论文检索

Daehoon Gwak, Minseo Jung, Junwoo Park, Minho Park, ChaeHun Park, Junha Hyung, Jaegul Choo

Masked diffusion models (MDMs) offer a promising non-autoregressive alternative for large language modeling. Standard decoding methods for MDMs, such as confidence-based sampling, select tokens independently based on individual token confidences at each diffusion step. However, we observe that this independent token selection often results in generation orders resembling sequential autoregressive processes, limiting the advantages of non-autoregressive modeling. To mitigate this pheonomenon, we propose Reward-Weighted Sampling (RWS), a novel decoding strategy that leverages an external reward model to provide a principled global signal during the iterative diffusion process. Specifically, at each diffusion step, RWS evaluates the quality of the entire intermediate sequence and scales token logits accordingly, guiding token selection by integrating global sequence-level coherence. This method selectively increases the confidence of tokens that initially have lower scores, thereby promoting a more non-autoregressive generation order. Furthermore, we provide theoretical justification showing that reward-weighted logit scaling induces beneficial rank reversals in token selection and consistently improves expected reward. Experiments demonstrate that RWS significantly promotes non-autoregressive generation orders, leading to improvements across multiple evaluation metrics. These results highlight the effectiveness of integrating global signals in enhancing both the non-autoregressive properties and overall performance of MDMs.

Joonho Ko, Jinheon Baek, Sung Ju Hwang

Large language models (LLMs) have shown remarkable performance across a wide range of natural language tasks. However, a critical challenge remains in that they sometimes generate factually incorrect answers. To address this, while many previous work has focused on identifying errors in their generation and further refining them, they are slow in deployment since they are designed to verify the response from LLMs only after their entire generation (from the first to last tokens) is done. Further, we observe that once LLMs generate incorrect tokens early on, there is a higher likelihood that subsequent tokens will also be factually incorrect. To this end, in this work, we propose Streaming-VR (Streaming Verification and Refinement), a novel approach designed to enhance the efficiency of verification and refinement of LLM outputs. Specifically, the proposed Streaming-VR enables on-the-fly verification and correction of tokens as they are being generated, similar to a streaming process, ensuring that each subset of tokens is checked and refined in real-time by another LLM as the LLM constructs its response. Through comprehensive evaluations on multiple datasets, we demonstrate that our approach not only enhances the factual accuracy of LLMs, but also offers a more efficient solution compared to prior refinement methods.

Gleb Kuzmin, Petr Strepetov, Maksim Stankevich, Natalia Chudova, Artem Shelmanov, Ivan Smirnov

This paper compares the effectiveness of traditional machine learning methods, encoder-based models, and large language models (LLMs) on the task of detecting depression and anxiety. Five Russian-language datasets were considered, each differing in format and in the method used to define the target pathology class. We tested AutoML models based on linguistic features, several variations of encoder-based Transformers such as BERT, and state-of-the-art LLMs as pathology classification models. The results demonstrated that LLMs outperform traditional methods, particularly on noisy and small datasets where training examples vary significantly in text length and genre. However, psycholinguistic features and encoder-based models can achieve performance comparable to language models when trained on texts from individuals with clinically confirmed depression, highlighting their potential effectiveness in targeted clinical applications.

Chenxin An, Zhihui Xie, Xiaonan Li, Ming Zhong, Shansan Gong, Lei Li, Jun Zhang, Jingjing Xu, Lingpeng Kong

Reasoning models have demonstrated remarkable performance on complex tasks by generating long reasoning traces prior to producing final answers. However, previous research on long-context scaling in language models has generally focused on managing lengthy input prompts instead of producing long outputs. To leverage the strong long context understanding abilities of current models, we introduce Understanding-to-Reasoning Transition (URT) fine-tuning, a sequence-level curriculum learning framework that gradually shifts a model’s focus from interpreting long chain-of-thoughts to generating them. By incorporating partial reasoning steps in the input context, URT naturally exposes the model to diverse prompt lengths during training, preserving its performance on long-context comprehension while developing advanced reasoning capabilities. Experiments on rigorous reasoning benchmarks, including AIME24 and GPQA Diamond, reveal that our approach surpasses standard fine-tuning by over 10%, while maintaining robust performance on the understanding tasks in RULER.

