Long reasoning models often struggle in multilingual settings: they tend to reason in English for non-English questions; when constrained to reasoning in the question language, accuracies drop substantially. The struggle is caused by the limited abilities for both multilingual question understanding and multilingual reasoning. To address both problems, we propose TRIT (Translation-Reasoning Integrated Training), a self-improving framework that integrates the training of translation into multilingual reasoning. Without external feedback or additional multilingual data, our method jointly enhances multilingual question understanding and response generation. On MMATH, our method outperforms multiple baselines by an average of 7 percentage points, improving both answer correctness and language consistency. Further analysis reveals that integrating translation training improves cross-lingual question alignment by over 10 percentage points and enhances translation quality for both mathematical questions and general-domain text, with gains up to 8.4 COMET points on FLORES-200.
Aidana Aidynkyzy, Oğuz Dikenelli, Oylum Alatlı
et al.
The scarcity of annotated datasets for clinical information extraction in non-English languages hinders the evaluation of large language model (LLM)-based methods developed primarily in English. In this study, we present the first comprehensive bilingual evaluation of LLMs for the clinical Relation Extraction (RE) task in both English and Turkish. To facilitate this evaluation, we introduce the first English-Turkish parallel clinical RE dataset, derived and carefully curated from the 2010 i2b2/VA relation classification corpus. We systematically assess a diverse set of prompting strategies, including multiple in-context learning (ICL) and Chain-of-Thought (CoT) approaches, and compare their performance to fine-tuned baselines such as PURE. Furthermore, we propose Relation-Aware Retrieval (RAR), a novel in-context example selection method based on contrastive learning, that is specifically designed to capture both sentence-level and relation-level semantics. Our results show that prompting-based LLM approaches consistently outperform traditional fine-tuned models. Moreover, evaluations for English performed better than their Turkish counterparts across all evaluated LLMs and prompting techniques. Among ICL methods, RAR achieves the highest performance, with Gemini 1.5 Flash reaching a micro-F1 score of 0.906 in English and 0.888 in Turkish. Performance further improves to 0.918 F1 in English when RAR is combined with a structured reasoning prompt using the DeepSeek-V3 model. These findings highlight the importance of high-quality demonstration retrieval and underscore the potential of advanced retrieval and prompting techniques to bridge resource gaps in clinical natural language processing.
Analysts require attribution, as nothing can be reported without knowing the source of the information. In this paper, we will focus on automatic methods for attribution, linking each sentence in the summary to a portion of the source text, which may be in one or more documents. We explore using a hybrid summarization, i.e., an automatic paraphrase of an extractive summary, to ease attribution. We also use a custom topology to identify the proportion of different categories of attribution-related errors.
Irina Tolstykh, Aleksandra Tsybina, Sergey Yakubson
et al.
The widespread use of human-like text from Large Language Models (LLMs) necessitates the development of robust detection systems. However, progress is limited by a critical lack of suitable training data; existing datasets are often generated with outdated models, are predominantly in English, and fail to address the increasingly common scenario of mixed human-AI authorship. Crucially, while some datasets address mixed authorship, none provide the character-level annotations required for the precise localization of AI-generated segments within a text. To address these gaps, we introduce LLMTrace, a new large-scale, bilingual (English and Russian) corpus for AI-generated text detection. Constructed using a diverse range of modern proprietary and open-source LLMs, our dataset is designed to support two key tasks: traditional full-text binary classification (human vs. AI) and the novel task of AI-generated interval detection, facilitated by character-level annotations. We believe LLMTrace will serve as a vital resource for training and evaluating the next generation of more nuanced and practical AI detection models. The project page is available at \href{https://sweetdream779.github.io/LLMTrace-info/}{iitolstykh/LLMTrace}.
Matthew W. Kenaston, Umair Ayub, Mihir Parmar
et al.
