abstracts sequencelengths 2 2 | id_1 stringlengths 10 10 | id_2 stringlengths 10 10 | pair_id stringlengths 21 21 |
|---|---|---|---|
[
" Large language models (LLMs) acquire general linguistic knowledge from\nmassive-scale pretraining. However, pretraining data mainly comprised of\nweb-crawled texts contain undesirable social biases which can be perpetuated or\neven amplified by LLMs. In this study, we propose an efficient yet effective\nannotati... | 2504.14212 | 2504.00780 | 2504.14212_2504.00780 |
[
" The evaluation of vision-language models (VLMs) has mainly relied on\nEnglish-language benchmarks, leaving significant gaps in both multilingual and\nmulticultural coverage. While multilingual benchmarks have expanded, both in\nsize and languages, many rely on translations of English datasets, failing to\ncaptur... | 2504.07072 | 2502.06252 | 2504.07072_2502.06252 |
[
" Transformer tends to overallocate attention to irrelevant context. In this\nwork, we introduce Diff Transformer, which amplifies attention to the relevant\ncontext while canceling noise. Specifically, the differential attention\nmechanism calculates attention scores as the difference between two separate\nsoftma... | 2410.05258 | 2503.03920 | 2410.05258_2503.03920 |
[
" Structured pruning is a promising approach to create smaller, faster LLMs.\nHowever, existing methods typically rely on backward passes, which can inflate\nmemory requirements and compute costs. In this work we introduce Bonsai, a\ngradient-free structured pruning method that eliminates the need for\nbackpropaga... | 2402.05406 | 2104.02496 | 2402.05406_2104.02496 |
[
" While researchers often study message features like moral content in text,\nsuch as party manifestos and social media, their quantification remains a\nchallenge. Conventional human coding struggles with scalability and intercoder\nreliability. While dictionary-based methods are cost-effective and\ncomputationall... | 2312.05990 | 2504.12558 | 2312.05990_2504.12558 |
[
" Supervised fine-tuning (SFT) is a common method to enhance the tool calling\ncapabilities of Large Language Models (LLMs), with the training data often\nbeing synthesized. The current data synthesis process generally involves\nsampling a set of tools, formulating a requirement based on these tools, and\ngenerati... | 2410.18447 | 2503.18562 | 2410.18447_2503.18562 |
[
" In Australia, post-contact language varieties, including creoles and local\nvarieties of international languages, emerged as a result of forced contact\nbetween Indigenous communities and English speakers. These contact varieties\nare widely used, yet are poorly supported by language technologies. This gap\npres... | 2503.03186 | 2410.09437 | 2503.03186_2410.09437 |
[
" Recently, model merging methods have demonstrated powerful strengths in\ncombining abilities on various tasks from multiple Large Language Models\n(LLMs). While previous model merging methods mainly focus on merging\nhomogeneous models with identical architecture, they meet challenges when\ndealing with Multimod... | 2503.23733 | 2405.19653 | 2503.23733_2405.19653 |
[
" As researchers continue to optimize AI agents for more effective task\nexecution within operating systems, they often overlook a critical security\nconcern: the ability of these agents to detect \"impostors\" within their\nenvironment. Through an analysis of the agents' operational context, we\nidentify a signif... | 2502.13053 | 2504.07072 | 2502.13053_2504.07072 |
[
" This work develops a probabilistic child language acquisition model to learn\na range of linguistic phenonmena, most notably long-range syntactic\ndependencies of the sort found in object wh-questions, among other\nconstructions. The model is trained on a corpus of real child-directed speech,\nwhere each utteran... | 2503.12832 | 2502.20560 | 2503.12832_2502.20560 |
[
