Large language models (LLMs) have demonstrated remarkable capabilities in many fields, especially in complex neural lagnuage processing tasks. Despite their impressive performance, the content generated by LLMs still suffers from the problem of hallucination, particularly in tasks that require real-time data or specialized domain knowledge. Knowledge graphs and multimodal unstructured data serve as important sources of knowledge that can help address the hallucination issues in LLMs. However, existing methods mostly utilize knowledge graphs or multimodal unstructured data in isolation, neglecting the interaction between the two and it is the interaction that contributes to the extraction of deep knowledge in the knowledge base. In this paper, we propose a novel framework called the Collaborative Framework of Multimodal Unstructured Data and Knowledge Graph (CoMuS-KG). This framework enhances the reasoning capabilities of LLMs by enabling interaction between multimodal unstructured data and knowledge graphs, extracting deep knowledge from unstructured data, and completing missing information in knowledge graphs. Specifically, CoMuS-KG first decompose the question posed to the LLMs into multiple sub-questions and convert these sub-questions into knowledge graph triplets with missing head entity, tail entity, or relation. And then the knowledge graph and multimodal unstructured data are used to complete these triplets. Finally, we use the completed triplets to answer the original question and the completed triplets can be updated back into the knowledge graph to assist in other reasoning tasks. Extensive experiments on three KGQA benchmark datasets demonstrate the question-answering performance and reasoning capabilities of CoMuS-KG. Our code is publicly available at: https://github.comlGuChongAnlCoMuS-KG

We propose a Dual-Layer Meta-Learning Network for Few-Shot Named Entity Recognition, where the network can selectively retain positive training signals from the memory chain to enhance the meta-model's learning capability and filter out interference from non-positive signals. Additionally, to mitigate the parameter explosion caused by the dual-layer network, we further use Chebyshev polynomials to fit the token classification function for entity span detection and employ the Kolmogorov-Arnold Network to fit the prototype-oriented classification function for entity span classification. This effectively reduces the runtime and GPU usage of the dual-layer structure.

The Relational Triple Extraction (RTE) is a fundamental and essential task in information extraction and knowledge graph construction. Recently, the table filling RTE methods have attracted more and more attention due to its good performance. However, there are still some problems with this kind of methods, such as only focusing on local features and not making full use of the regional information of triples. To overcome these deficiencies, we propose a Supervised contrastive learning and Global feature-oriented based Object detection framework for Relational Triple Extraction (SGORTE). Specifically, we convert the table filling RTE task into an object detection task, introduce multiple positive examples and a penalty term through a designed supervised contrastive learning method to enhance the robustness of the framework. In addition, we combine vertices-based bounding box detection and global relational region detection to fully utilize the relevant information of the triples for extraction. We conduct extensive experiments on two widely used datasets, and the experimental results show that the proposed framework performs better than the state-of-the-art baselines, and has obvious performance improvements in a variety of complex scenarios.

Optimizing the efficiency and adaptability of large language models (LLMs) for diverse downstream tasks remains a critical challenge. We propose FlexFFN (Flexible Feedforward Network), a novel framework that introduces hierarchical dynamic selection to enhance computational efficiency, flexibility, and performance in LLMs. At the macro level, FlexFFN leverages a Mixture of Experts (MoE) architecture to dynamically activate distinct FFN modules based on input characteristics. At the micro level, within each FFN module, a dynamic switching mechanism selects between KAN and traditional MLP, combining the rapid convergence capabilities of MLPs with the compositional learning and interpretability strengths of KAN. Additionally, FlexFFN integrates QLoRA (Quantized Low-Rank Adaptation) to significantly reduce memory requirements and computational costs during fine-tuning. By introducing these innovations, FlexFFN achieves a fine-grained balance between computational cost and model expressiveness, making it well-suited for large-scale training and deployment. Experimental results demonstrate that FlexFFN outperforms traditional architectures by reducing computational overhead while improving task-specific adaptability and model efficiency.

