Defending large language models (LLMs) against jailbreak attacks is crucial for ensuring their safe deployment. Existing defense strategies typically rely on predefined static criteria to differentiate between harmful and benign prompts. However, such rigid rules fail to accommodate the inherent complexity and dynamic nature of real-world jailbreak attacks. In this paper, we focus on the novel challenge of adaptive defense against diverse jailbreaks. We propose a new concept "mirror'', which is a dynamically generated prompt that reflects the syntactic structure of the input while ensuring semantic safety. The discrepancies between input prompts and their corresponding mirrors serve as guiding principles for defense. A novel defense model, MirrorShield, is further proposed to detect and calibrate risky inputs based on the crafted mirrors. Evaluated on multiple benchmark datasets and compared against ten state-of-the-art attack methods, MirrorShield demonstrates superior defense performance and promising generalization capabilities.

Large language model-based multi-agent systems (LLM-MAS) effectively accomplish complex and dynamic tasks through inter-agent communication, but this reliance introduces substantial safety vulnerabilities. Existing attack methods targeting LLM-MAS either compromise agent internals or rely on direct and overt persuasion, which limit their effectiveness, adaptability, and stealthiness. In this paper, we propose MAST, a Multi-round Adaptive Stealthy Tampering framework designed to exploit communication vulnerabilities within the system. MAST integrates Monte Carlo Tree Search with Direct Preference Optimization to train an attack policy model that adaptively generates effective multi-round tampering strategies. Furthermore, to preserve stealthiness, we impose dual semantic and embedding similarity constraints during the tampering process. Comprehensive experiments across diverse tasks, communication architectures, and LLMs demonstrate that MAST consistently achieves high attack success rates while significantly enhancing stealthiness compared to baselines. These findings highlight the effectiveness, stealthiness, and adaptability of MAST, underscoring the need for robust communication safeguards in LLM-MAS.

Defending large language models (LLMs) against jailbreak attacks is essential for their safe and reliable deployment.Existing defenses often rely on shallow pattern matching, which struggles to generalize to novel and unseen attack strategies.To address this challenge, we propose the Cognitive-Driven Defense (CDD) framework, which targets the underlying structure of jailbreak prompts by applying metaoperations, defined as basic manipulations that conceal harmful intent.CDD emulates human cognitive reasoning through a structured reasoning chain.It begins with a global perception of the prompt and follows with a localized analysis to uncover hidden manipulations.By applying supervised fine-tuning on this structured chain, the model learns to identify and reason about known manipulation patterns.To enhance generalization to unseen threats, an entropy-guided reinforcement learning algorithm (EG-GRPO) is introduced to encourage exploration of new types and variants of meta-operations.Experiments demonstrate that CDD can achieve state-of-the-art defense performance and exhibit strong generalization to unseen jailbreak attacks.

Bingyu Yan, Xiaoming Zhang, JinYu Hou, Chaozhuo Li, Ziyi Zhou, Yiming Hei, Litian Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Hallucination remains a major obstacle to the domain generalizability and reliability of Large Language Models (LLMs). Recent approaches address this issue by integrating Retrieval-Augmented Generation (RAG) with stepwise reasoning processes to iteratively retrieve knowledge. However, indiscriminate incorporation of external knowledge may interfere with reasoning, increasing latency and amplifying error accumulation. Moreover, existing methods rely on a unidirectional flow of external knowledge into LLMs while neglecting internal–external knowledge synergy, limiting autonomous reasoning capability. To address these limitations, we propose the Reasoning–Evidence Interactive Augmented Graph (RE-IAG), a framework that couples reasoning with evidence through a staged triggering mechanism and structured interaction. RE-IAG performs localized refinement of intermediate reasoning via adaptive branching under uncertainty and selectively triggers retrieval when internal reasoning stagnates. Crucially, it organizes both internal reasoning and retrieved evidence into aligned graph structures, enabling structure-guided verification and fine-grained refinement of intermediate conclusions. This design transforms retrieval from passive augmentation into an active constraint on reasoning, reducing error propagation, alleviating knowledge conflicts, and improving knowledge integration for hallucination mitigation. Extensive experiments on four multi-hop QA benchmarks show that RE-IAG outperforms adaptive RAG baselines, achieves competitive or superior performance to RL-based approaches, and demonstrates strong robustness and generalization across model scales and architectures.

