Detecting fake news is essential in natural language processing to verify news authenticity and prevent misinformation-driven social, political, and economic disruptions targeting specific groups. A major challenge in multimodal fake news detection is effectively integrating textual and visual modalities, as semantic gaps and contextual variations between images and text complicate alignment, interpretation, and the detection of subtle or blatant inconsistencies. To enhance accuracy in fake news detection, this article introduces an ensemble-based framework that integrates textual and visual data using ViLBERT’s two-stream architecture, incorporates VADER sentiment analysis to detect emotional language, and uses Image–Text Contextual Similarity to identify mismatches between visual and textual elements. These features are processed through the Bi-GRU classifier, Transformer-XL, DistilBERT, and XLNet, combined via a stacked ensemble method with soft voting, culminating in a T5 metaclassifier that predicts the outcome for robustness. Results on the Fakeddit and Weibo benchmarking datasets show that our method outperforms state-of-the-art models, achieving up to 96% and 94% accuracy in fake news detection, respectively. This study highlights the necessity for advanced multimodal fake news detection systems to address the increasing complexity of misinformation and offers a promising solution.

Conversational stance detection aims to infer a user's stance toward a specific target by modeling the conversa tion context across multiple interaction turns. However, as conversations become longer, coreference reasoning becomes increasingly challenging, and the topics discussed by users often shift implicitly over time. These phe nomena pose substantial challenges to conversational stance detection. To address these issues, we propose a Knowledge-Injected and Target-Aware Enhancement Framework for Conversational Stance Detection, termed KITE-CSD, which explicitly models the relationships between the topics discussed in the conversation context and the specific target. Specifically, we first employ an LLM-agent to perform contextual coreference resolu tion, effectively clarifying ambiguous pronouns, abbreviations, and slang expressions in user comments, thereby improving the precision of semantic modeling. We then introduce a JSON-based structured text representation to encode reply relations in multi-turn conversations. Furthermore, we adopt a two-stage fine-tuning strategy to extract and inject topic-target relation knowledge, enabling the model to better handle topic shifts during conver sations. Extensive experiments on two public benchmark datasets demonstrate the effectiveness and superiority of KITE-CSD, achieving SOTA performance. Our code is available at https://github.com/mengfeiyang11/KITE-CSD.

Paraphrasing uses different words, sentence structures, or expressions to convey similar semantics. It is an effective training data augmentation method to improve low-resource Natural Language Processing (NLP) tasks. Existing studies normally leverage parallel corpora to construct parabanks, regarding the Machine Translation (MT) results of source sentences as the paraphrases of the corresponding target sentences. As MT models are usually trained on the same parallel corpus, translation of the training set may suffer from overfitting, which leads to less diverse paraphrases. Training paraphrasers on the parabank generated via MT may also suffer from the information loss issue, as the parabank is derived from the parallel corpora, and the knowledge inside the parabank is a subset of that inside the parallel corpora. In this paper, we train bidirectional Multilingual Neural Machine Translation (MNMT) on the bi-directional bilingual parallel corpus, and use the MNMT model directly as a paraphrasing model by asking it to generate "translations" of the input language. As some source tokens also appear in the translation in the parallel corpus, we introduce "copy"/"not-copy" tags to indicate the existence/non-existence of source tokens in the target translation during training, and use the "not-copy" tag to encourage paraphrasing during inference. Manual and automatic evaluation results show that our ParaMNMT method can generate paraphrases of higher semantic consistency, literal fluency and sentential diversity compared to existing parabanks and LLMs. Our data augmentation experiments verify the effectiveness of ParaMNMT on improving low-resource NLP tasks.