In this paper, we propose IKG-Agent, an intent-driven knowledge graph agent framework designed to perform adaptive workflow reasoning for complex question answering. Unlike traditional approaches, IKG-Agent integrates semantic parsing (SP), subgraph retrieval (SR), and large language models (LLMs) into a unified system. When a user submits a query, IKG-Agent first identifies the underlying intent and constructs a tailored reasoning workflow. Based on task complexity and query requirements, the framework dynamically selects optimal reasoning paths and tools, adjusting workflows during execution. By leveraging a shared knowledge memory system to continuously evaluate information sufficiency at each step, IKG-Agent mitigates error accumulation in traditional SP/SR-based reasoning—particularly for long relational paths and complex multi-hop inference. Experimental results demonstrate that IKG-Agent outperforms state-of-the-art methods on multiple public datasets, achieving significant improvements in accuracy and reliability for tasks requiring multi-level reasoning. Our code and data will be publicly released.

Recommender systems have been widely adopted in real-world applications, yet they still face challenges in addressing the cold-start problem. Cross-domain recommendation (CDR) offers a promising solution to the cold-start problem by transferring user preferences across domains. However, existing methods often rely on overlapping users or biased embeddings, limiting their generalization and fairness. We propose PANDA-CDR, a unified CDR framework that integrates perturbation-guided contrastive bottleneck and dual-channel disentanglement to learn robust transferable semantics. To enhance fairness, especially under long-tail distributions, we further introduce adversarial domain alignment and exposure-aware reweighting to mitigate popularity bias. Extensive experiments on real-world benchmarks show that PANDA-CDR achieves state-of-the-art performance while improving cold-start and long-tail recommendation fairness in sparse, low-overlap scenarios.

Knowledge graph embedding, which aims to learn representations of entities and relations in large scale knowledge graphs, plays a crucial part in various downstream applications. The performance of knowledge graph embedding models mainly depends on the ability of modeling relation patterns, such as symmetry/antisymmetry, inversion and composition (commutative composition and non-commutative composition). Most existing methods fail in modeling the non-commutative composition patterns. Several methods support this kind of pattern by modeling in quaternion space or dihedral group. However, extending to such sophisticated spaces leads to a substantial increase in the amount of parameters, which greatly reduces the parameter efficiency. In this paper, we propose a new knowledge graph embedding method called dual complex number knowledge graph embeddings (DCNE), which maps entities to the dual complex number space, and represents relations as rotations in 2D space via dual complex number multiplication. The non-commutativity of the dual complex number multiplication empowers DCNE to model the non-commutative composition patterns. In the meantime, modeling relations as rotations in 2D space can effectively improve the parameter efficiency. Extensive experiments on multiple benchmark knowledge graphs empirically show that DCNE achieves significant performance in link prediction and path query answering.

News recommendation is critical for online news services. How to precisely match news content with users’ interests lies in the core of personalized news recommendation. Existing methods mainly learn a unified embedding vector for each user to represent his/her interests. However, users’ diverse interests can not be expressed adequately by a single embedding representation because of the lack of expressiveness. Additionally, incorporating news popularity into news recommendation can effectively improve accuracy since users with different interests are drawn to current popular news. In this paper, we propose a news recommendation method with multi-interest and popularity-aware modeling, named MIRec. We propose a novel news encoder with attentive learning to obtain unified representations of clicked news from the content and popularity. Furthermore, we exploit a popularity-aware multi-interest extractor to generate multi-interest representations of users and eliminate the bias of news popularity in preference modeling. Besides, we design a popularity predictor for candidate news, which measures its popularity based on the popularity features, recency, and interaction rate. Finally, we adopt a gated mechanism based on user multi-interest representation and the popularity of candidate news to make recommendations. Extensive experiments on large-scale benchmark dataset demonstrate that our approach significantly outperforms existing state-of-the-art methods.

