Akiko Aizawa

Mohit Gupta, Akiko Aizawa, Rajiv Ratn Shah

The emergence of large language models (LLMs) has significantly influenced numerous fields, including healthcare, by enhancing the capabilities of automated systems to process and generate human-like text. However, despite their advancements, the reliability and accuracy of LLMs in medical contexts remain critical concerns. Current evaluation methods often lack robustness and fail to provide a comprehensive assessment of LLM performance, leading to potential risks in clinical settings. In this work, we propose Med-CoDE, a specifically designed evaluation framework for medical LLMs to address these challenges. The framework leverages a critique-based approach to quantitatively measure the degree of disagreement between model-generated responses and established medical ground truths. This framework captures both accuracy and reliability in medical settings. The proposed evaluation framework aims to fill the existing gap in LLM assessment by offering a systematic method to evaluate the quality and trustworthiness of medical LLMs. Through extensive experiments and case studies, we illustrate the practicality of our framework in providing a comprehensive and reliable evaluation of medical LLMs.

Jiahao Huang, Fei Cheng, Junfeng Jiang et al.

Self-evolving agents improve over time by reflecting on past failures, but existing evaluation is limited in two ways: it measures only task scores, leaving reflection quality unknown, and it relies on agents' own episode runs, offering no mechanism to target specific failure patterns. We present \textbf{BenchTrace}, a benchmark for evaluating self-evolution ability in LLM agents. BenchTrace is built on a snapshot-reflection dataset of 1,821 annotated episodes spanning six diverse tasks, and comprises a \textbf{Reflection Evaluation} that probes failure identification through targeted QA tasks, and an \textbf{Evolution Evaluation} that tests whether past failure experience translates into avoidance behavior in a controlled self-evolution simulation. Building on BenchTrace, we propose \textbf{failure avoidance rate (FAR)}, a new evaluation metric measuring the fraction of test cases in which the agent successfully avoids the target failure instance. Experiments with Qwen3-32B and GPT-4.1 reveal that both models fall below a 30\% end-to-end pass rate on reflection evaluation, with diagnosis as the primary bottleneck. Evolution evaluation shows that self-evolution methods generally improve FAR over the non-evolving baseline, but agents forget early lessons as noise episodes accumulate, and agents fail to generalize their reflections beyond the specific context, causing negative transfer across task contexts. Our correlation analysis further reveals that only a fully correct reflection is strongly associated with higher FAR. BenchTrace exposes concrete limits of current self-evolution approaches and provides a controlled, model-agnostic framework for targeted evaluation.

Jiahao Huang, Fei Cheng, Junfeng Jiang et al.

Reasoning distillation transfers complex reasoning abilities from large language models (LLMs) to smaller ones, yet its success depends on how well the training data align with the student model. This paper introduces the Data-Model Compatibility (DMC) metric, which can be used to assess the suitability of a dataset for reasoning distillation on a student model. DMC provides an assessment by jointly considering data quality, relative difficulty, and student capability. We validated the effectiveness of DMC from two perspectives: (1) DMC exhibits a strong correlation with reasoning distillation performance; and (2) using DMC as the criterion for data selection leads to improved reasoning distillation performance. Both findings are consistently demonstrated across multiple student models and tasks. Moreover, since the DMC of each dataset dynamically changes during training, our experiments demonstrate that dynamically selecting datasets based on DMC can further enhance performance.

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/.

An Dao, Nhan Ly, Thao Tran et al.

Prion diseases are rare, rapidly progressive, and fatal neurodegenerative disorders that remain difficult to diagnose, particularly in their early stages because of nonspecific clinical presentations. However, to our knowledge, there is no publicly available prion-disease-focused dataset designed to capture a broad range of clinically relevant entities from the biomedical literature. We introduce PrionNER, a manually annotated named entity recognition dataset for prion disease clinical information in PubMed abstracts. The current release comprises 317 abstracts, 2,943 sentences, and 6,955 text-bound entity annotations spanning 15 coarse-grained and 31 fine-grained clinically oriented entity types covering diseases, symptoms, diagnostics, findings, anatomy, treatments, and temporal and statistical evidence. Inter-annotator agreement reaches 81.78 exact-match F1, indicating strong annotation consistency. We benchmark supervised BERT baselines, W2NER, and zero-shot extractors on PrionNER. W2NER is the strongest supervised model, and Gemma-4-31B is the strongest zero-shot model, but the benchmark remains challenging, especially for structurally complex mentions and fine-grained clinically adjacent label distinctions. PrionNER provides a clinically grounded benchmark for prion-disease information extraction and supports research on rare-disease biomedical NLP under low-resource, fine-grained, and non-flat extraction conditions. The dataset, annotation guidelines, and evaluation scripts are available at https://github.com/daotuanan/PrionNER/.