Sycophancy in large language models (LLMs) remains a critical yet underexplored alignment failure, especially in complex interactive settings. Existing benchmarks, often restricted to static and rudimentary tasks, fail to capture these deep alignment failures. Addressing this gap, we propose SycoPrism, a tri-facet benchmark that systematically characterizes sycophantic behavior across three dimensions: Opinion Emotion, Factual Concession, and a novel Instruction Hijacking scenario, which captures a previously underexplored yet more subtle failure mode. Based on this benchmark, we conduct a comprehensive evaluation of over ten mainstream LLMs, uncovering significant underlying deficiencies in current alignment paradigms. Specifically, we find that these paradigms drive models into “safety evasion” in subjective contexts, while strong task instructions compel them to ignore absurd premises and prioritize execution over factual adherence, exposing a severe distortion in their objective prioritization. We further find that scaling laws provide partial mitigation of sycophantic behavior; however, targeted preference alignment for sycophancy remains essential. For quantitative evaluation, we employ an LLM-as-a-Judge approach, but it relies on three specific prompts and produces only discrete classifications. To overcome this limitation, we propose a Cross-Model Preference Pooling strategy to construct a paired dataset, based on which we train a lightweight 8B-parameter reward model (RM) to generalize sycophancy detection across all dimensions. Experimental results show that the proposed RM significantly reduces inference costs while achieving an interception accuracy of up to 92.35 https://github.com/Xiaoyu7911/SycoPrism .

Text-to-SQL is a critical task in natural language processing that aims to transform natural language questions into accurate and executable SQL queries. In real-world scenarios, these reasoning tasks are often accompanied by complex mathematical computations, domain knowledge, and hypothetical reasoning scenarios. However, existing large-scale Text-to-SQL datasets typically focus on business logic and task logic, neglecting critical factors such as vertical domain knowledge, complex mathematical reasoning, and hypothetical reasoning, which are essential for realistically reflecting the reasoning demands in practical applications and completing data querying and analysis. To bridge this gap, we introduce LogicCat, the first Text-to-SQL benchmark dataset specifically designed for complex reasoning and chain-of-thought parsing, encompassing physics, arithmetic, commonsense, and hypothetical reasoning scenarios. LogicCat comprises 4,038 English questions paired 12,114 detailed chain-of-thought reasoning steps, spanning 45 databases across diverse domains, significantly surpassing existing datasets in complexity. Experimental results demonstrate that LogicCat substantially increases the task difficulty for current state-of-the-art models to at most 33.20% execution accuracy, indicating that this task remains exceptionally challenging. The advancement of LogicCat represents a crucial step toward developing systems suitable for real-world enterprise data analysis and autonomous query generation.

Real-world image–text posts mix subtle visual signals with short, informal language, making multimodal sentiment analysis susceptible to label noise, miscalibration, and brittle decisions. We present SentiGuard+, a reliability-aware model that processes images with a Swin Transformer and text with RoBERTa-large, fusing them via a bi-directional cross-attention module (CrossWeave). For each input, the system produces three interpretable outputs: class probabilities, a calibrated reliability score, and an accept/abstain decision for risk-controlled deployment. The fused representation is equipped with an evidential Dirichlet head and a hierarchical noisy-label channel (HD-SIT); we then apply reliability-preserving evidence calibration and a simple two-feature gate (P-Gate) that bases acceptance on predicted entropy and calibrated variance. Across standard multimodal sentiment benchmarks, SentiGuard+ improves overall accuracy and F1 while reducing calibration error, and it delivers stronger accepted-set performance at matched coverage. We find evidence that a single calibrated-and-gated operating point transfers between two datasets with minimal drift. These results indicate that combining Swin Transformer and RoBERTa-large with evidential modeling and selective gating yields better-conditioned probabilities, higher accepted-set F1, and portable operating points suitable for real-world use.

Social media has become a primary source of information, but its rapid growth has facilitated the spread of misinformation. Existing methods often rely on static graph neural networks (GNNs) to analyze news propagation patterns, overlooking temporal dynamics and semantic nuances. To address this, we propose a Temporal Graph Convolution Network with a Hierarchical Co-Attention (TGCN-HCA) model for effective fake news detection. TGCN-HCA leverages a dynamic graph representation to capture the evolving nature of news propagation. By incorporating temporal information, we can accurately model the sequential spread of news and identify patterns associated with fake news. To enhance semantic understanding, we employ GPT-2 to extract deep contextual features from news content. These features, combined with structural information from the graph, provide a rich representation of news propagation. A hierarchical co-attention mechanism further refines feature representation by jointly considering multiple perspectives: the importance of nodes within the graph and the relevance of semantic features to the task of fake news detection. By integrating these components, TGCN-HCA effectively captures the complex interplay between temporal dynamics, structural information, and semantic content. Extensive experiments on Twitter15, Twitter16, and Weibo datasets demonstrate the superior performance of TGCN-HCA over state-of-the-art methods. Our model achieves significant accuracy gains of 5.37