Colin Hong, Xu Guo, Anand Chaanan Singh, Esha Choukse, Dmitrii Ustiugov

Recently, Test-Time Scaling (TTS) has gained increasing attention for improving LLM reasoning performance at test time without retraining the model. A notable TTS technique is Self-Consistency (SC), which generates multiple reasoning chains in parallel and selects the final answer via majority voting. While effective, the order-of-magnitude computational overhead limits its broad deployment. Prior attempts to accelerate SC mainly rely on model-based confidence scores or heuristics with limited empirical support. For the first time, we theoretically and empirically analyze the inefficiencies of SC and reveal actionable opportunities for improvement. Building on these insights, we propose Slim-SC, a step-wise pruning strategy that identifies and removes redundant chains using inter-chain similarity at the thought level.Experiments on three STEM reasoning datasets and two recent LLM architectures show that Slim-SC reduces inference latency and KVC usage by up to 45% and 26%, respectively, with R1-Distill, while maintaining or improving accuracy, thus offering a simple yet efficient TTS alternative for SC.

Mounica Maddela, Lingjue Xie, Daniel Preotiuc-Pietro, Mausam

Our goal is to assess how well current Text2SQL systems support SQL analysts in their primary work of performing complex analytics on specialized relational databases. Although several benchmarks evaluate Text2SQL models, the complexity of questions (and the output SQL queries) in most datasets is inherently limited – they do not focus on intents involving analytics and reasoning. In response, we present STARQA, the first public human-created dataset focused on complex analytical questions and answers (involving nested joins, time series analytics, statistical operations, and more) on three specialized-domain databases. In addition to standard Text2SQL baselines, we also evaluate a novel approach (Text2SQLCode) that decomposes the task through a combination of SQL and Python: SQL is responsible for data fetch, and Python more naturally performs reasoning. Our results demonstrate that both existing Text2SQL systems and our Text2SQLCode approach find STARQA questions quite challenging, even though Text2SQLCode achieves better performance on the more difficult questions. Further analyses assess the typical errors made by existing systems and charts a research path for pushing the capabilities of real-world systems.

Padakanti Srijith, Khushbu Pahwa, Radhika Mamidi, Bapi Raju Surampudi, Manish Gupta, Subba Reddy Oota

Although speech language models are expected to align well with brain language processing during speech comprehension, recent studies have found that they fail to capture brain-relevant semantics beyond low-level features. Surprisingly, text-based language models exhibit stronger alignment with brain language regions, as they better capture brain-relevant semantics. However, no prior work has examined the alignment effectiveness of text/speech representations from multimodal models. This raises several key questions: Can speech embeddings from such multimodal models capture brain-relevant semantics through cross-modal interactions? Which modality can take advantage of this synergistic multimodal understanding to improve alignment with brain language processing? Can text/speech representations from such multimodal models outperform unimodal models? To address these questions, we systematically analyze multiple multimodal models, extracting both text- and speech-based representations to assess their alignment with MEG brain recordings during naturalistic story listening. We find that text embeddings from both multimodal and unimodal models significantly outperform speech embeddings from these models. Specifically, multimodal text embeddings exhibit a peak around 200 ms, suggesting that they benefit from speech embeddings, with heightened activity during this time period. However, speech embeddings from these multimodal models still show a similar alignment compared to their unimodal counterparts, suggesting that they do not gain meaningful semantic benefits over text-based representations. These results highlight an asymmetry in cross-modal knowledge transfer, where the text modality benefits more from speech information, but not vice versa.

Kyuyoung Kim, Jinwoo Shin, Jaehyung Kim

Personalization in language models aims to tailor model behavior to individual users or user groups. Prompt-based methods incorporate user preferences into queries, while training-based methods encode them into model parameters. Model merging has also been explored for personalization under limited data. However, existing methods often fail to directly optimize task-specific utility and lack explicit mechanisms for privacy preservation. To address the limitations, we propose Privacy-Preserving Model Merging via Evolutionary Algorithms (PriME), a novel personalization approach that employs gradient-free methods to directly optimize utility while reducing privacy risks. By integrating privacy preservation into the optimization objective, PriME creates personalized modules that effectively capture target user preferences while minimizing privacy risks for data-sharing users. Experiments on the LaMP benchmark show that PriME consistently outperforms a range of baselines, achieving up to a 45% improvement in task performance. Further analysis demonstrates that PriME achieves a superior privacy-utility trade-off compared to a prior state-of-the-art, with enhanced robustness to membership inference attacks and greater utility in capturing user preferences.