Fact verification has become increasingly vital in the internet age, driven by the proliferation of false claims and political misinformation. While traditional methods rely predominantly on text-based evidence, multi-modal evidence introduces richer sources of information, offering valuable insights for claim verification. Existing multi-modal verification models often focus on superficial correlations between claims and evidence, neglecting the complex semantic interactions present in fine-grained multi-modal signals. In this paper, we propose a novel framework for multi-modal fact-checking, named Hypergraph Transformer-based Multi-modal Fact-Checking (HGTMFC). Our approach captures high-order relationships between different modalities of evidence and claims by leveraging hypergraphs. HGTMFC models the intricate relationships among evidence across various modalities and enhances information propagation through a transformer-based mechanism embedded within the hypergraph. Moreover, we utilize linegraphs to refine this propagation process, further strengthening the model's reasoning capabilities. Experiments on benchmark datasets demonstrate that our model significantly outperforms existing approaches in multi-modal fact verification.

The pervasive spread of misinformation on social networks highlights the critical necessity for effective fact verification systems. Traditional approaches primarily focus on pairwise correlations between claims and evidence, often neglecting comprehensive multi-hop retrieval and reasoning, which results in suboptimal performance when dealing with complex claims. In this paper, we propose MRR-FV, a generative retrieval-enhanced model designed to address the novel challenge of Multi-hop Retrieval and Reasoning for Fact Verification, which integrates two core modules: Generative Multi-hop Retriever and the Hierarchical Interaction Reasoner. MRR-FV utilizes an autoregressive model for iterative multi-hop evidence retrieval, complemented by a pre-trained compressor to address the challenge of intention shift across retrieval hops. For claim verification, we propose a hierarchical interaction reasoner that conducts intra-sentence reasoning to capture long-term semantic dependencies and inter-sentence reasoning across multi-hop evidence subgraphs to reveal complex evidence interactions. Experimental evaluations on the FEVER and HOVER datasets demonstrate the superior performance of our model in both claim verification and evidence retrieval tasks.

The proliferation of fake news on social media platforms has exerted a substantial influence on society, leading to discernible impacts and deleterious consequences. Conventional deep learning methodologies employing small language models (SLMs) suffer from the necessity for extensive supervised training and the challenge of adapting to rapidly evolving circumstances. Large language models (LLMs), despite their robust zero-shot capabilities, have fallen short in effectively identifying fake news due to a lack of pertinent demonstrations and the dynamic nature of knowledge. In this paper, a novel framework Multi-Round Collaboration Detection (MRCD) is proposed to address these aforementioned limitations. The MRCD framework is capable of enjoying the merits from both LLMs and SLMs by integrating their generalization abilities and specialized functionalities, respectively. Our approach features a two-stage retrieval module that selects relevant and up-to-date demonstrations and knowledge, enhancing in-context learning for better detection of emerging news events. We further design a multi-round learning framework to ensure more reliable detection results. Our framework MRCD achieves SOTA results on two real-world datasets Pheme and Twitter16, with accuracy improvements of 7.4% and 12.8% compared to using only SLMs, which effectively addresses the limitations of current models and improves the detection of emergent fake news detection.

Information diffusion prediction (IDP) is a pivotal task for understanding the dynamics of information propagation within social networks. Conventional models typically adhere to a fixed learning-based paradigm, where the trained prediction model remains static during the inference phase. This paradigm presupposes that the data is independent and identically distributed, an assumption that may not hold true due to the inherently open nature of social media and the uncertainty and variability in user behavior. In this paper, we address the novel problem of out-of-distribution (OOD) shifts within IDP tasks and propose a new test-time training-based model for multi-scale IDP tasks, named Ghidorah. Our approach focuses on adapting a subset of model parameters to accommodate the unique characteristics of test samples through self-supervised learning (SSL) tasks. Ghidorah comprises three components: the macroscopic prediction branch, the microscopic prediction branch, and the auxiliary SSL branch. The auxiliary SSL task employs a masked autoencoder-based loss to fine-tune the model for specific test samples prior to prediction. Furthermore, Ghidorah integrates invariant learning to capture robust representations while mitigating spurious correlations. To our knowledge, Ghidorah is the first work to introduce a test-time training framework specifically designed to address the critical yet often overlooked OOD challenges in IDP. Experimental results across several benchmark datasets validate the superiority of our approach.