" Cooperative speech is purposive. From the speaker's perspective, one crucial\npurpose is the transmission of knowledge. Cooperative speakers care about\ngetting things right for their conversational partners. This attitude is a kind\nof respect. Cooperative speech is an ideal form of communication because\nparti... | 2504.16092 | 2406.11794 | 2504.16092_2406.11794 |
[
" Recent advancements in Large Language Models (LLMs) have expanded their\ncontext windows to unprecedented lengths, sparking debates about the necessity\nof Retrieval-Augmented Generation (RAG). To address the fragmented evaluation\nparadigms and limited cases in existing Needle-in-a-Haystack (NIAH), this paper\n... | 2503.00353 | 2504.18458 | 2503.00353_2504.18458 |
[
" The exponential growth of unstructured text data presents a fundamental\nchallenge in modern data management and information retrieval. While Large\nLanguage Models (LLMs) have shown remarkable capabilities in natural language\nprocessing, their potential to transform unstructured text into standardized,\nstruct... | 2503.02650 | 2503.05757 | 2503.02650_2503.05757 |
[
" Large Vision-Language Models (LVLMs) have made significant strides in\nmultimodal comprehension, thanks to extensive pre-training and fine-tuning on\nlarge-scale visual datasets. However, despite their robust textual safety\nmechanisms, they remain vulnerable to harmful visual inputs. Existing\nsafeguards-typica... | 2503.00037 | 2504.11441 | 2503.00037_2504.11441 |
[
" This paper introduces UniDive for Korean, an integrated framework that\nbridges Universal Dependencies (UD) and Universal Morphology (UniMorph) to\nenhance the representation and processing of Korean {morphosyntax}. Korean's\nrich inflectional morphology and flexible word order pose challenges for\nexisting fram... | 2503.21029 | 2503.09925 | 2503.21029_2503.09925 |
[
" Multi-head latent attention (MLA) is designed to optimize KV cache memory\nthrough low-rank key-value joint compression. Rather than caching keys and\nvalues separately, MLA stores their compressed latent representations, reducing\nmemory overhead while maintaining the performance. While MLA improves memory\neff... | 2503.11132 | 2504.08734 | 2503.11132_2504.08734 |
[
" Normalization layers are ubiquitous in modern neural networks and have long\nbeen considered essential. This work demonstrates that Transformers without\nnormalization can achieve the same or better performance using a remarkably\nsimple technique. We introduce Dynamic Tanh (DyT), an element-wise operation\n$DyT... | 2503.10622 | 2504.08274 | 2503.10622_2504.08274 |
[
" The rise of artificial intelligence (AI) technologies, particularly large\nlanguage models (LLMs), has brought significant advancements to the field of\neducation. Among various applications, automatic short answer grading (ASAG),\nwhich focuses on evaluating open-ended textual responses, has seen remarkable\npr... | 2504.05239 | 2412.00127 | 2504.05239_2412.00127 |
[
" A human author can write any length of story without losing coherence. Also,\nthey always bring the story to a proper ending, an ability that current\nlanguage models lack. In this work, we present the LongStory for coherent,\ncomplete, and length-controlled long story generation. LongStory introduces two\nnovel... | 2311.15208 | 2409.19804 | 2311.15208_2409.19804 |
[
" The integration of tools has extended the capabilities of language models\n(LMs) beyond vanilla text generation to versatile scenarios. However,\ntool-augmented language models (TaLMs) often assume 'perfect' information\naccess and tool availability, which may not hold in the real world. To\nsystematically study... | 2503.14227 | 2503.23924 | 2503.14227_2503.23924 |
[
" Large Language Models (LLMs) are reshaping the landscape of computer science\nresearch, driving significant shifts in research priorities across diverse\nconferences and fields. This study provides a comprehensive analysis of the\npublication trend of LLM-related papers in 77 top-tier computer science\nconferenc... | 2504.08619 | 2503.11084 | 2504.08619_2503.11084 |
[