Karim Ghonim, Pere-Lluís Huguet Cabot, Riccardo Orlando, Roberto Navigli

Entity Linking and Entity Disambiguation systems aim to link entity mentions to their corresponding entries, typically represented by descriptions within a predefined, static knowledge base. Current models assume that these knowledge bases are complete and up-to-date, rendering them incapable of handling entities not yet included therein. However, in an ever-evolving world, new entities emerge regularly, making these static resources insufficient for practical applications. To address this limitation, we introduce RAED, a model that retrieves external knowledge to improve factual grounding in entity descriptions. Using sources such as Wikipedia, RAED effectively disambiguates entities and bases their descriptions on factual information, reducing the dependence on parametric knowledge. Our experiments show that retrieval not only enhances overall description quality metrics, but also reduces hallucinations. Moreover, despite not relying on fixed entity inventories, RAED outperforms systems that require predefined candidate sets at inference time on Entity Disambiguation. Finally, we show that descriptions generated by RAED provide useful entity representations for downstream Entity Linking models, leading to improved performance in the extremely challenging Emerging Entity Linking task.

Karim Ghonim, Andrei Stefan Bejgu, Alberte Fernández-Castro, Roberto Navigli

Vision-language Models (VLMs), such as CLIP and SigLIP, have become the de facto standard for multimodal tasks, serving as essential building blocks for recent Multimodal Large Language Models, including LLaVA and PaliGemma. However, current evaluations for VLMs remain heavily anchored to ImageNet. In this paper, we question whether ImageNet’s coverage is still sufficiently challenging for modern VLMs, and investigate the impact of adding novel and varied concept categories, i.e., semantically grouped fine-grained synsets. To this end, we introduce Concept-pedia, a novel, large-scale, semantically-annotated multimodal resource covering more than 165,000 concepts. Leveraging a language-agnostic, automatic annotation pipeline grounded in Wikipedia, Concept-pedia expands the range of visual concepts, including diverse abstract categories. Building on Concept-pedia, we also present a manually-curated Visual Concept Recognition evaluation benchmark, Concept-10k, that spans thousands of concepts across a wide range of categories. Our experiments show that current models, although excelling on ImageNet, struggle with Concept-10k. Not only do these findings highlight a persistent bias toward ImageNet-centric concepts, but they also underscore the urgent need for more representative benchmarks. By offering a broader and semantically richer testbed, Concept-10k aims to support the development of multimodal systems that better generalize to the complexities of real-world visual concepts.

Sherrie Shen, Weixuan Wang, Alexandra Birch

The faithful transfer of contextually-embedded meaning continues to challenge contemporary machine translation (MT), particularly in the rendering of culture-bound terms—expressions or concepts rooted in specific languages or cultures, resisting direct linguistic transfer. Existing computational approaches to explicitating these terms have focused exclusively on in-text solutions, overlooking paratextual apparatus in the footnotes and endnotes employed by professional translators. In this paper, we formalize Genette’s (1987) theory of paratexts from literary and translation studies to introduce the task of paratextual explicitation for MT. We construct a dataset of 560 expert-aligned paratexts from four English translations of the classical Chinese short story collection Liaozhai and evaluate LLMs with and without reasoning traces on choice and content of explicitation. Experiments across intrinsic prompting and agentic retrieval methods establish the difficulty of this task, with human evaluation showing that LLM-generated paratexts improve audience comprehension, though remain considerably less effective than translator-authored ones. Beyond model performance, statistical analysis reveals that even professional translators vary widely in their use of paratexts, suggesting that cultural mediation is inherently open-ended rather than prescriptive. Our findings demonstrate the potential of paratextual explicitation in advancing MT beyond linguistic equivalence, with promising extensions to monolingual explanation and personalized adaptation.

Wenxuan Liu, Zixuan Li, Long Bai, Yuxin Zuo, Daozhu Xu, Xiaolong Jin, Jiafeng Guo, Xueqi Cheng

Developing a general-purpose system that can extract events with massive types is a long-standing target in Event Extraction (EE). In doing so, the basic challenge comes from the absence of an efficient and effective annotation framework to construct the corresponding datasets. In this paper, we propose an LLM-based collaborative annotation framework. Through collaboration among multiple LLMs and a subsequent voting process, it refines annotations of triggers from distant supervision and then carries out argument annotation. Finally, we create EEMT, the largest EE dataset to date, featuring over **200,000** samples, **3,465** event types, and **6,297** role types. Evaluation on human-annotated test set demonstrates that the proposed framework achieves the F1 scores of **90.1%** and **85.3%** for event detection and argument extraction, strongly validating its effectiveness. Besides, to alleviate the excessively long prompts caused by massive types, we propose an LLM-based Partitioning method for EE called LLM-PEE. It first recalls candidate event types and then splits them into multiple partitions for LLMs to extract. After fine-tuning on the EEMT training set, the distilled LLM-PEE with 7B parameters outperforms state-of-the-art methods by **5.4%** and **6.1%** in event detection and argument extraction. Besides, it also surpasses mainstream LLMs by **12.9%** on the unseen datasets, which strongly demonstrates the event diversity of the EEMT dataset and the generalization capabilities of the LLM-PEE method.