NLP models are shown to be vulnerable to adversarial examples. The usual attack methods in NLP fields mainly focus on word-level perturbations. However, the word-substitution based method is not suitable for short text. Short texts are more susceptible to word substitution than long texts, which makes semantic shifting more likely to occur, and the number of words in short texts can be modified is small, making the attack difficult to succeed and hard to guarantee naturality and fluency. To tackle the above problems, we present Polymorphic Adversarial Examples Generation (PAEG) attack, a generative method by combining pre-trained language model BERT and Variational Autoencoder. Compared to attack methods proposed in previous literature, the proposed approach can not only generate polymorphic adversarial examples with different forms but also improve the attack success rate significantly on two popular datasets. Our codes are released at https://github.com/YilingLiang/vMF-VAE-Bert.

Knowledge graph embedding, which aims to learn representations of entities and relations in knowledge graphs, finds applications in various downstream tasks. The key to success of knowledge graph embedding models are the ability to model relation patterns including symmetry/antisymmetry, inversion, commutative composition and non-commutative composition. Although existing methods fail in modeling the non-commutative composition patterns, several approaches support this pattern by modeling beyond Euclidean space and complex space. Nevertheless, expanding to complicated spaces such as quaternion can easily lead to a substantial increase in the amount of parameters, which greatly reduces the computational efficiency. In this paper, we propose a new knowledge graph embedding method called RotateCT, which first transforms the coordinates of each entity, and then represents each relation as a rotation from head entity to tail entity in complex space. By design, RotateCT can infer the non-commutative composition patterns and improve the computational efficiency. Experiments on multiple datasets empirically show that RotateCT outperforms most state-of-the-art methods on link prediction and path query answering.

Link prediction is to predict missing relations between entities for Knowledge Graph Completion (KGC). Convolution neural network has been used in much previous work on link prediction to capture fundamental data pattern of knowledge graph. However, because these models use low-dimensional convolution operation, which limits their performance, they learn fewer expressive features. Further more, they do not have the the capability of keeping the translation property of knowledge triplet, which is an important property for knowledge reasoning. Focusing on these problems, we propose Conv3D (3D Convolution Embedding), a neural network model for link prediction that uses 3D convolution. To capture deeper feature interactions in the knowledge graph, we employ 3D convolution instead of shallow 1D or 2D convolution for generating triplet scores. We conduct link prediction experiments on four general datasets (WN18, WN18RR, FB15k, FB15k-237) and get state-of-the-art (SOTA) results on WN18 and WN18RR. We also explore the influence of convolution parameters (reshaping dimension, number of filters, kernel size) on FB15k-237 and obtain quantitative analytical findings.

Knowledge representation learning is an automatic learning technique that can embeds a knowledge graph into a low-dimensional vector space. With use of this, knowledge becomes computable and various intelligent applications can be realized. Traditional semantic discriminating models suggest that the embeddings of entities should depend on the specific semantic environment. We find that the multiple latent information of relations has not been put to use by these models. In this paper, a gated semantic discriminating model (GSDM) is proposed to select useful latent information and neglect useless information according to the specific semantic environment for both entities and relations. Experiments show that GSDM achieves better performance than related state-of-the-art baselines on most indicators. The better trade-off between the discriminate parameter pressure and the model performance has proved the correctness and feasibility of semantic discriminating mechanism to some extent.

Transformer-based pre-trained language models (PLMs) mostly suffer from excessive overhead despite their advanced capacity.For resource-constrained devices, there is an urgent need for a spatially and temporally efficient model which retains the major capacity of PLMs.However, existing statically compressed models are unaware of the diverse complexities between input instances, potentially resulting in redundancy and inadequacy for simple and complex inputs.Also, miniature models with early exiting encounter challenges in the trade-off between making predictions and serving the deeper layers.Motivated by such considerations, we propose a collaborative optimization for PLMs that integrates static model compression and dynamic inference acceleration.Specifically, the PLM is slenderized in width while the depth remains intact, complementing layer-wise early exiting to speed up inference dynamically.To address the trade-off of early exiting, we propose a joint training approach that calibrates slenderization and preserves contributive structures to each exit instead of only the final layer.Experiments are conducted on GLUE benchmark and the results verify the Pareto optimality of our approach at high compression and acceleration rate with 1/8 parameters and 1/19 FLOPs of BERT.