Timo Spinde, Manuel Plank, Jan-David Krieger et al.

Media coverage has a substantial effect on the public perception of events. Nevertheless, media outlets are often biased. One way to bias news articles is by altering the word choice. The automatic identification of bias by word choice is challenging, primarily due to the lack of a gold standard data set and high context dependencies. This paper presents BABE, a robust and diverse data set created by trained experts, for media bias research. We also analyze why expert labeling is essential within this domain. Our data set offers better annotation quality and higher inter-annotator agreement than existing work. It consists of 3,700 sentences balanced among topics and outlets, containing media bias labels on the word and sentence level. Based on our data, we also introduce a way to detect bias-inducing sentences in news articles automatically. Our best performing BERT-based model is pre-trained on a larger corpus consisting of distant labels. Fine-tuning and evaluating the model on our proposed supervised data set, we achieve a macro F1-score of 0.804, outperforming existing methods.

Chuang Ma, Qianying Liu, Tomoyuki Obuchi et al.

Multimodal large language models (MLLMs) remain unreliable on spatial multiple-choice questions, and their failures are often attributed to poorly attended visual information. In this work, we identify a complementary failure mode, spatial lexical bias: adding a spatial relation word to the answer options can attract the model's decision and make the newly added option likely to be selected. Using nine open-weight MLLMs, we show that this phenomenon is widely observed. In particular, models can answer a binary spatial question correctly, yet consistently select an incorrect third spatial option once it is added to the answer set. We isolate such binary-stable but ternary-fragile cases as diagnostic examples and leverage mechanistic interpretability tools, revealing that a substantial part of the failure instead originates on the language side rather than the visual side: visual attention analyses and residual-stream probes show the correct spatial relation remains internally available on these failures, while irrelevant-option controls, activation patching, and sparse component interventions trace the bias to specific LLM-side channels and neurons. Based on this finding, 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.

Sunisth Kumar, Xanh Ho, Tim Schopf et al.

Multimodal LLMs are increasingly used to assist scientific peer review, where a core requirement is verifying whether claims in a paper are supported by its evidence. Prior work has shown that models perform substantially better at this task when the evidence is a table than when it is a chart of the same underlying data. This raises the question of whether models fail to extract information from charts, or do they extract it but fail to use it when forming their prediction? We study this question through layer-wise linear probing and attention analysis on three open-weight VLMs over table and chart evidence, representing the same underlying data. We find consistent evidence for the latter. Chart information is encoded in the models' intermediate representations but does not reach the prediction position, a gap that is absent for tables and holds across all conditions tested. Attention analysis further reveals that this disconnect takes two architecturally distinct forms across model families. These findings reframe the table-chart gap as a failure of how encoded visual information is routed at prediction time, rather than a failure of encoding itself.

Martin Wessel, Tomáš Horych, Terry Ruas et al.

Although media bias detection is a complex multi-task problem, there is, to date, no unified benchmark grouping these evaluation tasks. We introduce the Media Bias Identification Benchmark (MBIB), a comprehensive benchmark that groups different types of media bias (e.g., linguistic, cognitive, political) under a common framework to test how prospective detection techniques generalize. After reviewing 115 datasets, we select nine tasks and carefully propose 22 associated datasets for evaluating media bias detection techniques. We evaluate MBIB using state-of-the-art Transformer techniques (e.g., T5, BART). Our results suggest that while hate speech, racial bias, and gender bias are easier to detect, models struggle to handle certain bias types, e.g., cognitive and political bias. However, our results show that no single technique can outperform all the others significantly. We also find an uneven distribution of research interest and resource allocation to the individual tasks in media bias. A unified benchmark encourages the development of more robust systems and shifts the current paradigm in media bias detection evaluation towards solutions that tackle not one but multiple media bias types simultaneously.

Timo Spinde, Jan-David Krieger, Terry Ruas et al.

