论文检索

Horacio Jarquín Vásquez, Hugo Jair Escalante, Manuel Montes, Mario Ezra Aragon

Abusive language online creates toxic environments and exacerbates social tensions, underscoring the need for robust NLP models to interpret nuanced linguistic cues. This paper introduces GAttention, a novel Gated Attention mechanism that combines the strengths of Contextual attention and Self-attention mechanisms to address the limitations of existing attention models within the text classification task. GAttention capitalizes on local and global query vectors by integrating the internal relationships within a sequence (Self-attention) and the global relationships among distinct sequences (Contextual attention). This combination allows for a more nuanced understanding and processing of sequence elements, which is particularly beneficial in context-sensitive text classification tasks such as the case of abusive language detection. By applying this mechanism to transformer-based encoder models, we showcase how it enhances the model’s ability to discern subtle nuances and contextual clues essential for identifying abusive language, a challenging and increasingly relevant NLP task.

Xinyi Liu, Weiguang Wang, Hangfeng He

With the growing adoption of Large Language Models (LLMs) for open-ended tasks, accurately assessing epistemic uncertainty, which reflects a model’s lack of knowledge, has become crucial to ensuring reliable outcomes. However, quantifying epistemic uncertainty in such tasks is challenging due to the presence of aleatoric uncertainty, which arises from multiple valid answers. While bias can introduce noise into epistemic uncertainty estimation, it may also reduce noise from aleatoric uncertainty. To investigate this trade-off, we conduct experiments on Visual Question Answering (VQA) tasks and find that mitigating prompt-introduced bias improves uncertainty quantification in GPT-4o. Building on prior work showing that LLMs tend to copy input information when model confidence is low, we further analyze how these prompt biases affect measured epistemic and aleatoric uncertainty across varying bias-free confidence levels with GPT-4o and Qwen2-VL. We find that all considered biases have greater effects in both uncertainties when bias-free model confidence is lower. Moreover, lower bias-free model confidence is associated with greater bias-induced underestimation of epistemic uncertainty, resulting in overconfident estimates, whereas it has no significant effect on the direction of bias effect in aleatoric uncertainty estimation. These distinct effects deepen our understanding of bias mitigation for uncertainty quantification and potentially inform the development of more advanced techniques.

Jirui Qi, Shan Chen, Zidi Xiong, Raquel Fernández, Danielle Bitterman, Arianna Bisazza

Recent Large Reasoning Models (LRMs) with thinking traces have shown strong performance on English reasoning tasks. However, the extent to which LRMs can think in other languages is less studied. This is as important as answer accuracy for real-world applications since users may find the thinking trace useful for oversight only if expressed in their languages. In this work, we comprehensively evaluate two leading families of LRMs on our established benchmark XReasoning. Surprisingly, even the most advanced models often revert to English or produce fragmented reasoning in other languages, revealing a substantial gap in the capability of thinking in non-English languages. Promoting models to reason in the user’s language via prompt hacking enhances readability and oversight. This could gain user trust, but reduces answer accuracy, exposing an important trade-off. We further demonstrate that targeted post-training, even with just 100 instances, can mitigate this language mismatch, although accuracy is still degraded. Our results reveal the limited multilingual reasoning capabilities of current LRMs and suggest directions for future research. All code and datasets are released at https://github.com/Betswish/mCoT-XReasoning.

Guanyu Chen, Peiyang Wang, Yizhou Jiang, Yuqian Liu, Chujie Zhao, Ying Fang, Tianren Zhang, Feng Chen

Large language models (LLMs) have been able to perform various forms of reasoning tasks ina wide range of scenarios, but are they truly engaging in task abstraction and rule-based reasoning beyond mere memorization? To answer this question, we propose a novel experimentalapproach, Misleading Fine-Tuning (MisFT), to examine whether LLMs perform abstract reasoning by altering their original understanding of fundamental rules. In particular, by constructing datasets with math expressions or logical formulas that contradict correct principles, we fine-tune the model to learn those contradictory rules and assess its generalization ability on unseen test domains. Through a series of experiments, we find that current LLMs are capable of applying contradictory rules to solve practical math word problems and natural language reasoning tasks, implying the presence of an internal mechanism in LLMs that abstracts before reasoning.