The real-world image degradation in the super-resolution task is recently considered as a combination of Gaussian blur, down-sampling, and additional white Gaussian noise. To han-dle this degradation, previous methods estimate the Gaussian blur kernel or model the degradation based on a randomly selected image patch. However, these methods cannot han-dle degradations with high-level noise well as they ignore the spatial variability or even the existence of noise. Moreover, using image denoising networks to preprocess low-resolution images also fails due to the loss of important high-frequency information. In this paper, we propose a framework called EASE to flexibly handle real-world degradations. Specifi-cally, we develop a lightweight module to erase noise and blur simultaneously by learning from an image denoising and an image restoration network, which adapts to existing net-works that focus on handling bicubic down-sampling. Exten-sive experiments prove the superiority of our method, espe-cially when handling degradations with high-level noise.

Blind super-resolution task aims to restore low-resolution im“ages with unknown degradations to high-resolution counter-parts. Existing methods rely on degradations estimation to re-construct high-resolution images. However, they need human involvement to obtain the best results as they treat unknown types of degradations as known conditions and manually select corresponding trained models. Moreover, they cannot fully use estimated degradations and generate blurry artifacts as they ignore that the impact of degradations on images is re-lated to images contents. In this paper, we propose HIS-NEST which contains an automatic search strategy HIS and a net-work structure NEST. Specifically, to bypass manual partici-pation, HIS automatically selects the clearest image by esti-mating the qualities of generated images. Furthermore, NEST protects the connection between degradations and images by using no loss functions to limit the degradations estimation and analyzing degradations from the perspective of channel and space. Extensive experiments show that our method out-performs state-of-the-art methods.

The traditional knowledge representation learning (KRL) models treat each triplet in a knowledge base independently, so they can not make full use of the neighborhood information across triplets. KRL models based on graph attention networks (GAT) can not only capture feature interactions across triplets, but also further distinguish the importance of neighbor entities. Recently, we find that there are two flaws in GAT-based KRL models: (1) Ignoring the bidirectionality of information flows leads to insufficient utilization of entity neighborhood information. When encapsulating the neighborhood information, only the forward information flows flowing into the target entity are considered, but the backward information flows flowing out are neglected. (2) The unified update process for all relations causes the useful information related the few-shot relations to be diluted. We propose a multiple attention-based model considering bidirectional entity information flows and few-shot relations (BEFSR). In our model, a GAT-based attention framework is used to integrate forward information flows and backward information flows of each triplet respectively to capture feature interactions across triplets, and an LSTM-based attention framework is adopted to gradually aggregate the entity-pair information for few-shot relations’ updating. In BEFSR, entities and relations can be updated more appropriately. Experiments demonstrate that BEFSR outperforms state-of-the-art KRL models in knowledge base completion task.

Abstract DPA Contest is a world-famous side-channel competition aiming at analyzing and evaluating the implementing security of some latest countermeasures. Improved Rotating S-box Masking Scheme (RSM2.0) is one of the most popular countermeasures designed during DPA Contest V4.2, which arms with both Low Entropy Masking Schemes and shuffling strategy to ensure the software security of AES-128, particularly the non-profiled security. Up to now, conducting high efficient non-profiled attacking scheme with low resource costs is still a challenge. In this paper, we first propose general and non-profiled leakage fingerprint attacks (named NP-LFA) for secret cracking and make use of it to crack RSM2.0 random masks with almost 100% accuracy. Further, we analyze the hidden vulnerabilities embedded in RSM2.0 implementation, and utilize them to bypass the shuffling defense and perform the master key recovery. Official evaluation results show that NP-LFA is capable of compromising RSM2.0 within 14 traces, each of which only costs 60 ms processing time. Such result validates the high efficiency and light-weighted characteristics of our attacking scheme, which has ranked the first in the official website till now. In addition, we discuss and put forward some possible strategies to mitigate our NP-LFA threats.