Few-shot Knowledge Graph Completion (FKGC) refers to querying facts unseen in relations given a few-shot of reference entity pairs. Most FKGC methods model and measure representations of relations determined by few-shots, ignoring the impact of the semantic uncertainty of few-shot relations. To address this, we propose few-shot knowledge graph completion model based on Gaussian embedding metric learning (GEM-Net). GEM-Net screens entity neighbors based on relation relevance, aggregates neighbors using a cross-attention mechanism to enhance entity representations. On one hand, adaptive aggregation supports entity pairs to obtain representations of few-shot relations, designing Gaussian encoders to construct Gaussian distributions of few-shot relations, sampling latent variables as uncertainty representations, and using a stochastic ManifoldE decoder to measure uncertainty. On the other hand, it generates query-specific relation prototypes to measure determined feature representations. Link prediction evaluations of GEM-Net on two public datasets, NELL and Wiki, demonstrate that GEM-Net outperforms baseline models.

Red-teaming Vision-Language Models is essential for identifying vulnerabilities where adversarial image-text inputs trigger toxic outputs. Existing approaches treat image generation as a black box, returning only terminal toxicity scores and leaving open the question of when and how toxic semantics emerge during multi-step synthesis. We introduce STARE, a hierarchical reinforcement learning framework that treats the denoising trajectory itself as the attack surface, under a direct white-box T2I and query-only black-box VLM setting. By coupling a high-level prompt editor with low-level T2I fine-tuning via Group Relative Policy Optimization (GRPO), STARE attains a 68% improvement in Attack Success Rate over state-of-the-art black-box and white-box baselines. More importantly, this trajectory-level view surfaces the Optimization-Induced Phase Alignment phenomenon: vanilla models exhibit diffuse toxicity, whereas adversarial optimization concentrates conceptual harms into early semantic phases and detail-oriented harms into late refinement. Targeted perturbations of either window selectively suppress different toxicity categories, indicating that this temporal structure is a genuine causal handle rather than a side effect of the hierarchical design. The phenomenon turns toxicity formation from a chaotic process into a small set of predictable vulnerability windows, providing both a potent attack engine and a basis for phase-aware safety mechanisms. Content warning: This paper contains examples of toxic content that may be offensive or disturbing.

Schema linking is a critical component of Text-to-SQL systems, but existing approaches often trade off contextual modeling capacity, score-based controllability, and inference efficiency. We introduce AttnLink, an attention-based framework that converts LLMs' internal attention into continuous relevance scores for schema items. AttnLink extracts the attention from the generation-start position to candidate schema spans, enabling all candidates to be ranked in a single prefill pass without autoregressive decoding. We develop two variants: AttnLink-U, which directly probes pretrained attention without parameter updates, and AttnLink-S, which aligns the attention distribution with gold schema items through direct supervision. To improve coverage of multiple relevant schema items, AttnLink-S combines a set-mass objective with an adaptive probability-floor regularizer. The resulting scores support post-hoc precision-recall control through temperature scaling and cumulative-mass selection. Experiments on Spider, BIRD, and Spider2-SQLite show that AttnLink-S achieves mAP scores of 99.22

The rapid advancement of large language models has significantly enhanced the quality of AI-generated text, making it increasingly difficult for detection systems to distinguish from human-written content. Existing detection methods, such as statistical, linguistic, machine learning, and deep learning approaches, often exhibit a decline in performance when applied to new or previously unseen large language models. Additionally, they tend to become outdated due to their static frameworks and inability to adapt to emerging patterns in generative text. To address this limitation, we introduce a novel dynamic fusion framework that integrates multi-scale feature fusion to capture diverse text patterns and employs continual learning with Elastic Weight Consolidation (EWC) to adapt to new models while mitigating catastrophic forgetting. This is the first attempt, to the best of our knowledge, to develop such a dynamic framework for AI-generated text detection. Evaluated on the TuringBench and DeepfakeTextDetect benchmark datasets, our framework achieves an average accuracy of 95.78% and 92.39%, outperforming the standard model by 5.88% and 7.98%, respectively, in distinguishing AI-generated from human-written text across various language generative model architectures. The continual learning ensures that the model remains adaptive and accurate over time, which is essential for practical applications in dynamic environments. This dynamic and adaptive approach paves the way for resilient AI-generated text detection systems capable of evolving alongside the rapidly advancing landscape of generative language technologies.

