natural language understanding (NLU) studies often exaggerate or underestimate the capabilities of systems, thereby limiting the reproducibility of their findings. These erroneous evaluations can be attributed to the difficulty of defining and testing NLU adequately. In this position paper, we reconsider this challenge by identifying two types of researcher degrees of freedom. We revisit Turing's original interpretation of the Turing test and indicate that an NLU test does not provide an operational definition; it merely provides inductive evidence that the test subject understands the language sufficiently well to meet stakeholder objectives. In other words, stakeholders are free to arbitrarily define NLU through their objectives. To use the test results as inductive evidence, stakeholders must carefully assess if the interpretation of test scores is valid or not. However, designing and using NLU tests involve other degrees of freedom, such as specifying target skills and defining evaluation metrics. As a result, achieving consensus among stakeholders becomes difficult. To resolve this issue, we propose a validity argument, which is a framework comprising a series of validation criteria across test components. By demonstrating that current practices in NLU studies can be associated with those criteria and organizing them into a comprehensive checklist, we prove that the validity argument can serve as a coherent guideline for designing credible test sets and facilitating scientific communication.

Johannes Mario Meissner, Napat Thumwanit, Saku Sugawara et al.

Natural Language Inference (NLI) datasets contain examples with highly ambiguous labels. While many research works do not pay much attention to this fact, several recent efforts have been made to acknowledge and embrace the existence of ambiguity, such as UNLI and ChaosNLI. In this paper, we explore the option of training directly on the estimated label distribution of the annotators in the NLI task, using a learning loss based on this ambiguity distribution instead of the gold-labels. We prepare AmbiNLI, a trial dataset obtained from readily available sources, and show it is possible to reduce ChaosNLI divergence scores when finetuning on this data, a promising first step towards learning how to capture linguistic ambiguity. Additionally, we show that training on the same amount of data but targeting the ambiguity distribution instead of gold-labels can result in models that achieve higher performance and learn better representations for downstream tasks.

Nikita Nangia, Saku Sugawara, Harsh Trivedi et al.

Crowdsourcing is widely used to create data for common natural language understanding tasks. Despite the importance of these datasets for measuring and refining model understanding of language, there has been little focus on the crowdsourcing methods used for collecting the datasets. In this paper, we compare the efficacy of interventions that have been proposed in prior work as ways of improving data quality. We use multiple-choice question answering as a testbed and run a randomized trial by assigning crowdworkers to write questions under one of four different data collection protocols. We find that asking workers to write explanations for their examples is an ineffective stand-alone strategy for boosting NLU example difficulty. However, we find that training crowdworkers, and then using an iterative process of collecting data, sending feedback, and qualifying workers based on expert judgments is an effective means of collecting challenging data. But using crowdsourced, instead of expert judgments, to qualify workers and send feedback does not prove to be effective. We observe that the data from the iterative protocol with expert assessments is more challenging by several measures. Notably, the human--model gap on the unanimous agreement portion of this data is, on average, twice as large as the gap for the baseline protocol data.

Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara et al.

A multi-hop question answering (QA) dataset aims to test reasoning and inference skills by requiring a model to read multiple paragraphs to answer a given question. However, current datasets do not provide a complete explanation for the reasoning process from the question to the answer. Further, previous studies revealed that many examples in existing multi-hop datasets do not require multi-hop reasoning to answer a question. In this study, we present a new multi-hop QA dataset, called 2WikiMultiHopQA, which uses structured and unstructured data. In our dataset, we introduce the evidence information containing a reasoning path for multi-hop questions. The evidence information has two benefits: (i) providing a comprehensive explanation for predictions and (ii) evaluating the reasoning skills of a model. We carefully design a pipeline and a set of templates when generating a question-answer pair that guarantees the multi-hop steps and the quality of the questions. We also exploit the structured format in Wikidata and use logical rules to create questions that are natural but still require multi-hop reasoning. Through experiments, we demonstrate that our dataset is challenging for multi-hop models and it ensures that multi-hop reasoning is required.

Saku Sugawara, Pontus Stenetorp, Akiko Aizawa

Machine reading comprehension (MRC) has received considerable attention as a benchmark for natural language understanding. However, the conventional task design of MRC lacks explainability beyond the model interpretation, i.e., reading comprehension by a model cannot be explained in human terms. To this end, this position paper provides a theoretical basis for the design of MRC datasets based on psychology as well as psychometrics, and summarizes it in terms of the prerequisites for benchmarking MRC. We conclude that future datasets should (i) evaluate the capability of the model for constructing a coherent and grounded representation to understand context-dependent situations and (ii) ensure substantive validity by shortcut-proof questions and explanation as a part of the task design.

Saku Sugawara, Pontus Stenetorp, Kentaro Inui et al.

Existing analysis work in machine reading comprehension (MRC) is largely concerned with evaluating the capabilities of systems. However, the capabilities of datasets are not assessed for benchmarking language understanding precisely. We propose a semi-automated, ablation-based methodology for this challenge; By checking whether questions can be solved even after removing features associated with a skill requisite for language understanding, we evaluate to what degree the questions do not require the skill. Experiments on 10 datasets (e.g., CoQA, SQuAD v2.0, and RACE) with a strong baseline model show that, for example, the relative scores of a baseline model provided with content words only and with shuffled sentence words in the context are on average 89.2% and 78.5% of the original score, respectively. These results suggest that most of the questions already answered correctly by the model do not necessarily require grammatical and complex reasoning. For precise benchmarking, MRC datasets will need to take extra care in their design to ensure that questions can correctly evaluate the intended skills.

Saku Sugawara, Kentaro Inui, Satoshi Sekine et al.

A challenge in creating a dataset for machine reading comprehension (MRC) is to collect questions that require a sophisticated understanding of language to answer beyond using superficial cues. In this work, we investigate what makes questions easier across recent 12 MRC datasets with three question styles (answer extraction, description, and multiple choice). We propose to employ simple heuristics to split each dataset into easy and hard subsets and examine the performance of two baseline models for each of the subsets. We then manually annotate questions sampled from each subset with both validity and requisite reasoning skills to investigate which skills explain the difference between easy and hard questions. From this study, we observed that (i) the baseline performances for the hard subsets remarkably degrade compared to those of entire datasets, (ii) hard questions require knowledge inference and multiple-sentence reasoning in comparison with easy questions, and (iii) multiple-choice questions tend to require a broader range of reasoning skills than answer extraction and description questions. These results suggest that one might overestimate recent advances in MRC.