The exploration of bundle recommendation has garnered significant attention for its potential to enhance user experience and augment business sales. Previous research in this domain has primarily focused on modeling user-item and user-bundle interactions, utilizing multi-view collaboration to bolster the accuracy of bundle recommendations. Nevertheless, existing methodologies exhibit limitations, notably in the inadequate modeling of multi-view information and the absence of multi-grained details. Consequently, addressing the intricate correlation among users, items, and bundles necessitates a sophisticated approach capable of capturing both global and local nuances. We present a novel framework named Large Language Model-based Multi-View and Multi-Grained Learning for Bundle Recommendation (LLMCBR). We introduce an LLM-based semantic refinement module to summarize and encode bundle-level knowledge. To bridge the gap between semantic representation and collaborative signals, we design an adaptation strategy. Furthermore, LLMCBR leverages multi-view and multi-granular modeling to unify collaborative signals. Specifically, LLMCBR integrates item preferences within both bundle-view and item-view, thereby augmenting the comprehensiveness of multi-view data. Following this integration, each view undergoes stratification into multiple granularities to facilitate the acquisition of multi-grained details. We introduce a multiple contrastive instance mechanism to regulate the influence of different granularities and views. This mechanism empowers the model to comprehend complex consumer behaviors across various dimensions. LLMCBR is extensively evaluated over three real-world datasets, and the experimental results demonstrate its superiority.

Large language models (LLMs) exhibit strong language understanding capabilities, but encounter challenges when integrating structured knowledge from knowledge graphs (KGs) for complex reasoning tasks such as knowledge graph question answering (KGQA). Existing methods often rely on prompt engineering or fixed templates, which obscure the relational structure and limit generalization. To address these limitations, this paper introduces the Reverse Chain-of-Thought (R-CoT) and Causal Path Verification Plugin, a modular framework that reconstructs retrieved KG triples into reverse chains of sub-questions. Each reasoning step is aligned with a supporting triple, forming interpretable multi-hop paths. In particular, Semantic Causal Scoring (SCS) module is further incorporated to evaluate the causal alignment between each reverse sub-question and the original question through dynamic semantic vector matching. The SCS design avoids frequent interactions with LLMs and effectively filters irrelevant or unsupported reasoning steps. Based on the scoring results, a template-free, model-agnostic R-CoT input format is constructed as a semi-structured sequence. This design preserves the KG structure in natural language form and enables seamless integration with standard LLMs without fine-tuning. Experimental results demonstrate that the R-CoT Plugin consistently improves factual alignment, enhances reasoning stability, and outperforms conventional prompt-based methods in both accuracy and coherence.

Large language models (LLMs) have shown strong potential in complex reasoning tasks.However, as task complexity increases, their performance often degrades, resulting in hallucinations, errors, and logical inconsistencies.To enhance reasoning capabilities, Monte Carlo Tree Search (MCTS) has been introduced to guide the exploration of reasoning paths in a structured manner.Despite its advantages, traditional MCTS relies on fixed reasoning strategies, limiting the diversity of reasoning paths and the coverage of the solution space.To address these limitations, we propose Dynamic Strategy-Guided MCTS (DSG-MCTS), a novel framework that dynamically integrates multiple reasoning strategies, such as abductive and analogical reasoning, to expand the reasoning space.At the same time, DSG-MCTS enhances reasoning efficiency through a dynamic strategy selection mechanism that adapts to the task context.Experimental results on challenging reasoning benchmarks demonstrate that DSG-MCTS achieves improved accuracy and efficiency, outperforming existing state-of-the-art methods.