Recent Text-to-SQL systems increasingly rely on multi-turn interaction, execution feedback, and reinforcement learning. However, most existing methods use execution correctness only as a trajectory-level reward, which provides limited guidance for identifying the SQL decisions responsible for success or failure. We propose SERL-SQL, a selective execution-grounded reinforcement learning framework for multi-turn Text-to-SQL agents. SERL-SQL samples on-policy SQL interaction trajectories and uses a training-only teacher to re-score student actions with execution feedback. The resulting teacher–student likelihood gap is converted into bounded, masked weights that reweight GRPO advantages only on SQL and tool-action tokens. In this way, task rewards preserve the optimization direction, while execution hindsight provides localized credit assignment. Experiments on BIRD, Spider, and cross-domain benchmarks show that SERL-SQL achieves competitive performance, reaching 76.56

Text classification is one of the most common tasks in the field of natural language processing. Currently, and the pre-trained fine-tuning paradigm has achieved outstanding results in text classification. However, for fine-tuning on downstream tasks, an ample amount of training data is still crucial. When manually annotated training data is insufficient, a common solution is data augmentation, using machine translation techniques for back translation. We generate sentence embeddings using pre-trained models and apply a similarity-based filtering method to automatically screen back-translated samples, thereby improving back-translation quality. By incorporating LoRa fine-tuning on the large model Baichuan2, we further improve text classification accuracy. Compared to the traditional BERT model, our approach achieved a 0.5

Medical relation extraction aims to extract pairs of entities and their corresponding relations from unstructured text, which faces the challenge of a scarcity of labeled data. Zero-shot relation extraction can extract new relations not observed during training and alleviate the problem of scarce medical data. However, existing zero-shot relation extraction methods use semantic similarity matching with limited domain representation capability. In this paper, we introduces an Unified framework for zero-shot Medicial Relation Extraction with Large Language Models (UniMRE), which leverages Large Language Models’ (LLMs) advanced contextual understanding capabilities to extract relation triplets in zero-shot setting. UniMRE employs a knowledge injection strategy to infuse medical knowledge into LLMs, which enables the generation of silver labels. These labels are used to retrieve relevant samples and relation rules, which are processed by a relation extraction agent. Based on their evaluation scores, high-confidence labels are incorporated into the sample library as gold labels, while low-confidence labels are refined and regenerated based on identified errors. Extensive experiments on medical datasets demonstrate that UniMRE outperforms baseline models, validating its effectiveness in extracting structured medical knowledge.

Document-level Relation Extraction (DocRE) aims to identify the relation types between entity pairs within a document. Large language models (LLMs) have garnered attention due to their strong reasoning capabilities. However, directly applying LLMs to DocRE often results in suboptimal performance, primarily due to long-context inputs inhibiting its reasoning capability. Logical rules can inject refined information into LLMs due to their explicit reasoning process. In this paper, we propose logical Rule-constrained LLMs or Ru-LLM for short, a novel framework that learns logical rules to enhance the performance of LLMs in DocRE. Specifically, Ru-LLM leverages frequent pattern mining and LLM-assisted completion to obtain the candidate predicate pool. To further refine the rule search process, Ru-LLM incorporates the Monte Carlo Tree Search algorithm to efficiently search for and generate logical rules through a four-stage process: selection, expansion, simulation, and backpropagation. Ru-LLM transforms logical rules into concise prompts to guide initial result generation, which is then subjected to a Chain-of-Thought verification process that retains only those triples satisfying all rules. Experimental results on the three datasets DWIE, DocRED, and Re-DocRED demonstrate that Ru-LLM outperforms existing implicit reasoning models, LLM-based models, and rule-based frameworks.