" Modern AI models are increasingly being used as theoretical tools to study\nhuman cognition. One dominant approach is to evaluate whether human-derived\nmeasures (such as offline judgments or real-time processing) are predicted by a\nmodel's output: that is, the end-product of forward pass(es) through the\nnetwo... | 2504.14107 | 2503.04869 | 2504.14107_2503.04869 |
[
" How capable are large language models (LLMs) in the domain of taxation?\nAlthough numerous studies have explored the legal domain in general, research\ndedicated to taxation remain scarce. Moreover, the datasets used in these\nstudies are either simplified, failing to reflect the real-world complexities,\nor una... | 2503.03444 | 2504.13818 | 2503.03444_2504.13818 |
[
" Retrieval-augmented generation (RAG) provides LLMs with relevant documents.\nAlthough previous studies noted that retrieving many documents can degrade\nperformance, they did not isolate how the quantity of documents affects\nperformance while controlling for context length. We evaluate various language\nmodels ... | 2503.04388 | 2503.01263 | 2503.04388_2503.01263 |
[
" This study addresses the gap in the literature concerning the comparative\nperformance of LLMs in interpreting different types of figurative language\nacross multiple languages. By evaluating LLMs using two multilingual datasets\non simile and idiom interpretation, we explore the effectiveness of various\nprompt... | 2410.16461 | 2502.15359 | 2410.16461_2502.15359 |
[
" We propose Low-Rank Sparse Attention (Lorsa), a sparse replacement model of\nTransformer attention layers to disentangle original Multi Head Self Attention\n(MHSA) into individually comprehensible components. Lorsa is designed to\naddress the challenge of attention superposition to understand\nattention-mediated... | 2504.20938 | 2412.12997 | 2504.20938_2412.12997 |
[
" Evaluating the step-by-step reliability of large language model (LLM)\nreasoning, such as Chain-of-Thought, remains challenging due to the difficulty\nand cost of obtaining high-quality step-level supervision. In this paper, we\nintroduce Self-Play Critic (SPC), a novel approach where a critic model evolves\nits... | 2504.19162 | 2503.08990 | 2504.19162_2503.08990 |
[
" The evaluation of machine-generated image captions is a complex and evolving\nchallenge. With the advent of Multimodal Large Language Models (MLLMs), image\ncaptioning has become a core task, increasing the need for robust and reliable\nevaluation metrics. This survey provides a comprehensive overview of\nadvanc... | 2503.14604 | 2503.16212 | 2503.14604_2503.16212 |
[
" Open domain question answering systems frequently rely on information\nretrieved from large collections of text (such as the Web) to answer questions.\nHowever, such collections of text often contain conflicting information, and\nindiscriminately depending on this information may result in untruthful and\ninaccu... | 2410.12311 | 2503.02854 | 2410.12311_2503.02854 |
[
" Semantic similarity measures (SSMs) are widely used in biomedical research\nbut remain underutilized in pharmacovigilance. This study evaluates six\nontology-based SSMs for clustering MedDRA Preferred Terms (PTs) in drug safety\ndata. Using the Unified Medical Language System (UMLS), we assess each method's\nabi... | 2503.20737 | 2503.09219 | 2503.20737_2503.09219 |
[
" This research introduces a comprehensive system based on state-of-the-art\nnatural language processing, semantic embedding, and efficient search\ntechniques for retrieving similarities and thus generating actionable insights\nfrom raw textual information. The system automatically extracts and aggregates\nnormali... | 2503.10094 | 2406.05516 | 2503.10094_2406.05516 |
[
" Trust plays a fundamental role in shaping the willingness of users to engage\nand collaborate with artificial intelligence (AI) systems. Yet, measuring user\ntrust remains challenging due to its complex and dynamic nature. While\ntraditional survey methods provide trust levels for long conversations, they\nfail ... | 2503.07279 | 2504.02674 | 2503.07279_2504.02674 |
[