Despite high performance on clinical benchmarks, large language models may reach correct conclusions through faulty reasoning, a failure mode with safety implications for oncology decision support that is not captured by accuracy-based evaluation. In this two-cohort retrospective study, we developed a hierarchical taxonomy of reasoning errors from GPT-4 chain-of-thought responses to real oncology notes and tested its clinical relevance. Using breast and pancreatic cancer notes from the CORAL dataset, we annotated 600 reasoning traces to define a three-tier taxonomy mapping computational failures to cognitive bias frameworks. We validated the taxonomy on 822 responses from prostate cancer consult notes spanning localized through metastatic disease, simulating extraction, analysis, and clinical recommendation tasks. Reasoning errors occurred in 23 percent of interpretations and dominated overall errors, with confirmation bias and anchoring bias most common. Reasoning failures were associated with guideline-discordant and potentially harmful recommendations, particularly in advanced disease management. Automated evaluators using state-of-the-art language models detected error presence but could not reliably classify subtypes. These findings show that large language models may provide fluent but clinically unsafe recommendations when reasoning is flawed. The taxonomy provides a generalizable framework for evaluating and improving reasoning fidelity before clinical deployment.
We present HADA (Human-AI Agent Decision Alignment), a protocol- and framework agnostic reference architecture that keeps both large language model (LLM) agents and legacy algorithms aligned with organizational targets and values. HADA wraps any algorithm or LLM in role-specific stakeholder agents -- business, data-science, audit, ethics, and customer -- each exposing conversational APIs so that technical and non-technical actors can query, steer, audit, or contest every decision across strategic, tactical, and real-time horizons. Alignment objectives, KPIs, and value constraints are expressed in natural language and are continuously propagated, logged, and versioned while thousands of heterogeneous agents run on different orchestration stacks. A cloud-native proof of concept packages a production credit-scoring model (getLoanDecision) and deploys it on Docker/Kubernetes/Python; five scripted retail-bank scenarios show how target changes, parameter tweaks, explanation requests, and ethics triggers flow end to end through the architecture. Evaluation followed the Design-Science Research Methodology. Walkthrough observation and log inspection demonstrated complete coverage of six predefined objectives: every role could invoke conversational control, trace KPIs and value constraints, detect and mitigate ZIP-code bias, and reproduce full decision lineage, independent of the underlying LLM or agent library. Contributions: (1) an open-source HADA architecture, (2) a mid-range design theory for human-AI alignment in multi-agent systems, and (3) empirical evidence that framework-agnostic, protocol-compliant stakeholder agents improve accuracy, transparency, and ethical compliance in real-world decision pipelines.
Hailay Teklehaymanot, Dren Fazlija, Niloy Ganguly
et al.
The absence of explicitly tailored, accessible annotated datasets for educational purposes presents a notable obstacle for NLP tasks in languages with limited resources.This study initially explores the feasibility of using machine translation (MT) to convert an existing dataset into a Tigrinya dataset in SQuAD format. As a result, we present TIGQA, an expert annotated educational dataset consisting of 2.68K question-answer pairs covering 122 diverse topics such as climate, water, and traffic. These pairs are from 537 context paragraphs in publicly accessible Tigrinya and Biology books. Through comprehensive analyses, we demonstrate that the TIGQA dataset requires skills beyond simple word matching, requiring both single-sentence and multiple-sentence inference abilities. We conduct experiments using state-of-the art MRC methods, marking the first exploration of such models on TIGQA. Additionally, we estimate human performance on the dataset and juxtapose it with the results obtained from pretrained models.The notable disparities between human performance and best model performance underscore the potential for further enhancements to TIGQA through continued research. Our dataset is freely accessible via the provided link to encourage the research community to address the challenges in the Tigrinya MRC.
ChatGPT explores a strategic blueprint of question answering (QA) in delivering medical diagnosis, treatment recommendations, and other healthcare support. This is achieved through the increasing incorporation of medical domain data via natural language processing (NLP) and multimodal paradigms. By transitioning the distribution of text, images, videos, and other modalities from the general domain to the medical domain, these techniques have expedited the progress of medical domain question answering (MDQA). They bridge the gap between human natural language and sophisticated medical domain knowledge or expert manual annotations, handling large-scale, diverse, unbalanced, or even unlabeled data analysis scenarios in medical contexts. Central to our focus is the utilizing of language models and multimodal paradigms for medical question answering, aiming to guide the research community in selecting appropriate mechanisms for their specific medical research requirements. Specialized tasks such as unimodal-related question answering, reading comprehension, reasoning, diagnosis, relation extraction, probability modeling, and others, as well as multimodal-related tasks like vision question answering, image caption, cross-modal retrieval, report summarization, and generation, are discussed in detail. Each section delves into the intricate specifics of the respective method under consideration. This paper highlights the structures and advancements of medical domain explorations against general domain methods, emphasizing their applications across different tasks and datasets. It also outlines current challenges and opportunities for future medical domain research, paving the way for continued innovation and application in this rapidly evolving field.