Murathan Kurfali, Robert Östling

Large Language Models (LLMs) have demonstrated an impressive ability to retrieve and summarize complex information, but their reliability in conflicting contexts remains poorly understood. We introduce an adversarial extension of the Needle-in-a-Haystack framework in which three mutually exclusive “needles” are embedded within long documents. By systematically manipulating factors such as position, repetition, layout, and domain relevance, we evaluate how LLMs handle contradictions. We find that models almost always fail to signal uncertainty and instead confidently select a single answer, exhibiting strong and consistent biases toward repetition, recency, and particular surface forms. We further analyze whether these patterns persist across model families and sizes, and we evaluate both probability-based and generation-based retrieval. Our framework highlights critical limitations in the robustness of current LLMs—including commercial systems—to contradiction. These limitations reveal potential shortcomings in RAG systems’ ability to handle noisy or manipulated inputs and exposes risks for deployment in high-stakes applications.

Alaa Aljabari, Mohammed Khalilia, Mustafa Jarrar

Relation extraction (RE) is a core task in natural language processing, crucial for semantic understanding, knowledge graph construction, and enhancing downstream applications. Existing work on Arabic RE remains limited due to the language’s rich morphology and syntactic complexity, and the lack of large, high-quality datasets. In this paper, we present \mathrm{Wojood^{Relations}}, the largest and most diverse Arabic RE corpus to date, containing over 33K sentences (\sim550K tokens) annotated with \sim15K relation triples across 40 relation types. The corpus is built on top of Wojood NER dataset with manual relation annotations carried out by expert annotators, achieving a Cohen’s \kappa of 0.92, indicating high reliability. In addition, we propose two methods: NLI-RE, which formulates RE as a binary natural language inference problem using relation-aware templates, and GPT-Joint, a few-shot LLM framework for joint entity and RE via relation-aware retrieval. Finally, we benchmark the dataset using both supervised models and in-context learning with LLMs. Supervised models achieve 92.89% F1 for RE, while LLMs obtain 72.73% F1 for joint entity and RE. These results establish strong baselines, highlight key challenges, and provide a foundation for advancing Arabic RE research.

Maggie Mi, Aline Villavicencio, Nafise Sadat Moosavi

Language models often struggle with idiomatic, figurative, or context-sensitive inputs, not because they produce flawed outputs, but because they misinterpret the input from the outset. We propose an input-only method for anticipating such failures using token-level likelihood features inspired by surprisal and the Uniform Information Density hypothesis. These features capture localized uncertainty in input comprehension and outperform standard baselines across five linguistically challenging datasets. We show that span-localized features improve error detection for larger models, while smaller models benefit from global patterns. Our method requires no access to outputs or hidden activations, offering a lightweight and generalizable approach to pre-generation error prediction.

Pawitsapak Akarajaradwong, Pirat Pothavorn, Chompakorn Chaksangchaichot, Panuthep Tasawong, Thitiwat Nopparatbundit, Keerakiat Pratai, Sarana Nutanong

Large language models (LLMs) show promise in legal question answering (QA), yet Thai legal QA systems face challenges due to limited data and complex legal structures. We introduce NitiBench, a novel benchmark featuring two datasets: (1) NitiBench-CCL, covering Thai financial laws, and (2) NitiBench-Tax, containing Thailand’s official tax rulings. Our benchmark also consists of specialized evaluation metrics suited for Thai legal QA. We evaluate retrieval-augmented generation (RAG) and long-context LLM (LCLM) approaches across three key dimensions: (1) the benefits of domain-specific techniques like hierarchy-aware chunking and cross-referencing, (2) comparative performance of RAG components, e.g., retrievers and LLMs, and (3) the potential of long-context LLMs to replace traditional RAG systems. Our results reveal that domain-specific components slightly improve over naive methods. At the same time, existing retrieval models still struggle with complex legal queries, and long-context LLMs have limitations in consistent legal reasoning. Our study highlights current limitations in Thai legal NLP and lays a foundation for future research in this emerging domain.