Media has a substantial impact on the public perception of events. A one-sided or polarizing perspective on any topic is usually described as media bias. One of the ways how bias in news articles can be introduced is by altering word choice. Biased word choices are not always obvious, nor do they exhibit high context-dependency. Hence, detecting bias is often difficult. We propose a Transformer-based deep learning architecture trained via Multi-Task Learning using six bias-related data sets to tackle the media bias detection problem. Our best-performing implementation achieves a macro $F_{1}$ of 0.776, a performance boost of 3\% compared to our baseline, outperforming existing methods. Our results indicate Multi-Task Learning as a promising alternative to improve existing baseline models in identifying slanted reporting.

Przemyslaw Joniak, Akiko Aizawa

Language model debiasing has emerged as an important field of study in the NLP community. Numerous debiasing techniques were proposed, but bias ablation remains an unaddressed issue. We demonstrate a novel framework for inspecting bias in pre-trained transformer-based language models via movement pruning. Given a model and a debiasing objective, our framework finds a subset of the model containing less bias than the original model. We implement our framework by pruning the model while fine-tuning it on the debiasing objective. Optimized are only the pruning scores - parameters coupled with the model's weights that act as gates. We experiment with pruning attention heads, an important building block of transformers: we prune square blocks, as well as establish a new way of pruning the entire heads. Lastly, we demonstrate the usage of our framework using gender bias, and based on our findings, we propose an improvement to an existing debiasing method. Additionally, we re-discover a bias-performance trade-off: the better the model performs, the more bias it contains.

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.

Junfeng Jiang, Jiahao Huang, Akiko Aizawa

Recent developments in Japanese large language models (LLMs) primarily focus on general domains, with fewer advancements in Japanese biomedical LLMs. One obstacle is the absence of a comprehensive, large-scale benchmark for comparison. Furthermore, the resources for evaluating Japanese biomedical LLMs are insufficient. To advance this field, we propose a new benchmark including eight LLMs across four categories and 20 Japanese biomedical datasets across five tasks. Experimental results indicate that: (1) LLMs with a better understanding of Japanese and richer biomedical knowledge achieve better performance in Japanese biomedical tasks, (2) LLMs that are not mainly designed for Japanese biomedical domains can still perform unexpectedly well, and (3) there is still much room for improving the existing LLMs in certain Japanese biomedical tasks. Moreover, we offer insights that could further enhance development in this field. Our evaluation tools tailored to our benchmark as well as the datasets are publicly available in https://huggingface.co/datasets/Coldog2333/JMedBench to facilitate future research.

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.

Moritz Schubotz, Ankit Satpute, Andre Greiner-Petter et al.

Small to medium-scale data science experiments often rely on research software developed ad-hoc by individual scientists or small teams. Often there is no time to make the research software fast, reusable, and open access. The consequence is twofold. First, subsequent researchers must spend significant work hours building upon the proposed hypotheses or experimental framework. In the worst case, others cannot reproduce the experiment and reuse the findings for subsequent research. Second, suppose the ad-hoc research software fails during often long-running computationally expensive experiments. In that case, the overall effort to iteratively improve the software and rerun the experiments creates significant time pressure on the researchers. We suggest making caching an integral part of the research software development process, even before the first line of code is written. This article outlines caching recommendations for developing research software in data science projects. Our recommendations provide a perspective to circumvent common problems such as propriety dependence, speed, etc. At the same time, caching contributes to the reproducibility of experiments in the open science workflow. Concerning the four guiding principles, i.e., Findability, Accessibility, Interoperability, and Reusability (FAIR), we foresee that including the proposed recommendation in a research software development will make the data related to that software FAIRer for both machines and humans. We exhibit the usefulness of some of the proposed recommendations on our recently completed research software project in mathematical information retrieval.

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.

Taichi Iki, Akiko Aizawa

Pre-trained Transformers are good foundations for unified multi-task models owing to their task-agnostic representation. Pre-trained Transformers are often combined with text-to-text framework to execute multiple tasks by a single model. Performing a task through a graphical user interface (GUI) is another candidate to accommodate various tasks, including multi-step tasks with vision and language inputs. However, few papers combine pre-trained Transformers with performing through GUI. To fill this gap, we explore a framework in which a model performs a task by manipulating the GUI implemented with web pages in multiple steps. We develop task pages with and without page transitions and propose a BERT extension for the framework. We jointly trained our BERT extension with those task pages, and made the following observations. (1) The model learned to use both task pages with and without page transition. (2) In four out of five tasks without page transitions, the model performs greater than 75% of the performance of the original BERT, which does not use browsers. (3) The model did not generalize effectively on unseen tasks. These results suggest that we can fine-tune BERTs to multi-step tasks through GUIs, and there is room for improvement in their generalizability. Code will be available online.