Priyanka Dey, Aayush Bothra, Yugal Khanter, Jieyu Zhao, Emilio Ferrara

As LLMs become central to interactive applications, ranging from tutoring to mental health, the ability to express personality in culturally appropriate ways is increasingly important. While recent works have explored personality evaluation of LLMs, they largely overlook the interplay between culture and personality. To address this, we introduce , the first large-scale benchmark with human validation for evaluating LLMs’ personality expression in culturally grounded, behaviorally rich contexts. Our dataset spans 3,000 scenario-based questions across six diverse countries, designed to elicit personality through everyday scenarios rooted in local values. We evaluate how closely three models’ personality distributions align to real human populations through two evaluation settings: multiple-choice and open-ended response formats. Our results show– improves alignment with country-specific human personality distributions (over a 20% reduction in Wasserstein distance across models and countries) and elicits more expressive, culturally coherent outputs compared to existing benchmarks. surfaces meaningful modulate trait outputs in response to culturally grounded prompts, offering new directions for aligning LLMs to global norms of behavior. By bridging personality expression and cultural nuance, we envision that will pave the way for more socially intelligent and globally adaptive LLMs. Datasets and code are available at: https://github.com/limenlp/CulturalPersonas.

Bohao Yang, Dong Liu, Chenghao Xiao, Kun Zhao, Chen Tang, Chao Li, Lin Yuan, Yang Guang, Chenghua Lin

Large Language Models (LLMs) demonstrate remarkable ability to comprehend instructions and generate human-like text, enabling sophisticated agent simulation beyond basic behavior replication. However, the potential for creating freely customisable characters remains underexplored. We introduce the Customisable Conversation Agent Framework, which employs LLMs to simulate real-world characters through personalised characteristic feature injection, enabling diverse character creation according to user preferences.We propose the SimsConv dataset, comprising 68 customised characters and 13,971 multi-turn role-playing dialogues across 1,360 real-world scenes. Characters are initially customised using pre-defined elements (career, aspiration, traits, skills), then expanded through personal and social profiles. Building on this, we present SimsChat, a freely customisable role-playing agent incorporating various realistic settings and topic-specified character interactions.Experimental results on both SimsConv and WikiRoleEval datasets demonstrate SimsChat’s superior performance in maintaining character consistency, knowledge accuracy, and appropriate question rejection compared to existing models. Comprehensive ablation studies validate each component’s contribution to overall performance, with the pre-defined aspects framework and scene construction showing particularly significant impact. Our framework provides valuable insights for developing more accurate and customisable human simulacra.Our data and code are publicly available at https://github.com/Bernard-Yang/SimsChat.

Mayank Kulkarni, Vittorio Mazzia, Judith Gaspers, Chris Hench, Jack FitzGerald

We present MASSIVE-Agents, a new benchmark for assessing multilingual function calling across 52 languages. We created MASSIVE-Agents by cleaning the original MASSIVE dataset and then reformatting it for evaluation within the Berkeley Function-Calling Leaderboard (BFCL) framework. The full benchmark comprises 47,020 samples with an average of 904 samples per language, covering 55 different functions and 286 arguments. We benchmarked 21 models using Amazon Bedrock and present the results along with associated analyses. MASSIVE-Agents is challenging, with the top model Nova Premier achieving an average Abstract Syntax Tree (AST) Accuracy of 34.05% across all languages, with performance varying significantly from 57.37% for English to as low as 6.81% for Amharic. Some models, particularly smaller ones, yielded a score of zero for the more difficult languages. Additionally, we provide results from ablations using a custom 1-shot prompt, ablations with prompts translated into different languages, and comparisons based on model latency.

Yi-Pei Chen, Noriki Nishida, Hideki Nakayama, Yuji Matsumoto

Multi-session persona-based dialogue generation presents challenges in maintaining long-term consistency and generating diverse, personalized responses. While large language models (LLMs) excel in single-session dialogues, they struggle to preserve persona fidelity and conversational coherence across extended interactions. Existing methods typically retrieve persona information before response generation, which can constrain diversity and result in generic outputs. We propose Post Persona Alignment (PPA), a novel two-stage framework that reverses this process. PPA first generates a general response based solely on dialogue context, then retrieves relevant persona memories using the response as a query, and finally refines the response to align with the speaker’s persona. This post-hoc alignment strategy promotes naturalness and diversity while preserving consistency and personalization. Experiments on multi-session LLM-generated dialogue data demonstrate that PPA significantly outperforms prior approaches in consistency, diversity, and persona relevance, offering a more flexible and effective paradigm for long-term personalized dialogue generation.