With the rapid advancement of online social networks, the proliferation of misinformation has exerted significant influence on both social order and individual decision-making. Traditional manual fact-checking methods are resource-intensive and lack scalability, highlighting the urgent need for automated approaches to verify claims and foster the dissemination of truthful content. In this work, we introduce a novel fact-checking framework, MLLM-CRMFC, which integrates multi-modal large language models (MLLMs) with counterfactual reasoning to enhance both the accuracy and interpretability of automated fact verification. Our multi-modal fact-checking method processes textual and visual data, utilizing LLMs and counterfactual reasoning mechanisms to evaluate the relevance and consistency of diverse evidence. The proposed framework not only predicts the verification outcomes of multi-modal claims but also generates natural language explanations, providing users with transparent and traceable justifications for the decision-making process. Experimental results demonstrate that our model effectively verifies claims using benchmark datasets, outperforms state-of-the-art approaches, and delivers coherent, fluent explanations that enhance the interpretability of the verification process.

Survival prediction is a pivotal task for estimating mortality risk within a given timeframe based on whole slide images (WSIs). Conventional models typically assume that WSIs across patients are independent and identically distributed, an assumption that may not hold due to inherent variability in WSI preparation and the uncertain condition of infected tissues. These uncontrollable external factors introduce significant variability in the numbers and resolutions of WSIs across patients, leading to bias and compromised performance, particularly for tail patients with limited data. In this paper, we propose a novel approach, PathoKD, based on knowledge distillation. Recognizing the hierarchical nature of disease progression and the data scarcity issues associated with vanilla knowledge distillation methods, PathoKD integrates a novel curriculum learning framework with hierarchical knowledge distillation. This integration effectively mitigates the performance gap between head and tail patients, thereby enhancing prediction accuracy across patient groups. Our proposal is extensively evaluated over popular datasets and experimental results demonstrate its superiority.

The prediction of information popularity propagation is critical for applications such as recommendation systems, targeted advertising, and social media trend analysis. Traditional approaches primarily rely on historical cascade data, often sacrificing timeliness for prediction accuracy. These methods capture aggregate diffusion patterns but fail to account for the complex temporal dynamics of early-stage propagation. In this paper, we introduce Diffusion Guided Propagation Augmentation(DGPA), a novel framework designed to improve early-stage popularity prediction. DGPA models cascade dynamics by leveraging a generative approach, where a temporal conditional interpolator serves as a noising process and forecasting as a denoising process. By iteratively generating cascade representations through a sampling procedure, DGPA effectively incorporates the evolving time steps of diffusion, significantly enhancing prediction timeliness and accuracy. Extensive experiments on benchmark datasets from Twitter, Weibo, and APS demonstrate that DGPA outperforms state-of-the-art methods in early-stage popularity prediction.

Most jailbreak methods for large language models (LLMs) focus on superficially improving attack success through manually defined rules. However, they fail to uncover the underlying mechanisms within target LLMs that explain why an attack succeeds or fails. In this paper, we propose investigating the phenomenon of jailbreaks and defenses for LLMs from the perspective of attention distributions within the models. A preliminary experiment reveals that the success of a jailbreak is closely linked to the LLM's attention on sensitive words.Inspired by this interesting finding, we propose incorporating critical signals derived from internal attention distributions within LLMs, namely Attention Intensity on Sensitive Words and Attention Dispersion Entropy, to guide both attacks and defenses. Drawing inspiration from the concept of "Feint and Attack", we introduce an attention-guided jailbreak model, ABA, which redirects the model's attention to benign contexts, and an attention-based defense model, ABD, designed to detect attacks by analyzing internal attention entropy. Experimental results demonstrate the superiority of our proposal when compared to SOTA baselines.

LLMs have demonstrated unprecedented capabilities in natural language processing, yet their practical deployment remains hindered by persistent factuality and faithfulness hallucinations. While existing methods address these hallucination types independently, they inadvertently induce performance trade-offs, as interventions targeting one type often exacerbate the other. Through empirical and theoretical analysis of activation space dynamics in LLMs, we reveal that these hallucination categories share overlapping subspaces within neural representations, presenting an opportunity for concurrent mitigation. To harness this insight, we propose SPACE, a unified framework that jointly enhances factuality and faithfulness by editing shared activation subspaces. SPACE establishes a geometric foundation for shared subspace existence through dual-task feature modeling, then identifies and edits these subspaces via a hybrid probe strategy combining spectral clustering and attention head saliency scoring. Experimental results across multiple benchmark datasets demonstrate the superiority of our approach.