Yihua Zhu, Qianying Liu, Fei Cheng et al.

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.

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.

Yihua Zhu, Qianying Liu, Jiaxin Wang et al.

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.

Ankit Satpute, Noah Giessing, Andre Greiner-Petter et al.

Large Language Models (LLMs) have demonstrated exceptional capabilities in various natural language tasks, often achieving performances that surpass those of humans. Despite these advancements, the domain of mathematics presents a distinctive challenge, primarily due to its specialized structure and the precision it demands. In this study, we adopted a two-step approach for investigating the proficiency of LLMs in answering mathematical questions. First, we employ the most effective LLMs, as identified by their performance on math question-answer benchmarks, to generate answers to 78 questions from the Math Stack Exchange (MSE). Second, a case analysis is conducted on the LLM that showed the highest performance, focusing on the quality and accuracy of its answers through manual evaluation. We found that GPT-4 performs best (nDCG of 0.48 and P@10 of 0.37) amongst existing LLMs fine-tuned for answering mathematics questions and outperforms the current best approach on ArqMATH3 Task1, considering P@10. Our Case analysis indicates that while the GPT-4 can generate relevant responses in certain instances, it does not consistently answer all questions accurately. This paper explores the current limitations of LLMs in navigating complex mathematical problem-solving. Through case analysis, we shed light on the gaps in LLM capabilities within mathematics, thereby setting the stage for future research and advancements in AI-driven mathematical reasoning. We make our code and findings publicly available for research: \url{https://github.com/gipplab/LLM-Investig-MathStackExchange}

Kon Woo Kim, Jin-Dong Kim, Akiko Aizawa

While Large Language Models (LLMs) demonstrate remarkable performance on zero-shot annotation tasks, they often struggle with the specialized conventions of gold-standard benchmarks. We propose the systematic reuse and refinement of annotation guidelines as an alignment mechanism, introducing an iterative moderation framework that simulates the early phases of annotation projects. We evaluate three hypotheses: (1) the efficacy of guideline integration, (2) the advantage of reasoning optimized models, and (3) the viability of moderation under minimal supervision. Testing across biomedical NER tasks (NCBI Disease, BC5CDR, BioRED) with three LLM families (GPT, Gemini, DeepSeek), our results empirically confirm all three hypotheses. While the iterative moderation framework shows good potential in effectively refining guidelines, our analysis also reveals substantial room for improvement.

Florian Boudin, Akiko Aizawa

Adapting keyphrase generation models to new domains typically involves few-shot fine-tuning with in-domain labeled data. However, annotating documents with keyphrases is often prohibitively expensive and impractical, requiring expert annotators. This paper presents silk, an unsupervised method designed to address this issue by extracting silver-standard keyphrases from citation contexts to create synthetic labeled data for domain adaptation. Extensive experiments across three distinct domains demonstrate that our method yields high-quality synthetic samples, resulting in significant and consistent improvements in in-domain performance over strong baselines.

Xanh Ho, Yun-Ang Wu, Sunisth Kumar et al.

With the growing number of submitted scientific papers, there is an increasing demand for systems that can assist reviewers in evaluating research claims. Experimental results are a core component of scientific work, often presented in varying formats such as tables or charts. Understanding how robust current multimodal large language models (multimodal LLMs) are at verifying scientific claims across different evidence formats remains an important and underexplored challenge. In this paper, we design and conduct a series of experiments to assess the ability of multimodal LLMs to verify scientific claims using both tables and charts as evidence. To enable this evaluation, we adapt two existing datasets of scientific papers by incorporating annotations and structures necessary for a multimodal claim verification task. Using this adapted dataset, we evaluate 12 multimodal LLMs and find that current models perform better with table-based evidence while struggling with chart-based evidence. We further conduct human evaluations and observe that humans maintain strong performance across both formats, unlike the models. Our analysis also reveals that smaller multimodal LLMs (under 8B) show weak correlation in performance between table-based and chart-based tasks, indicating limited cross-modal generalization. These findings highlight a critical gap in current models' multimodal reasoning capabilities. We suggest that future multimodal LLMs should place greater emphasis on improving chart understanding to better support scientific claim verification.