Minghang Liu, Yinghan Shen, Zihe Huang, Yuanzhuo Wang, Xuhui Jiang, Huawei Shen

Multimodal Knowledge Graphs (MMKGs) enhance knowledge representations by integrating structural and multimodal information of entities. Recently, MMKGs have proven effective in tasks such as information retrieval, knowledge discovery, and question answering. Current methods typically utilize pre-trained visual encoders to extract features from images associated with each entity, emphasizing complex cross-modal interactions. However, these approaches often overlook the varying relevance of visual information across entities. Specifically, not all entities benefit from visual data, and not all associated images are pertinent, with irrelevant images introducing noise and potentially degrading model performance. To address these issues, we propose the Differentiated Vision for Multimodal Knowledge Graphs (DVMKG) model. DVMKG evaluates the necessity of visual modality for each entity based on its intrinsic attributes and assesses image quality through representativeness and diversity. Leveraging these metrics, DVMKG dynamically adjusts the influence of visual data during feature integration, tailoring it to the specific needs of different entity types. Extensive experiments on multiple benchmark datasets confirm the effectiveness of DVMKG, demonstrating significant improvements over existing methods.

Bohao Yang, Kun Zhao, Dong Liu, Chen Tang, Liang Zhan, Chenghua Lin

Automatic open-domain dialogue evaluation has attracted increasing attention, yet remains challenging due to the complexity of assessing response appropriateness. Traditional evaluation metrics, typically trained with true positive and randomly selected negative responses, tend to assign higher scores to responses that share greater content similarity with contexts. However, adversarial negative responses, despite possessing high lexical overlap with contexts, can be semantically incongruous. Consequently, existing metrics struggle to evaluate such responses effectively, resulting in low correlations with human judgments. While recent studies have demonstrated the effectiveness of Large Language Models (LLMs) for open-domain dialogue evaluation, they still face challenges in handling adversarial negative examples. We propose a novel evaluation framework that integrates Abstract Meaning Representation (AMR) enhanced domain-specific language models (SLMs) with LLMs. Our SLMs explicitly incorporate AMR graph information through a gating mechanism for enhanced semantic representation learning, while both SLM predictions and AMR knowledge are integrated into LLM prompts for robust evaluation. Extensive experiments on open-domain dialogue evaluation tasks demonstrate the superiority of our method compared to state-of-the-art baselines, particularly in discriminating adversarial negative responses. Our framework achieves strong correlations with human judgments across multiple datasets, establishing a new benchmark for dialogue evaluation. Our code and data are publicly available at https://github.com/Bernard-Yang/SIMAMR.

Chongxin Li, Hanzhang Wang, Yuchun Fang

Adversarial vulnerabilities in vision-language models pose a critical challenge to the reliability of large language systems, where typographic manipulations and adversarial perturbations can effectively bypass language model defenses. We introduce Attack as Defense (AsD), the first approach to proactively defend at the cross-modality level, embedding protective perturbations in vision to disrupt attacks before they propagate to the language model. By leveraging the semantic alignment between vision and language, AsD enhances adversarial robustness through model perturbations and system-level prompting. Unlike prior work that focuses on text-stage defenses, our method integrates visual defenses to reinforce prompt-based protections, mitigating jailbreaking attacks across benchmarks. Experiments on the LLaVA-1.5 show that AsD reduces attack success rates from 56.7% to 12.6% for typographic attacks and from 89.0% to 47.5% for adversarial perturbations. Further analysis reveals that the key bottleneck in vision-language security lies not in isolated model vulnerabilities, but in cross-modal interactions, where adversarial cues in the vision model fail to consistently activate the defense mechanisms of the language model.

Frederic Kirstein, Sonu Kumar, Terry Ruas, Bela Gipp

Meeting summarization with large language models (LLMs) remains error-prone, often producing outputs with hallucinations, omissions, and irrelevancies. We present FRAME, a modular pipeline that reframes summarization as a semantic enrichment task. FRAME extracts and scores salient facts, organizes them thematically, and uses these to enrich an outline into an abstractive summary. To personalize summaries, we introduce SCOPE, a reason-out-loud protocol that has the model build a reasoning trace by answering nine questions before content selection. For evaluation, we propose P-MESA, a multi-dimensional, reference-free evaluation framework to assess if a summary fits a target reader. P-MESA reliably identifies error instances, achieving \geq 89% balanced accuracy against human annotations and strongly aligned with human severity ratings (\rho \geq 0.70). On QMSum and FAME, FRAME reduces hallucination and omission by 2 out of 5 points (measured with MESA), while SCOPE improves knowledge fit and goal alignment over prompt-only baselines. Our findings advocate for rethinking summarization to improve control, faithfulness, and personalization.