" Large Multimodal Models (LMMs) have demonstrated exceptional performance\nacross a wide range of domains. This paper explores their potential in\npronunciation assessment tasks, with a particular focus on evaluating the\ncapabilities of the Generative Pre-trained Transformer (GPT) model,\nspecifically GPT-4o. Ou... | 2503.11229 | 2504.14669 | 2503.11229_2504.14669 |
[
" The rise of Large Reasoning Models (LRMs) signifies a paradigm shift toward\nadvanced computational reasoning. Yet, this progress disrupts traditional agent\nframeworks, traditionally anchored by execution-oriented Large Language Models\n(LLMs). To explore this transformation, we propose the LaRMA framework,\nen... | 2503.11074 | 2504.04238 | 2503.11074_2504.04238 |
[
" Large Language models (LLMs) have been prominent for language translation,\nincluding low-resource languages. There has been limited study about the\nassessment of the quality of translations generated by LLMs, including Gemini,\nGPT and Google Translate. In this study, we address this limitation by using\nseman... | 2503.21393 | 2504.08002 | 2503.21393_2504.08002 |
[
" Online fake news moderation now faces a new challenge brought by the\nmalicious use of large language models (LLMs) in fake news production. Though\nexisting works have shown LLM-generated fake news is hard to detect from an\nindividual aspect, it remains underexplored how its large-scale release will\nimpact th... | 2504.20013 | 2404.07304 | 2504.20013_2404.07304 |
[
" We introduce a new problem KTRL+F, a knowledge-augmented in-document search\ntask that necessitates real-time identification of all semantic targets within\na document with the awareness of external sources through a single natural\nquery. KTRL+F addresses following unique challenges for in-document search:\n1)u... | 2311.08329 | 2503.01424 | 2311.08329_2503.01424 |
[
" Large Language Models (LLMs) have revolutionized natural language processing,\nyet concerns persist regarding their tendency to reflect or amplify social\nbiases present in their training data. This study introduces a novel evaluation\nframework to uncover gender biases in LLMs, focusing on their occupational\nn... | 2503.15904 | 2504.16134 | 2503.15904_2504.16134 |
[
" When applying pre-trained large language models (LLMs) to address anomaly\ndetection tasks, the multivariate time series (MTS) modality of anomaly\ndetection does not align with the text modality of LLMs. Existing methods\nsimply transform the MTS data into multiple univariate time series sequences,\nwhich can c... | 2504.09504 | 2312.00326 | 2504.09504_2312.00326 |
[
" Recent breakthroughs in scale have enabled the emergence of powerful\ngenerative language models, and the ability to fine-tune these models on\nvarious tasks by casting them into prompts or instructions. In this landscape,\nthe problem of Unsupervised Domain Adaptation (UDA), or the problem of\nleveraging knowle... | 2401.17514 | 2502.21321 | 2401.17514_2502.21321 |
[
" The limitations sections of scientific articles play a crucial role in\nhighlighting the boundaries and shortcomings of research, thereby guiding\nfuture studies and improving research methods. Analyzing these limitations\nbenefits researchers, reviewers, funding agencies, and the broader academic\ncommunity. We... | 2503.10658 | 2503.03064 | 2503.10658_2503.03064 |
[
" As LLM-as-a-Judge emerges as a new paradigm for assessing large language\nmodels (LLMs), concerns have been raised regarding the alignment, bias, and\nstability of LLM evaluators. While substantial work has focused on alignment\nand bias, little research has concentrated on the stability of LLM evaluators.\nIn t... | 2502.10709 | 2401.09727 | 2502.10709_2401.09727 |
[
" Multi-head latent attention (MLA) is designed to optimize KV cache memory\nthrough low-rank key-value joint compression. Rather than caching keys and\nvalues separately, MLA stores their compressed latent representations, reducing\nmemory overhead while maintaining the performance. While MLA improves memory\neff... | 2503.11132 | 2410.18194 | 2503.11132_2410.18194 |