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

Junfeng Jiang, Chengzhang Dong, Sadao Kurohashi et al.

Dialogue segmentation is a crucial task for dialogue systems allowing a better understanding of conversational texts. Despite recent progress in unsupervised dialogue segmentation methods, their performances are limited by the lack of explicit supervised signals for training. Furthermore, the precise definition of segmentation points in conversations still remains as a challenging problem, increasing the difficulty of collecting manual annotations. In this paper, we provide a feasible definition of dialogue segmentation points with the help of document-grounded dialogues and release a large-scale supervised dataset called SuperDialseg, containing 9,478 dialogues based on two prevalent document-grounded dialogue corpora, and also inherit their useful dialogue-related annotations. Moreover, we provide a benchmark including 18 models across five categories for the dialogue segmentation task with several proper evaluation metrics. Empirical studies show that supervised learning is extremely effective in in-domain datasets and models trained on SuperDialseg can achieve good generalization ability on out-of-domain data. Additionally, we also conducted human verification on the test set and the Kappa score confirmed the quality of our automatically constructed dataset. We believe our work is an important step forward in the field of dialogue segmentation. Our codes and data can be found from: https://github.com/Coldog2333/SuperDialseg.

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.

Florian Boudin, Ygor Gallina, Akiko Aizawa

Sequence-to-sequence models have lead to significant progress in keyphrase generation, but it remains unknown whether they are reliable enough to be beneficial for document retrieval. This study provides empirical evidence that such models can significantly improve retrieval performance, and introduces a new extrinsic evaluation framework that allows for a better understanding of the limitations of keyphrase generation models. Using this framework, we point out and discuss the difficulties encountered with supplementing documents with -- not present in text -- keyphrases, and generalizing models across domains. Our code is available at https://github.com/boudinfl/ir-using-kg

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.

Joeran Beel, Andrew Collins, Akiko Aizawa

Recommender systems in academia are not widely available. This may be in part due to the difficulty and cost of developing and maintaining recommender systems. Many operators of academic products such as digital libraries and reference managers avoid this effort, although a recommender system could provide significant benefits to their users. In this paper, we introduce Mr. DLib's "Recommendations as-a-Service" (RaaS) API that allows operators of academic products to easily integrate a scientific recommender system into their products. Mr. DLib generates recommendations for research articles but in the future, recommendations may include call for papers, grants, etc. Operators of academic products can request recommendations from Mr. DLib and display these recommendations to their users. Mr. DLib can be integrated in just a few hours or days; creating an equivalent recommender system from scratch would require several months for an academic operator. Mr. DLib has been used by GESIS Sowiport and by the reference manager JabRef. Mr. DLib is open source and its goal is to facilitate the application of, and research on, scientific recommender systems. In this paper, we present the motivation for Mr. DLib, the architecture and details about the effectiveness. Mr. DLib has delivered 94m recommendations over a span of two years with an average click-through rate of 0.12%.

Junjun Jiang, Yi Yu, Suhua Tang et al.

Face hallucination is a technique that reconstruct high-resolution (HR) faces from low-resolution (LR) faces, by using the prior knowledge learned from HR/LR face pairs. Most state-of-the-arts leverage position-patch prior knowledge of human face to estimate the optimal representation coefficients for each image patch. However, they focus only the position information and usually ignore the context information of image patch. In addition, when they are confronted with misalignment or the Small Sample Size (SSS) problem, the hallucination performance is very poor. To this end, this study incorporates the contextual information of image patch and proposes a powerful and efficient context-patch based face hallucination approach, namely Thresholding Locality-constrained Representation and Reproducing learning (TLcR-RL). Under the context-patch based framework, we advance a thresholding based representation method to enhance the reconstruction accuracy and reduce the computational complexity. To further improve the performance of the proposed algorithm, we propose a promotion strategy called reproducing learning. By adding the estimated HR face to the training set, which can simulates the case that the HR version of the input LR face is present in the training set, thus iteratively enhancing the final hallucination result. Experiments demonstrate that the proposed TLcR-RL method achieves a substantial increase in the hallucinated results, both subjectively and objectively. Additionally, the proposed framework is more robust to face misalignment and the SSS problem, and its hallucinated HR face is still very good when the LR test face is from the real-world. The MATLAB source code is available at https://github.com/junjun-jiang/TLcR-RL

Ankit Satpute, André Greiner-Petter, Noah Gießing et al.