Bioinformatics has witnessed a paradigm shift with the increasing integration of artificial intelligence (AI), particularly through the adoption of foundation models (FMs). These AI techniques have rapidly advanced, addressing historical challenges in bioinformatics such as the scarcity of annotated data and the presence of data noise. FMs are particularly adept at handling large-scale, unlabeled data, a common scenario in biological contexts due to the time-consuming and costly nature of experimentally determining labeled data. This characteristic has allowed FMs to excel and achieve notable results in various downstream validation tasks, demonstrating their ability to represent diverse biological entities effectively. Undoubtedly, FMs have ushered in a new era in computational biology, especially in the realm of deep learning. The primary goal of this survey is to conduct a systematic investigation and summary of FMs in bioinformatics, tracing their evolution, current research status, and the methodologies employed. Central to our focus is the application of FMs to specific biological problems, aiming to guide the research community in choosing appropriate FMs for their research needs. We delve into the specifics of the problem at hand including sequence analysis, structure prediction, function annotation, and multimodal integration, comparing the structures and advancements against traditional methods. Furthermore, the review analyses challenges and limitations faced by FMs in biology, such as data noise, model explainability, and potential biases. Finally, we outline potential development paths and strategies for FMs in future biological research, setting the stage for continued innovation and application in this rapidly evolving field. This comprehensive review serves not only as an academic resource but also as a roadmap for future explorations and applications of FMs in biology.
Large language models (LLMs) now exhibit near human-level performance in various tasks, but their performance drops drastically after a handful of high-resource languages due to the imbalance in pre-training data. Inspired by the human process of second language acquisition, particularly code-switching$\unicode{x2014}$the practice of language alternation in a conversation$\unicode{x2014}$we propose code-switching curriculum learning (CSCL) to enhance cross-lingual transfer for LLMs. CSCL mimics the stages of human language learning by progressively training models with a curriculum consisting of 1) token-level code-switching, 2) sentence-level code-switching, and 3) monolingual corpora. Using Qwen 2 as our underlying model, we demonstrate the efficacy of the CSCL in improving language transfer to Korean, achieving significant performance gains compared to monolingual continual pre-training methods. Ablation studies reveal that both token- and sentence-level code-switching significantly enhance cross-lingual transfer and that curriculum learning amplifies these effects. We also extend our findings into various languages, including Japanese (high-resource) and Indonesian (low-resource), and using two additional models (Gemma 2 and Phi 3.5). We further show that CSCL mitigates spurious correlations between language resources and safety alignment, presenting a robust, efficient framework for more equitable language transfer in LLMs. We observe that CSCL is effective for low-resource settings where high-quality, monolingual corpora for language transfer are hardly available.
The lead-free perovskites Cs3 B 2 X 9 (B = Sb, Bi; X = Cl, Br, I) as the popular photoelectric materials have excellent optical properties with lower toxicity. In this study, we systematically investigate the stable monolayer Cs3 B 2 X 9 and bilayer vertical heterostructure Cs3 B 2 X 9/Cs3 B 2 ′ X 9 (B, B′ = Sb, Bi; X = Cl, Br, I) via first-principles simulations. By exploring the electrical structures and band edge positions, we find the band gap reduction and the band type transition in the heterostructure Cs3 B 2 X 9/Cs3 B 2 ′ X 9 due to the charge transfer between layers. Furthermore, the results of optical properties reveal light absorption from the visible light to UV region, especially monolayer Cs3Sb2I9 and heterostructure Cs3Sb2I9/Cs3Bi2I9, which have absorption peaks in the visible light region, leading to the possibility of photocatalytic water splitting. These results provide insights for more two-dimensional semiconductors applied in the optoelectronic and photocatalytic fields.
Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma
et al.