[
" Existing Multimodal Large Language Models (MLLMs) are predominantly trained\nand tested on consistent visual-textual inputs, leaving open the question of\nwhether they can handle inconsistencies in real-world, layout-rich content. To\nbridge this gap, we propose the Multimodal Inconsistency Reasoning (MMIR)\nben... | 2502.16033 | 2503.18293 | 2502.16033_2503.18293 |
[
" Recent years have witnessed the success of Multimodal Large Language Models\n(MLLMs) in the vision understanding domain. The success of these models can\nlargely be attributed to the dominant scaling law, which states that larger\nparameter sizes and data volumes contribute to better performance. Notably,\ndata ... | 2411.19951 | 2503.23798 | 2411.19951_2503.23798 |
[
" Large models such as Large Language Models (LLMs) and Vision Language Models\n(VLMs) have transformed artificial intelligence, powering applications in\nnatural language processing, computer vision, and multimodal learning. However,\nfully fine-tuning these models remains expensive, requiring extensive\ncomputat... | 2504.14117 | 2504.10277 | 2504.14117_2504.10277 |
[
" Large Language Models are built on the so-called distributional semantic\napproach to linguistic meaning that has the distributional hypothesis at its\ncore. The distributional hypothesis involves a holistic conception of word\nmeaning: the meaning of a word depends upon its relations to other words in the\nmode... | 2405.12084 | 2410.03767 | 2405.12084_2410.03767 |
[
" Large language models (LLMs) have recently demonstrated remarkable\ncapabilities across domains, tasks, and languages (e.g., ChatGPT and GPT-4),\nreviving the research of general autonomous agents with human-like cognitive\nabilities. Such human-level agents require semantic comprehension and\ninstruction-follow... | 2503.03459 | 2504.15573 | 2503.03459_2504.15573 |
[
" Existing approaches to zero-shot Named Entity Recognition (NER) for\nlow-resource languages have primarily relied on machine translation, whereas\nmore recent methods have shifted focus to phonemic representation. Building\nupon this, we investigate how reducing the phonemic representation gap in IPA\ntranscript... | 2503.07214 | 2503.04840 | 2503.07214_2503.04840 |
[
" A binary decision task, like yes-no questions or answer verification,\nreflects a significant real-world scenario such as where users look for\nconfirmation about the correctness of their decisions on specific issues. In\nthis work, we observe that language models exhibit a negative bias in the\nbinary decisions... | 2408.00137 | 2503.15235 | 2408.00137_2503.15235 |
[
" Recent advancements in Large Language Models (LLMs) have expanded their\ncontext windows to unprecedented lengths, sparking debates about the necessity\nof Retrieval-Augmented Generation (RAG). To address the fragmented evaluation\nparadigms and limited cases in existing Needle-in-a-Haystack (NIAH), this paper\n... | 2503.00353 | 2407.00342 | 2503.00353_2407.00342 |
[
" A key feature of human collaboration is the ability to iteratively refine the\nconcepts we have communicated. In contrast, while generative AI excels at the\n\\textit{generation} of content, it often struggles to make specific\nlanguage-guided \\textit{modifications} of its prior outputs. To bridge the gap\nbetw... | 2504.20294 | 2401.01759 | 2504.20294_2401.01759 |
[
" The rise of Large Language Models (LLMs) as evaluators offers a scalable\nalternative to human annotation, yet existing Supervised Fine-Tuning (SFT) for\njudges approaches often fall short in domains requiring complex reasoning. In\nthis work, we investigate whether LLM judges truly benefit from enhanced\nreason... | 2504.00050 | 2503.10648 | 2504.00050_2503.10648 |
[
" Unsupervised sentence embedding representation has become a hot research\ntopic in natural language processing. As a tensor, sentence embedding has two\ncritical properties: direction and norm. Existing works have been limited to\nconstraining only the orientation of the samples' representations while\nignoring ... | 2503.12739 | 2503.02442 | 2503.12739_2503.02442 |