Rumana Ferdous Munne, Md Mostafizur Rahman, Yuji Matsumoto

Entity alignment (EA) involves identifying and linking equivalent entities across different knowledge graphs (KGs). While knowledge graphs provide structured information about real-world entities, only a small fraction of these entities are aligned. The entity alignment process is challenging due to heterogeneity in KGs, such as differences in structure, terminology, and attribute details. Traditional EA methods use multi-aspect entity embeddings to align entities. Although these methods perform well in certain scenarios, their effective- ness is often constrained by sparse or incomplete data in knowledge graphs and the limitations of embedding techniques. We propose ProLEA ( Profile Generation and Reasoning with LLMs for Entity Alignment) an entity alignment method that combines large language models (LLMs) with entity embed- dings. LLMs generate contextual profiles for entities based on their properties. Candidate entities identified by entity embedding techniques are then re-evaluated by the LLMs, using its background knowledge and the generated profile. A thresholding mechanism is introduced to resolve conflicts between LLMs predictions and embedding-based alignments. This method enhances alignment accuracy, robustness, and explainability, particularly for complex, het- erogeneous knowledge graphs. Furthermore, ProLEA is a generalized framework. Its profile generation and LLM-enhanced entity align- ment components can improve the performance of existing entity alignment models.

Filippo Momentè, Alessandro Suglia, Mario Giulianelli, Ambra Ferrari, Alexander Koller, Oliver Lemon, David Schlangen, Raquel Fernández, Raffaella Bernardi

We examine three evaluation paradigms: standard benchmarks (e.g., MMLU and BBH), interactive games (e.g., Signalling Games or Taboo), and cognitive tests (e.g., for working memory or theory of mind). First, we investigate which of the former two—benchmarks or games—is most effective at discriminating LLMs of varying quality. Then, inspired by human cognitive assessments, we compile a suite of targeted tests that measure cognitive abilities deemed essential for effective language use, and we investigate their correlation with model performance in benchmarks and games. Our analyses reveal that interactive games are superior to standard benchmarks in discriminating models. Causal and logical reasoning correlate with both static and interactive tests, while differences emerge regarding core executive functions and social/emotional skills, which correlate more with games. We advocate for the development of new interactive benchmarks and targeted cognitive tasks inspired by assessing human abilities but designed specifically for LLMs.

Davide Testa, Giovanni Bonetta, Raffaella Bernardi, Alessandro Bondielli, Alessandro Lenci, Alessio Miaschi, Lucia Passaro, Bernardo Magnini

We introduce MAIA (Multimodal AI Assessment), a native-Italian benchmark designed for fine-grained investigation of the reasoning abilities of visual language models on videos. MAIA differs from other available video benchmarks for its design, its reasoning categories, the metric it uses, and the language and culture of the videos. MAIA evaluates Vision Language Models (VLMs) on two aligned tasks: a visual statement verification task, and an open-ended visual question-answering task, both on the same set of video-related questions. It considers twelve reasoning categories that aim to disentangle language and vision relations by highlighting the role of the visual input. Thanks to its carefully taught design, it evaluates VLMs’ consistency and visually grounded natural language comprehension and generation simultaneously through an aggregated metric revealing low results that highlight models’ fragility. Last but not least, the video collection has been carefully selected to reflect the Italian culture, and the language data are produced by native-speakers.Data available at *[GitHub](https://github.com/Caput97/MAIA-Multimodal_AI_Assessment.git).*

Sangmin Song, Juhwan Choi, JungMin Yun, YoungBin Kim

Large language models (LLMs) have demonstrated remarkable performance in zero-shot dialogue state tracking (DST), reducing the need for task-specific training. However, conventional DST benchmarks primarily focus on structured user-agent conversations, failing to capture the complexities of real-world multi-user interactions. In this study, we assess the robustness of LLMs in multi-user DST while minimizing dataset construction costs. Inspired by recent advances in LLM-based data annotation, we extend an existing DST dataset by generating utterances of a second user based on speech act theory. Our methodology systematically incorporates a second user’s utterances into conversations, enabling a controlled evaluation of LLMs in multi-user settings. Experimental results reveal a significant performance drop compared to single-user DST, highlighting the limitations of current LLMs in extracting and tracking dialogue states amidst multiple speakers. Our findings emphasize the need for future research to enhance LLMs for multi-user DST scenarios, paving the way for more realistic and robust DST models.