Automating the fact checking (FC) process relies on information obtained from external sources. In this work, we posit that it is crucial for FC models to make veracity predictions only when there is sufficient evidence and otherwise indicate when it is not enough. To this end, we are the first to study what information FC models consider sufficient by introducing a novel task and advancing it with three main contributions. First, we conduct an in-depth empirical analysis of the task with a new fluency-preserving method for omitting information from the evidence at the constituent and sentence level. We identify when models consider the remaining evidence (in)sufficient for FC, based on three trained models with different Transformer architectures and three FC datasets. Second, we ask annotators whether the omitted evidence was important for FC, resulting in a novel diagnostic dataset, SufficientFacts, for FC with omitted evidence. We find that models are least successful in detecting missing evidence when adverbial modifiers are omitted (21% accuracy), whereas it is easiest for omitted date modifiers (63% accuracy). Finally, we propose a novel data augmentation strategy for contrastive self-learning of missing evidence by employing the proposed omission method combined with tri-training. It improves performance for Evidence Sufficiency Prediction by up to 17.8 F1 score, which in turn improves FC performance by up to 2.6 F1 score.
This paper presents Z-Code++, a new pre-trained language model optimized for abstractive text summarization. The model extends the state of the art encoder-decoder model using three techniques. First, we use a two-phase pre-training process to improve model's performance on low-resource summarization tasks. The model is first pre-trained using text corpora for language understanding, and then is continually pre-trained on summarization corpora for grounded text generation. Second, we replace self-attention layers in the encoder with disentangled attention layers, where each word is represented using two vectors that encode its content and position, respectively. Third, we use fusion-in-encoder, a simple yet effective method of encoding long sequences in a hierarchical manner. Z-Code++ creates new state of the art on 9 out of 13 text summarization tasks across 5 languages. Our model is parameter-efficient in that it outperforms the 600x larger PaLM-540B on XSum, and the finetuned 200x larger GPT3-175B on SAMSum. In zero-shot and few-shot settings, our model substantially outperforms the competing models.
This article explores the sentiment dynamics present in narratives and their contribution to literary appreciation. Specifically, we investigate whether a certain type of sentiment development in a literary narrative correlates with its quality as perceived by a large number of readers. While we do not expect a story's sentiment arc to relate directly to readers' appreciation, we focus on its internal coherence as measured by its sentiment arc's level of fractality as a potential predictor of literary quality. To measure the arcs' fractality we use the Hurst exponent, a popular measure of fractal patterns that reflects the predictability or self-similarity of a time series. We apply this measure to the fairy tales of H.C. Andersen, using GoodReads' scores to approximate their level of appreciation. Based on our results we suggest that there might be an optimal balance between predictability and surprise in a sentiment arcs' structure that contributes to the perceived quality of a narrative text.
History of scholarship and learning. The humanities, Bibliography. Library science. Information resources
Developing a system to detect online offensive language is very important to the health and the security of online users. Studies have shown that cyberhate, online harassment and other misuses of technology are on the rise, particularly during the global Coronavirus pandemic in 2020. According to the latest report by the Anti-Defamation League (ADL), 35% of online users reported online harassment related to their identity-based characteristics, which is a 3% increase over 2019. Applying advanced techniques from the Natural Language Processing (NLP) field to support the development of an online hate-free community is a critical task for social justice. Transfer learning enhances the performance of the classifier by allowing the transfer of knowledge from one domain or one dataset to others that have not been seen before, thus, supporting the classifier to be more generalizable. In our study, we apply the principles of transfer learning cross multiple Arabic offensive language datasets to compare the effects on system performance. This study aims at investigating the effects of fine-tuning and training Bidirectional Encoder Representations from Transformers (BERT) model on multiple Arabic offensive language datasets individually and testing it using other datasets individually. Our experiment starts with a comparison among multiple BERT models to guide the selection of the main model that is used for our study. The study also investigates the effects of concatenating all datasets to be used for fine-tuning and training BERT model. Our results demonstrate the limited effects of transfer learning on the performance of the classifiers, particularly for highly dialectic comments.