Jeongyeon Hwang, Junyoung Park, Hyejin Park, Dongwoo Kim, Sangdon Park, Jungseul Ok

Retrieval-Augmented Generation (RAG) is an effective approach to enhance the factual accuracy of large language models (LLMs) by retrieving information from external databases, which are typically composed of diverse sources, to supplement the limited internal knowledge of LLMs. However, the standard RAG often risks retrieving incorrect information, as it relies solely on relevance between a query and a document, overlooking the heterogeneous reliability of these sources. To address this issue, we propose Reliability-Aware RAG (RA-RAG), a new multi-source RAG framework that estimates the reliability of sources and leverages this information to prioritize highly reliable and relevant documents, ensuring more robust and accurate response generation. Specifically, RA-RAG first estimates source reliability by cross-checking information across multiple sources. It then retrieves documents from the top-\kappa reliable and relevant sources and aggregates their information using weighted majority voting (WMV), where the selective retrieval ensures scalability while not compromising the performance. Comprehensive experiments show that RA-RAG consistently outperforms baselines in scenarios with heterogeneous source reliability while scaling efficiently as the number of sources increases. Furthermore, we demonstrate the ability of RA-RAG to estimate real-world sources’ reliability, highlighting its practical applicability. Our code and data are available at RA-RAG.

Giuliano Martinelli, Bruno Gatti, Roberto Navigli

Current coreference resolution systems are typically tailored for short- or medium-sized texts and struggle to scale to very long documents due to architectural limitations and implied memory costs.However, a few available solutions can be applied by inputting documents split into smaller windows. This is inherently similar to what happens in the cross-document setting, in which systems infer coreference relations between mentions that are found in separate documents.In this paper, we unify these two challenging settings under the general framework of cross-context coreference, and introduce xCoRe, a new unified approach designed to efficiently handle short-, long-, and cross-document coreference resolution.xCoRe adopts a three-step pipeline that first identifies mentions, then creates clusters within individual contexts, and finally merges clusters across contexts.In our experiments, we show that our formulation enables joint training on shared long- and cross-document resources, increasing data availability and particularly benefiting the challenging cross-document task.Our model achieves new state-of-the-art results on cross-document benchmarks and strong performance on long-document data, while retaining top-tier results on traditional datasets, positioning it as a robust, versatile solution that can be applied across all end-to-end coreference settings.We release our models and code at http://github.com/sapienzanlp/xcore.

Numaan Naeem, Abdellah El Mekki, Muhammad Abdul-Mageed

Large language models (LLMs) are transforming education by answering questions, explaining complex concepts, and generating content across a wide range of subjects. Despite strong performance on academic benchmarks, they often fail to tailor responses to students’ grade levels. This is a critical need in K-12 education, where age-appropriate vocabulary and explanation are essential for effective learning. Existing models frequently produce outputs that are too advanced or vague for younger learners, and there are no standardized benchmarks to evaluate their ability to adjust across cognitive and developmental stages. To address this gap, we introduce EduAdapt, a benchmark of nearly 48k grade-labeled QA pairs across nine science subjects, spanning Grades 1-12 and grouped into four grade levels. We evaluate a diverse set of open-source LLMs on EduAdapt and find that while larger models generally perform better, they still struggle with generating suitable responses for early-grade students (Grades 1-5). Our work presents the first dataset and evaluation framework for assessing grade-level adaptability in LLMs, aiming to foster more developmentally aligned educational AI systems through better training and prompting strategies. EduAdapt code and datasets are publicly available at https://github.com/NaumanNaeem/EduAdapt.

Shweta Verma, Abhinav Anand, Mira Mezini

Although transformers dominate many code-specific tasks, they have significant limitations. This paper explores State Space Models (SSMs) as a promising alternative for code understanding tasks such as retrieval, classification, and clone detection. We introduce CodeSSM, the first SSM-based model trained on code corpora to assess its effectiveness. Our results demonstrate that SSMs are more sample-efficient and can extrapolate to longer contexts beyond the pretraining length. Extensive experiments show that SSMs offer a viable alternative to transformers, addressing several their limitations. Additionally, CodeSSM reduces memory usage by up to 64% compared to transformers at a context length of 2048, with greater savings as context length grows.The code is available [here](https://github.com/abx04/CodeSSM).