Content-based research paper recommendation (CbRPR) has seen advances in computer science and biomedicine, but remains unexplored for mathematics, where paper relatedness is more conceptual than explicit textual or citation-based similarity. Mathematics papers may be connected through shared proof techniques, logical implications, or natural generalizations, yet exhibit minimal textual or citation overlap, rendering existing CbRPR ineffective. To address this gap, we first conduct an expert-driven study characterizing mathematical recommendations, revealing that relevance is inherently \textit{aspect}-driven. Grounded in this insight, we introduce GoldRiM (small, expert-annotated) and SilverRiM (large, automatically derived), the first datasets for \textit{aspect}-aware CbRPR in mathematics. Recognizing that LLM embeddings of mathematical content alone yield suboptimal representation, we propose AchGNN, an \textit{aspect}-conditioned heterogeneous GNN that jointly models textual semantics, citation structure, and author lineage. Across GoldRiM and SilverRiM, AchGNN consistently outperforms prior \textit{aspect}-based CbRPR methods, achieving substantial gains across all evaluated \textit{aspects}. We conduct ablation studies to analyze the contributions of individual \textit{aspect} supervision, authorship lineage, and graph-structural signals to AchGNN's performance. To assess domain generality, we further evaluate AchGNN on the \textit{Papers with Code} dataset of machine learning publications, demonstrating that our \textit{aspect}-aware approach effectively transfers beyond mathematics. We deploy our system on the MaRDI platform to help mathematicians with recommendations and release datasets and code publicly for reproducibility.

Hung-Nghiep Tran, Akiko Aizawa, Atsuhiro Takasu

This paper reviews, analyzes, and proposes a new perspective on the bi-encoder architecture for neural search. While the bi-encoder architecture is widely used due to its simplicity and scalability at test time, it has some notable issues such as low performance on seen datasets and weak zero-shot performance on new datasets. In this paper, we analyze these issues and summarize two main critiques: the encoding information bottleneck problem and limitations of the basic assumption of embedding search. We then construct a thought experiment to logically analyze the encoding and searching operations and challenge the basic assumptions of embedding search. Building on these observations, we propose a new perspective on the bi-encoder architecture called the \textit{encoding--searching separation} perspective, which conceptually and practically separates the encoding and searching operations. This framework is applied to explain the root cause of existing issues and suggest mitigation strategies, potentially lowering training costs and improving retrieval performance. Finally, we discuss the broader implications of the ideas underlying this perspective, the new design surface it exposes, and potential research directions arising from it.

Tomas Horych, Christoph Mandl, Terry Ruas et al.

High annotation costs from hiring or crowdsourcing complicate the creation of large, high-quality datasets needed for training reliable text classifiers. Recent research suggests using Large Language Models (LLMs) to automate the annotation process, reducing these costs while maintaining data quality. LLMs have shown promising results in annotating downstream tasks like hate speech detection and political framing. Building on the success in these areas, this study investigates whether LLMs are viable for annotating the complex task of media bias detection and whether a downstream media bias classifier can be trained on such data. We create annolexical, the first large-scale dataset for media bias classification with over 48000 synthetically annotated examples. Our classifier, fine-tuned on this dataset, surpasses all of the annotator LLMs by 5-9 percent in Matthews Correlation Coefficient (MCC) and performs close to or outperforms the model trained on human-labeled data when evaluated on two media bias benchmark datasets (BABE and BASIL). This study demonstrates how our approach significantly reduces the cost of dataset creation in the media bias domain and, by extension, the development of classifiers, while our subsequent behavioral stress-testing reveals some of its current limitations and trade-offs.

Xanh Ho, Jiahao Huang, Florian Boudin et al.

Extractive reading comprehension question answering (QA) datasets are typically evaluated using Exact Match (EM) and F1-score, but these metrics often fail to fully capture model performance. With the success of large language models (LLMs), they have been employed in various tasks, including serving as judges (LLM-as-a-judge). In this paper, we reassess the performance of QA models using LLM-as-a-judge across four reading comprehension QA datasets. We examine different families of LLMs and various answer types to evaluate the effectiveness of LLM-as-a-judge in these tasks. Our results show that LLM-as-a-judge is highly correlated with human judgments and can replace traditional EM/F1 metrics. By using LLM-as-a-judge, the correlation with human judgments improves significantly, from 0.22 (EM) and 0.40 (F1-score) to 0.85. These findings confirm that EM and F1 metrics underestimate the true performance of the QA models. While LLM-as-a-judge is not perfect for more difficult answer types (e.g., job), it still outperforms EM/F1, and we observe no bias issues, such as self-preference, when the same model is used for both the QA and judgment tasks.