Xiaoyi Wang, Jiwei Zhang, Guangtao Zhang, Honglei Guo

Synthetic therapy dialogues generated by large language models (LLMs) are increasingly used in mental health NLP to simulate counseling scenarios, train models, and supplement limited real-world data. However, it remains unclear whether these synthetic conversations capture the nuanced emotional dynamics of real therapy. In this work, we introduce RealCBT, a dataset of authentic cognitive behavioral therapy (CBT) dialogues, and conduct the first comparative analysis of emotional arcs between real and LLM-generated CBT sessions. We adapt the Utterance Emotion Dynamics framework to analyze fine-grained affective trajectories across valence, arousal, and dominance dimensions. Our analysis spans both full dialogues and individual speaker roles (counselor and client), using real sessions from the RealCBT dataset and synthetic dialogues from the CACTUS dataset. We find that while synthetic dialogues are fluent and structurally coherent, they diverge from real conversations in key emotional properties: real sessions exhibit greater emotional variability, more emotion-laden language, and more authentic patterns of reactivity and regulation. Moreover, emotional arc similarity remains low across all pairings, with especially weak alignment between real and synthetic speakers. These findings underscore the limitations of current LLM-generated therapy data and highlight the importance of emotional fidelity in mental health applications. To support future research, our dataset RealCBT is released at https://gitlab.com/xiaoyi.wang/realcbt-dataset.

Hangdi Xing, Feiyu Gao, Qi Zheng, Zhaoqing Zhu, Zirui Shao, Ming Yan

In the daily work, vast amounts of documents are stored in pixel-based formats such as images and scanned PDFs, posing challenges for efficient database management and data processing. Existing methods often fragment the parsing process into the pipeline of separated subtasks on the layout element level, resulting in incomplete semantics and error propagation. Even though models based on multi-modal large language models (MLLMs) mitigate the issues to some extent, they also suffer from absent or sub-optimal grounding ability for visual information. To address these challenges, we introduce the Intelligent Document Parsing (IDP) framework, an end-to-end document parsing framework leveraging the vision-language priors of MLLMs, equipped with an elaborately designed document representation and decoding mechanism to decouple the content parsing and layout grounding to fully activate the potential of MLLMs for document parsing. Experimental results demonstrate that the IDP method surpasses existing methods, significantly advancing MLLM-based document parsing.

Lujie Niu, Haochen Sun, Fangkun Zhao, Sheng Chen, Zimeng Bai, Jiawei Zhang, Caixia Yuan, Xiaojie Wang

Chain-of-Thought (CoT) distillation has emerged as a promising paradigm to enhance the reasoning ability of small language models by imitating the reasoning and outputs of larger teacher models. However, existing approaches suffer from a critical limitation: a distribution mismatch between teacher-generated training trajectories and the student model’s own generative distribution. This mismatch leads to exposure bias during inference and often induces mode collapse or mode averaging, thereby degrading the student model’s generative diversity and robustness. To address these issues, we propose CoTD-PO (Chain-of-Thought Distillation with Preference Optimization), a reinforcement learning framework that shifts the training paradigm from passive imitation to active trajectory exploration. Instead of forcing the student to imitate exact teacher traces, our method enables the student to sample its own answer paths. To support training with non-open-source teacher models, we approximate the teacher’s output distribution through preference-based scoring. Furthermore, we adopt an offline iterative training procedure that enables stable and efficient optimization. Experiments on diverse open-ended generation tasks demonstrate that CoTD-PO significantly outperforms standard CoT distillation baselines, achieving higher output quality while mitigating mode collapse and preserving semantic diversity.

Paramita Mirza, Lucas Weber, Fabian Küch

Recent work shows that post-training datasets for LLMs can be substantially downsampled without noticeably deteriorating performance. However, data selection often incurs high computational costs or is limited to narrow domains. In this paper, we demonstrate that data selection can be both—efficient and universal—by using a multi-step pipeline in which we efficiently bin data points into groups, estimate quality using specialized models, and score difficulty with a robust, lightweight method. Task-based categorization allows us to control the composition of our final data—crucial for finetuning multi-purpose models. To guarantee diversity, we improve upon previous work using embedding models and a clustering algorithm. This integrated strategy enables high-performance fine-tuning with minimal overhead.