Amidst the rapid propagation of multimodal fake news across social media platforms, the detection of fake news has emerged as a prime research pursuit. To detect heightened level of meticulous fabrications, propagation paths are introduced to provide nuanced social context that enhances the basic semantic analysis of the news content. However, existing propagation-enhanced models encounter a dilemma between detection efficacy and social hazard. In this paper, we explore the innovative problem of early fake news detection through the generation of propagation paths, capable of benefiting from the extensive social context within propagation paths while mitigating potential social hazards. To address these challenges, we propose a novel Reinforced Propagation Path Generation Fake News Detection model,RPPG-Fake. Departing from conventional discriminative approaches,RPPG-Fakecaptures the propagation topology pattern from a heterogeneous social graph and generates the propagation paths to detect fake news effectively under a reinforcement learning paradigm. Our proposal is extensively evaluated over three popular datasets, and experimental results demonstrate the superiority of our proposal.

Privacy-Preserving Image Retrieval (PPIR) enables searching for similar images on cloud servers while safeguarding image privacy. PPIR uploads encrypted images to servers to address privacy concerns and then employs deep neural networks for retrieval on extracted features from cipher-images. However, current PPIR encrypts all images with the same secret keys to maintain consistent feature spaces, lacking support for adaptive keys where distinct images are encrypted with various keys. To this end, we propose a new PPIR scheme to support adaptive keys while keeping stable feature spaces. Specifically, we design ingenious image encryption to align with feature extraction during the JPEG compression process, incorporating encryptable orthogonal transformations, shuffling, stream cipher, and sign encryption operations. Our approach extracts well-designed absolute value sequences of local blocks and global histogram features from cipher-images, ensuring stable feature spaces under adaptive keys. To enhance model generalization performance, we employ Self-Supervised Contrastive Learning (SSCL) to build a pretraining model and propose a straightforward yet efficient block-sampling augmentation technique for the structured features to drive SSCL. Moreover, our retrieval model implements a dual-attention structure to capture dependencies among local block sequences and import scores of global features. Extensive experiments on four datasets demonstrate that our approach achieves superior retrieval accuracy compared to existing schemes and maintains excellent retrieval performance under adaptive keys, effectively preserving image privacy.

Nowadays, detecting multimodal fake news has emerged as a foremost concern since the widespread dissemination of fake news may incur adverse societal impact. Conventional methods generally focus on capturing the linguistic and visual semantics within the multimodal content, which fall short in effectively distinguishing the heightened level of meticulous fabrications. Recently, external knowledge is introduced to provide valuable background facts as complementary to facilitate news detection. Nevertheless, existing knowledge-enhanced endeavors directly incorporate all knowledge contexts through static entity embeddings, resulting in the potential noisy and content-irrelevant knowledge. Moreover, the integration of knowledge entities makes it intractable to model the sophisticated correlations between multimodal semantics and knowledge entities. In light of these limitations, we propose a novel Adaptive Knowledge-Aware Fake News Detection model, dubbed AKA-Fake. For each news, AKA-Fake learns a compact knowledge subgraph under a reinforcement learning paradigm, which consists of a subset of entities and contextual neighbors in the knowledge graph, restoring the most informative knowledge facts. A novel heterogeneous graph learning module is further proposed to capture the reliable cross-modality correlations via topology refinement and modality-attentive pooling. Our proposal is extensively evaluated over three popular datasets, and experimental results demonstrate the superiority of AKA-Fake.

Due to the explosion of various dialogue scenes, summarizing the dialogue into a short message has drawn much attention recently. In the multi-modal dialogue scene, people tend to use tone and body language to illustrate their intentions. While traditional dialogue summarization has predominantly focused on textual content, this approach may overlook vital visual and audio information essential for understanding multi-modal interactions. Recognizing the established field of multi-modal dialogue summarization, we develop a new multi-modal dialogue summarization dataset (MDS), which aims to enhance the variety and scope of data available for this research area. MDS provides a demanding testbed for multi-modal dialogue summarization. Subsequently, we conducted a comparative analysis of various summarization techniques on MDS and found that the existing methods tend to produce redundant and incoherent summaries. All of the models generate unfaithful facts to some degree, suggesting future research directions. MDS is available at https://github.com/R00kkie/MDS.