Xanh Ho, Anh Khoa Duong Nguyen, An Tuan Dao et al.

The number of Language Models (LMs) dedicated to processing scientific text is on the rise. Keeping pace with the rapid growth of scientific LMs (SciLMs) has become a daunting task for researchers. To date, no comprehensive surveys on SciLMs have been undertaken, leaving this issue unaddressed. Given the constant stream of new SciLMs, appraising the state-of-the-art and how they compare to each other remain largely unknown. This work fills that gap and provides a comprehensive review of SciLMs, including an extensive analysis of their effectiveness across different domains, tasks and datasets, and a discussion on the challenges that lie ahead.

Tomáš Horych, Martin Wessel, Jan Philip Wahle et al.

Media bias detection poses a complex, multifaceted problem traditionally tackled using single-task models and small in-domain datasets, consequently lacking generalizability. To address this, we introduce MAGPIE, the first large-scale multi-task pre-training approach explicitly tailored for media bias detection. To enable pre-training at scale, we present Large Bias Mixture (LBM), a compilation of 59 bias-related tasks. MAGPIE outperforms previous approaches in media bias detection on the Bias Annotation By Experts (BABE) dataset, with a relative improvement of 3.3% F1-score. MAGPIE also performs better than previous models on 5 out of 8 tasks in the Media Bias Identification Benchmark (MBIB). Using a RoBERTa encoder, MAGPIE needs only 15% of finetuning steps compared to single-task approaches. Our evaluation shows, for instance, that tasks like sentiment and emotionality boost all learning, all tasks enhance fake news detection, and scaling tasks leads to the best results. MAGPIE confirms that MTL is a promising approach for addressing media bias detection, enhancing the accuracy and efficiency of existing models. Furthermore, LBM is the first available resource collection focused on media bias MTL.

Léane Jourdan, Nicolas Hernandez, Richard Dufour et al.

Revision is a crucial step in scientific writing, where authors refine their work to improve clarity, structure, and academic quality. Existing approaches to automated writing assistance often focus on sentence-level revisions, which fail to capture the broader context needed for effective modification. In this paper, we explore the impact of shifting from sentence-level to paragraph-level scope for the task of scientific text revision. The paragraph level definition of the task allows for more meaningful changes, and is guided by detailed revision instructions rather than general ones. To support this task, we introduce ParaRev, the first dataset of revised scientific paragraphs with an evaluation subset manually annotated with revision instructions. Our experiments demonstrate that using detailed instructions significantly improves the quality of automated revisions compared to general approaches, no matter the model or the metric considered.

Xanh Ho, Sunisth Kumar, Yun-Ang Wu et al.

Scientific claim verification against tables typically requires predicting whether a claim is supported or refuted given a table. However, we argue that predicting the final label alone is insufficient: it reveals little about the model's reasoning and offers limited interpretability. To address this, we reframe table-text alignment as an explanation task, requiring models to identify the table cells essential for claim verification. We build a new dataset by extending the SciTab benchmark with human-annotated cell-level rationales. Annotators verify the claim label and highlight the minimal set of cells needed to support their decision. After the annotation process, we utilize the collected information and propose a taxonomy for handling ambiguous cases. Our experiments show that (i) incorporating table alignment information improves claim verification performance, and (ii) most LLMs, while often predicting correct labels, fail to recover human-aligned rationales, suggesting that their predictions do not stem from faithful reasoning.

Yihua Zhu, Qianying Liu, Akiko Aizawa et al.

Knowledge Base Question Answering (KBQA) aims to answer natural language questions using structured knowledge from KBs. While LLM-only approaches offer generalization, they suffer from outdated knowledge, hallucinations, and lack of transparency. Chain-based KG-RAG methods address these issues by incorporating external KBs, but are limited to simple chain-structured questions due to the absence of planning and logical structuring. Inspired by semantic parsing methods, we propose PDRR: a four-stage framework consisting of Predict, Decompose, Retrieve, and Reason. Our method first predicts the question type and decomposes the question into structured triples. Then retrieves relevant information from KBs and guides the LLM as an agent to reason over and complete the decomposed triples. Experimental results demonstrate that PDRR consistently outperforms existing methods across various LLM backbones and achieves superior performance on both chain-structured and non-chain complex questions.