This paper presents a new pre-trained language model, DeBERTaV3, which improves the original DeBERTa model by replacing mask language modeling (MLM) with replaced token detection (RTD), a more sample-efficient pre-training task. Our analysis shows that vanilla embedding sharing in ELECTRA hurts training efficiency and model performance. This is because the training losses of the discriminator and the generator pull token embeddings in different directions, creating the "tug-of-war" dynamics. We thus propose a new gradient-disentangled embedding sharing method that avoids the tug-of-war dynamics, improving both training efficiency and the quality of the pre-trained model. We have pre-trained DeBERTaV3 using the same settings as DeBERTa to demonstrate its exceptional performance on a wide range of downstream natural language understanding (NLU) tasks. Taking the GLUE benchmark with eight tasks as an example, the DeBERTaV3 Large model achieves a 91.37% average score, which is 1.37% over DeBERTa and 1.91% over ELECTRA, setting a new state-of-the-art (SOTA) among the models with a similar structure. Furthermore, we have pre-trained a multi-lingual model mDeBERTa and observed a larger improvement over strong baselines compared to English models. For example, the mDeBERTa Base achieves a 79.8% zero-shot cross-lingual accuracy on XNLI and a 3.6% improvement over XLM-R Base, creating a new SOTA on this benchmark. We have made our pre-trained models and inference code publicly available at https://github.com/microsoft/DeBERTa.
We describe a strategy for identifying the universe of research publications relevant to the application and development of artificial intelligence. The approach leverages the arXiv corpus of scientific preprints, in which authors choose subject tags for their papers from a set defined by editors. We compose a functional definition of AI relevance by learning these subjects from paper metadata, and then inferring the arXiv-subject labels of papers in larger corpora: Clarivate Web of Science, Digital Science Dimensions, and Microsoft Academic Graph. This yields predictive classification $F_1$ scores between .75 and .86 for Natural Language Processing (cs.CL), Computer Vision (cs.CV), and Robotics (cs.RO). For a single model that learns these and four other AI-relevant subjects (cs.AI, cs.LG, stat.ML, and cs.MA), we see precision of .83 and recall of .85. We evaluate the out-of-domain performance of our classifiers against other sources of topic information and predictions from alternative methods. We find that a supervised solution can generalize to identify publications that belong to the high-level fields of study represented on arXiv. This offers a method for identifying AI-relevant publications that updates at the pace of research output, without reliance on subject-matter experts for query development or labeling.
Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma
et al.
Recent developments in machine learning have introduced models that approach human performance at the cost of increased architectural complexity. Efforts to make the rationales behind the models' predictions transparent have inspired an abundance of new explainability techniques. Provided with an already trained model, they compute saliency scores for the words of an input instance. However, there exists no definitive guide on (i) how to choose such a technique given a particular application task and model architecture, and (ii) the benefits and drawbacks of using each such technique. In this paper, we develop a comprehensive list of diagnostic properties for evaluating existing explainability techniques. We then employ the proposed list to compare a set of diverse explainability techniques on downstream text classification tasks and neural network architectures. We also compare the saliency scores assigned by the explainability techniques with human annotations of salient input regions to find relations between a model's performance and the agreement of its rationales with human ones. Overall, we find that the gradient-based explanations perform best across tasks and model architectures, and we present further insights into the properties of the reviewed explainability techniques.
Recent progress in pre-trained neural language models has significantly improved the performance of many natural language processing (NLP) tasks. In this paper we propose a new model architecture DeBERTa (Decoding-enhanced BERT with disentangled attention) that improves the BERT and RoBERTa models using two novel techniques. The first is the disentangled attention mechanism, where each word is represented using two vectors that encode its content and position, respectively, and the attention weights among words are computed using disentangled matrices on their contents and relative positions, respectively. Second, an enhanced mask decoder is used to incorporate absolute positions in the decoding layer to predict the masked tokens in model pre-training. In addition, a new virtual adversarial training method is used for fine-tuning to improve models' generalization. We show that these techniques significantly improve the efficiency of model pre-training and the performance of both natural language understanding (NLU) and natural langauge generation (NLG) downstream tasks. Compared to RoBERTa-Large, a DeBERTa model trained on half of the training data performs consistently better on a wide range of NLP tasks, achieving improvements on MNLI by +0.9% (90.2% vs. 91.1%), on SQuAD v2.0 by +2.3% (88.4% vs. 90.7%) and RACE by +3.6% (83.2% vs. 86.8%). Notably, we scale up DeBERTa by training a larger version that consists of 48 Transform layers with 1.5 billion parameters. The significant performance boost makes the single DeBERTa model surpass the human performance on the SuperGLUE benchmark (Wang et al., 2019a) for the first time in terms of macro-average score (89.9 versus 89.8), and the ensemble DeBERTa model sits atop the SuperGLUE leaderboard as of January 6, 2021, out performing the human baseline by a decent margin (90.3 versus 89.8).