[
" While current large language models have achieved a remarkable success, their\ndata efficiency remains a challenge to overcome. Recently it has been suggested\nthat child-directed speech (CDS) can improve training data efficiency of modern\nlanguage models based on Transformer neural networks. However, it is not... | 2411.09587 | 2406.08598 | 2411.09587_2406.08598 |
[
" Retrieval-Augmented Generation (RAG) has emerged as a reliable external\nknowledge augmentation technique to mitigate hallucination issues and\nparameterized knowledge limitations in Large Language Models (LLMs). Existing\nadaptive RAG (ARAG) systems excel at in-depth exploration within a single\nsource but stru... | 2411.00689 | 2503.20641 | 2411.00689_2503.20641 |
[
" We propose the Mixture of Frozen Experts (MoFE) architecture, which\nintegrates Parameter-efficient Fine-tuning (PEFT) and the Mixture of Experts\n(MoE) architecture to enhance both training efficiency and model scalability.\nBy freezing the Feed Forward Network (FFN) layers within the MoE framework,\nMoFE signi... | 2503.06491 | 2503.04827 | 2503.06491_2503.04827 |
[
" We present Autonomous Data Selection (AutoDS), a method that leverages base\nlanguage models themselves as zero-shot \"generative classifiers\" to\nautomatically curate high-quality mathematical texts. Unlike prior approaches\nthat require human annotations or training a dedicated data filter, AutoDS\nrelies sol... | 2402.07625 | 2502.11492 | 2402.07625_2502.11492 |
[
" Intent detection, a critical component in task-oriented dialogue (TOD)\nsystems, faces significant challenges in adapting to the rapid influx of\nintegrable tools with complex interrelationships. Existing approaches, such as\nzero-shot reformulations and LLM-based dynamic recognition, struggle with\nperformance ... | 2504.13592 | 2503.03592 | 2504.13592_2503.03592 |
[
" Pre-tokenization, the initial step in many modern tokenization pipelines,\nsegments text into smaller units called pretokens, typically splitting on\nwhitespace and punctuation. While this process encourages having full,\nindividual words as tokens, it introduces a fundamental limitation in most\ntokenization al... | 2504.00178 | 2504.18992 | 2504.00178_2504.18992 |
[
" Retrieval-augmented generation (RAG) systems respond to queries by retrieving\nrelevant documents from a knowledge database and applying an LLM to the\nretrieved documents. We demonstrate that RAG systems that operate on databases\nwith untrusted content are vulnerable to denial-of-service attacks we call\njammi... | 2406.05870 | 2501.11790 | 2406.05870_2501.11790 |
[
" In this work, we explore the potential of large language models (LLMs) for\ngenerating functional test scripts, which necessitates understanding the\ndynamically evolving code structure of the target software. To achieve this, we\npropose a case-based reasoning (CBR) system utilizing a 4R cycle (i.e.,\nretrieve,... | 2503.20576 | 2503.04784 | 2503.20576_2503.04784 |
[
" We introduce Skywork R1V, a multimodal reasoning model extending the an\nR1-series Large language models (LLM) to visual modalities via an efficient\nmultimodal transfer method. Leveraging a lightweight visual projector, Skywork\nR1V facilitates seamless multimodal adaptation without necessitating retraining\nof... | 2504.05599 | 2504.00163 | 2504.05599_2504.00163 |
[
" Large Language Models (LLMs) have been increasingly used to optimize the\nanalysis and synthesis of legal documents, enabling the automation of tasks\nsuch as summarization, classification, and retrieval of legal information. This\nstudy aims to conduct a systematic literature review to identify the state of\nth... | 2504.00725 | 2504.00255 | 2504.00725_2504.00255 |
[