Tomás Vergara-Browne, Álvaro Soto, Akiko Aizawa

We introduce eigenpruning, a method that removes singular values from weight matrices in an LLM to improve its performance in a particular task. This method is inspired by interpretability methods designed to automatically find subnetworks of a model which solve a specific task. In our tests, the pruned model outperforms the original model by a large margin, while only requiring minimal computation to prune the weight matrices. In the case of a small synthetic task in integer multiplication, the Phi-2 model can improve its accuracy in the test set from 13.75% to 97.50%. Interestingly, these results seem to indicate the existence of a computation path that can solve the task very effectively, but it was not being used by the original model. Finally, we publicly release our implementation.

Huy Quoc To, Ming Liu, Guangyan Huang et al.

Summarization for scientific text has shown significant benefits both for the research community and human society. Given the fact that the nature of scientific text is distinctive and the input of the multi-document summarization task is substantially long, the task requires sufficient embedding generation and text truncation without losing important information. To tackle these issues, in this paper, we propose SKT5SciSumm - a hybrid framework for multi-document scientific summarization (MDSS). We leverage the Sentence-Transformer version of Scientific Paper Embeddings using Citation-Informed Transformers (SPECTER) to encode and represent textual sentences, allowing for efficient extractive summarization using k-means clustering. We employ the T5 family of models to generate abstractive summaries using extracted sentences. SKT5SciSumm achieves state-of-the-art performance on the Multi-XScience dataset. Through extensive experiments and evaluation, we showcase the benefits of our model by using less complicated models to achieve remarkable results, thereby highlighting its potential in advancing the field of multi-document summarization for scientific text.

Juan Junqueras, Florian Boudin, May-Myo Zin et al.

Hate speech (HS) is a critical issue in online discourse, and one promising strategy to counter it is through the use of counter-narratives (CNs). Datasets linking HS with CNs are essential for advancing counterspeech research. However, even flagship resources like CONAN (Chung et al., 2019) annotate only a sparse subset of all possible HS-CN pairs, limiting evaluation. We introduce FC-CONAN (Fully Connected CONAN), the first dataset created by exhaustively considering all combinations of 45 English HS messages and 129 CNs. A two-stage annotation process involving nine annotators and four validators produces four partitions-Diamond, Gold, Silver, and Bronze-that balance reliability and scale. None of the labeled pairs overlap with CONAN, uncovering hundreds of previously unlabelled positives. FC-CONAN enables more faithful evaluation of counterspeech retrieval systems and facilitates detailed error analysis. The dataset is publicly available.

Xin Zhao, Naoki Yoshinaga, Yuma Tsuta et al.

Multilingual domain adaptation (ML-DA) is widely used to learn new domain knowledge across languages into large language models (LLMs). Although many methods have been proposed to improve domain adaptation, the mechanisms of multilingual knowledge acquisition, how domain knowledge is learned within a language and transferred across languages, remain underexplored. This gap leads to suboptimal performance, particularly in low-resource settings. This work examines the learning dynamics of LLMs during ML-DA. Because prior ML-DA studies often train and evaluate on datasets with mismatched knowledge coverage, we propose AdaXEval, an adaptive evaluation method that builds multiple-choice QA datasets from the same bilingual domain corpus used for training, thereby directly studying multilingual knowledge acquisition. Through continual training of LLMs with diverse data recipes, we track how LLMs acquire domain facts and pinpoint the mechanism behind the transformation process from domain training data to knowledge. Our experiments on a 13B English-Japanese bilingual LLM reveal that cross-lingual transfer remains challenging despite a high-quality bilingual corpus. The code has been released.

Kon Woo Kim, Rezarta Islamaj, Jin-Dong Kim et al.

This study investigates how existing annotation guidelines can be repurposed to instruct large language model (LLM) annotators for text annotation tasks. Traditional guidelines are written for human annotators who internalize training, while LLMs require explicit, structured instructions. We propose a moderation-oriented guideline repurposing method that transforms guidelines into clear directives for LLMs through an LLM moderation process. Using the NCBI Disease Corpus as a case study, our experiments show that repurposed guidelines can effectively guide LLM annotators, while revealing several practical challenges. The results highlight the potential of this workflow to support scalable and cost-effective refinement of annotation guidelines and automated annotation.