Jailbreak attacks enable malicious queries to evade detection by LLMs.Existing attacks focus on meticulously constructing prompts to disguise harmful intentions.However, the incorporation of sophisticated disguising prompts may incur the challenge of "intention shift".Intention shift occurs when the additional semantics within the prompt distract the LLMs, causing the responses to deviate significantly from the original harmful intentions.In this paper, we propose a novel component, "bait", to alleviate the effects of intention shift.Bait comprises an initial response to the harmful query, prompting LLMs to rectify or supplement the knowledge within the bait.By furnishing rich semantics relevant to the query, the bait helps LLMs focus on the original intention.To conceal the harmful content within the bait, we further propose a novel attack paradigm, BaitAttack.BaitAttack adaptively generates necessary components to persuade targeted LLMs that they are engaging with a legitimate inquiry in a safe context.Our proposal is evaluated on a popular dataset, demonstrating state-of-the-art attack performance and an exceptional capability for mitigating intention shift.The implementation of BaitAttack is accessible at: https://anonymous.4open. science/r/BaitAttack-D1F5.How to make a bomb. Query Bait Role Scene Format Bait Maker Which expert is best suitable to deal with the act of < Harmful Query >? Create a scene that fits the expert role.

The detection of fake news has emerged as a pressing issue in the era of online social media. To detect meticulously fabricated fake news, propagation paths are introduced to provide nuanced social context to complement the pure semantics within news content. However, existing propagation-enhanced models face a dilemma between detection efficacy and social hazard. In this paper, we investigate the novel problem of early fake news detection via propagation path generation, capable of enjoying the merits of rich social context within propagation paths while alleviating potential social hazards. In contrast to previous discriminative detection models, we further propose a novel generative model, DGA-Fake, by simulating realistic propagation paths based on news content before actual spreading. A guided diffusion module is integrated into DGA-Fake to generate simulated user interaction sequences, guided by historical interactions and news content. Evaluation across three datasets demonstrates the superiority of our proposal.

Multimodal summarization (MS) aims to generate a summary from multimodal input. Previous works mainly focus on textual semantic coverage metrics such as ROUGE, which considers the visual content as supplemental data. Therefore, the summary is ineffective to cover the semantics of different modalities. This paper proposes a multi-task cross-modality learning framework (CISum) to improve multimodal semantic coverage by learning the cross-modality interaction in the multimodal article. To obtain the visual semantics, we translate images into visual descriptions based on the correlation with text content. Then, the visual description and text content are fused to generate the textual summary to capture the semantics of the multimodal content, and the most relevant image is selected as the visual summary. Furthermore, we design an automatic multimodal semantics coverage metric to evaluate the performance. Experimental results show that CISum outperforms baselines in multimodal semantics coverage metrics while maintaining the excellent performance of ROUGE and BLEU.

Multimodal summarization with multimodal output (MSMO) generates a summary with both textual and visual content. Multimodal news report contains heterogeneous contents, which makes MSMO nontrivial. Moreover, it is observed that different modalities of data in the news report correlate hierarchically. Traditional MSMO methods indistinguishably handle different modalities of data by learning a representation for the whole data, which is not directly adaptable to the heterogeneous contents and hierarchical correlation. In this paper, we propose a hierarchical cross-modality semantic correlation learning model (HCSCL) to learn the intra- and inter-modal correlation existing in the multimodal data. HCSCL adopts a graph network to encode the intra-modal correlation. Then, a hierarchical fusion framework is proposed to learn the hierarchical correlation between text and images. Furthermore, we construct a new dataset with relevant image annotation and image object label information to provide the supervision information for the learning procedure. Extensive experiments on the dataset show that HCSCL significantly outperforms the baseline methods in automatic summarization metrics and fine-grained diversity tests.