Hybrid tabular-textual question answering (QA) typically requires numerical reasoning over heterogeneous data, with the reasoning program first generated and then executed to obtain the final answer. However, in most existing hybrid QA benchmarks, each step only relies on the numbers in the input or the calculation result of the preceding step. Questions that require long-distance numerical reasoning (where the calculation of a step relies on the results calculated a number of steps previously) are rare. However, they may be required to solve some complicated financial problems; these questions require a more complex model capable of capturing long-distance dependencies. To study more challenging hybrid tabular-textual QA, we construct a new large-scale hybrid tabular-textual dataset, COLD-QA, COmplex Long-Distance numerical reasoning Question Answering dataset. We also conduct extensive experiments with multiple baselines. The COLD-QA dataset is significantly more difficult than previous work, according to experiment results.

This paper focuses on Dialogue Aspect-based Sentiment Quadruple (DiaASQ) analysis, aiming to extract structured quadruples from multi-turn conversations. Applying Large Language Models (LLMs) for this specific task presents two primary challenges: the accurate extraction of multiple elements and the understanding of complex dialogue reply structure. To tackle these issues, we propose a novel LLM-based multi-task approach, named T ask- a ware Co ntrastive M ixture o f E xperts ( TaCoMoE ), to tackle the DiaASQ task by integrating expert-level contrastive loss within task-oriented mixture of experts layer. TaCoMoE minimizes the distance between the representations of the same expert in the semantic space while maximizing the distance between the representations of different experts to efficiently learn representations of different task samples. Additionally, we design a Graph-Centric Dialogue Structuring strategy for representing dialogue reply structure and perform non-opinion utterances detection to enhance the performance of quadruple extraction. Extensive experiments are conducted on the DiaASQ dataset, demonstrating that our method significantly outperforms existing parameter-efficient fine-tuning techniques in terms of both accuracy and computational efficiency. The code is available at https://github.com/he2720/TaCoMoE.

This paper presents a joint learning framework for fake news detection, introducing an Enhanced BERT model that integrates named entity recognition, relational feature classification, and Stance Detection through a unified multi-task approach. The model incorporates task-specific masking and hierarchical attention mechanisms to capture both fine-grained and high-level contextual relationships across headlines and body text. Cross-task consistency losses are applied to ensure coherence and alignment with external factual knowledge. We analyse the average distance from components to the centroid of a news sample to differentiate genuine information from falsehoods in large-scale text data effectively. Experiments on two FakeNewsNet datasets show that our framework outperforms state-of-the-art models, with accuracy improvements of 2.17% and 1.03%. These results indicate the potential for applications needing detailed text processing, like automatic summarisation and misinformation detection.

Argument mining in Chinese essays is a challenging task due to the diversity of argumentative structures. In NLPCC 2025 Shared Task 5, we explored the use of large language models for comprehensive argument analysis. For argumentative component detection, we decouple the task into two stages. We first perform coarse-grained classification, then refine the results with fine-grained classification, both accomplished through fine-tuning large language models. For argument relation identification, we apply automated prompt engineering, beginning with a manually designed seed prompt and iteratively expanding and testing candidate prompts on simple data to select and optimize the best template. Our approach enhances contextual understanding and achieves strong performance in both argumentative component detection and argument relation identification, with our method achieving first place in the evaluation.

Argument Mining is crucial for uncovering the logical structure and reasoning process within texts, particularly in educational scenarios such as argumentative essay analysis. The NLPCC 2025 Shared Task on Argument Mining in Chinese argumentative essays introduces two subtasks: Argument Component Detection (ACD) and Argument Relation Identification (ARI). In this paper, we propose two corresponding methods—Track1 and Track2—to address these challenges. For Track1, we design a fine-grained data augmentation strategy by refining category labels with domain-specific and contextual information, enabling Large Language Models (LLMs) to generate diverse, high-quality samples. These are then used to fine-tune models through LoRA to improve category understanding. For Track2, we introduce a knowledge-enhanced relation identification method that combines prompt-based LLM generation with rule-based scoring and filtering to improve relation precision. In both tracks, we apply a majority voting strategy across multiple strong LLMs to enhance robustness. Experimental results demonstrate that our approach achieves state-of-the-art performance, validating the effectiveness of the proposed framework for fine-grained argument mining in Chinese essays.