" The ability to understand and predict the mental states of oneself and\nothers, known as the Theory of Mind (ToM), is crucial for effective social\ninteractions. Recent research has emerged to evaluate whether Large Language\nModels (LLMs) exhibit a form of ToM. Although recent studies have evaluated ToM\nin LLM... | 2502.21017 | 2410.23746 | 2502.21017_2410.23746 |
[
" This paper leverages past sentence processing studies to investigate whether\nmonolingual and multilingual LLMs show human-like preferences when presented\nwith examples of relative clause attachment ambiguities in Italian and English.\nFurthermore, we test whether these preferences can be modulated by lexical\n... | 2504.09886 | 2504.13172 | 2504.09886_2504.13172 |
[
" Unsupervised commonsense reasoning (UCR) is becoming increasingly popular as\nthe construction of commonsense reasoning datasets is expensive, and they are\ninevitably limited in their scope. A popular approach to UCR is to fine-tune\nlanguage models with external knowledge (e.g., knowledge graphs), but this\nus... | 2305.15932 | 2503.18792 | 2305.15932_2503.18792 |
[
" Reinforcement Learning from Human Feedback (RLHF) has become a cornerstone\nfor aligning large language models (LLMs) with human values. However, existing\napproaches struggle to capture the multi-dimensional, distributional nuances of\nhuman preferences. Methods such as RiC that directly inject raw reward value... | 2503.10669 | 2503.19309 | 2503.10669_2503.19309 |
[
" Patients are increasingly turning to online health Q&A communities for social\nsupport to improve their well-being. However, when this support received does\nnot align with their specific needs, it may prove ineffective or even\ndetrimental. This necessitates a model capable of identifying the social\nsupport ne... | 2503.17421 | 2503.13503 | 2503.17421_2503.13503 |
[
" Phishing attacks remain a critical cybersecurity threat. Attackers constantly\nrefine their methods, making phishing emails harder to detect. Traditional\ndetection methods, including rule-based systems and supervised machine learning\nmodels, either rely on predefined patterns like blacklists, which can be\nbyp... | 2503.22038 | 2504.20643 | 2503.22038_2504.20643 |
[
" Reinforcement Learning with Human Feedback (RLHF) has achieved great success\nin aligning large language models (LLMs) with human preferences. Prevalent RLHF\napproaches are reward-based, following the Bradley-Terry (BT) model assumption,\nwhich may not fully capture the complexity of human preferences. In this ... | 2407.00617 | 2504.18884 | 2407.00617_2504.18884 |
[
" Syntax connects words to each other in very specific ways. Two words are\nsyntactically connected if they depend\n directly on each other. Syntactic connections usually happen within a\nsentence. Gathering all those connection\n across several sentences gives birth to syntax networks. Earlier studies in\nthe f... | 2503.06724 | 2411.17116 | 2503.06724_2411.17116 |
[
" Hallucinations pose a significant obstacle to the reliability and widespread\nadoption of language models, yet their accurate measurement remains a\npersistent challenge. While many task- and domain-specific metrics have been\nproposed to assess faithfulness and factuality concerns, the robustness and\ngeneraliz... | 2504.18114 | 2503.04381 | 2504.18114_2503.04381 |
[
" LLMs have paved the way for truly simple document-level machine translation,\nbut challenges such as omission errors remain. In this paper, we study a simple\nmethod for handling document-level machine translation, by leveraging previous\ncontexts in a multi-turn conversational manner. Specifically, by decomposi... | 2503.10494 | 2503.00955 | 2503.10494_2503.00955 |
[
" LLMs are often claimed to be capable of Natural Language Inference (NLI),\nwhich is widely regarded as a cornerstone of more complex forms of reasoning.\nHowever, recent works show that LLMs still suffer from hallucinations in NLI\ndue to attestation bias, where LLMs overly rely on propositional memory to\nbuild... | 2503.11614 | 2503.06074 | 2503.11614_2503.06074 |
End of preview. Expand in Data Studio
README.md exists but content is empty.
- Downloads last month
- 8