Video super resolution (SR) is an important task, which recovers high-resolution (HR) frames from consecutive low-resolution (LR) couterparts. The most advanced works achieved good performance to this day. However, most of them has largely focussed on making a breakthrough in accuracy and speed, which has neglect that how to recover the finer texture details. Therefore, in this paper, we first present an Video SR model combined generative adversarial network and recurrent neural network (GAN-RNN) structure. It is forced by the self-attention mechanism to pay great attention to the high-frequency information of the LR frames. The perceptual loss is introduced to retain the high-frequency detail which is different from other video SR network. A great deal of evaluations and comparisons with previous methods have confirmed the merits of the proposed framework which can significantly outperform the current state of the art.

Single image dehazing is challenging without scene airlight and transmission map. Most of existing dehazing algorithms tend to estimate key parameters based on manual designed priors or statistics, which may be invalid in some scenarios. Although deep learning-based dehazing methods provide an effective solution, most of them rely on paired training datasets, which are prohibitively difficult to be collected in real world. In this paper, we propose an effective end-to-end generative adversarial network for single image dehazing, named SIDGAN. The proposed SIDGAN adopts a U-net architecture with a novel color-consistency loss derived from dark channel prior and perceptual loss, which can be trained in an unsupervised fashion without paired synthetic datasets. We create a RealHaze dataset for network training, including 4,000 outdoor hazy images and 4,000 haze-free images. Extensive experiments demonstrate that our proposed SIDGAN achieves better performance than existing state-of-the-art methods on both synthetic datasets and real-world datasets in terms of PSNR, SSIM, and subjective visual experience.

Anomalies detection is to spot the objects whose patterns singularly differ from the reference majority. Attributed networks often contain node attributes and network structure, which are widely used for real-life applications. Meanwhile, how to detect anomalies on attributed networks has caused a lot of attention. Most existing works on anomaly detection attempt to incorporate node attributes with the network structure. However, there may exist structurally irrelevant attributes in the networks, which may have adverse effects on the detection results. Besides, the heterogeneity of node attributes and network structure may further make the detection of anomalies difficult. In order to overcome the above challenges, in this paper, we propose a Robust Embedding Method for Anomaly Detection on Attributed Networks, called REMAD. Methodologically, the proposed REMAD combines network embedding and residual analysis together. By performing network embedding on the network, REMAD obtains the representative attributes that are closely coherent with the network structure. Simultaneously, by adopting the residual analysis, REMAD characterizes and analyzes the residuals of attribute information to discover anomalies. Experimental results on both synthetic and real-world datasets demonstrate the advantages of our proposed method REMAD against the state-of-the-art anomaly detection methods.

Network embedding, which aims to learn the low-dimensional representations of nodes, has attracted increasing attention in various fields such as social networks, paper citation networks and knowledge graphs. At present, most of the network embedding works are based on static networks, that is, the evolution of networks over time is not taken into account. It is more realistic to consider temporal information in network embedding and it could also make the embedding get more abundant information. In this paper, we propose a dynamic network embedding model DynSEM with semantic evolution, to train node embeddings in a sequence of networks over time. The advantage of our method is that it presents an effective inheritance of historical information. Our method uses nonrandom initialization and orthogonal procrustes method to align the node embeddings into common space which makes node embedding able to inheritance information. In particular, in the common space, we train a model to capture the dynamics information of the networks and smooth temporal node embeddings. We evaluate our method comparing it with other methods on three real-world datasets. The experimental results prove the effectiveness of dynamic network embeddings generated by DynSEM model.

Thinking about cross-domain image-to-image translation problems, where an input image belonging to domain U is transformed into an output image belonging to another domain V. A series of typical tasks, such as style transformation, colorization, super-resolution, can be seen as cross-domain image-to-image translation tasks. Recent methods such as Conditional Generative Adversarial Networks (cGANs) make big progress in this field, but they require paired image data, which is hard to obtain. The DualGAN (Unsupervised Dual Learning for Image-to-Image Translation) architecture was proposed to solve the issue of lack of paired data. But the pixel-level reconstruction losses of DualGAN are simple. In this paper, we replace the pixel-level reconstruction losses with the perceptual reconstruction losses, and propose a more advanced framework for cross-domain image-to-image translation named perceptual-DualGAN. The perceptual reconstruction losses consist of feature reconstruction losses and style reconstruction losses, both of them are computed from pretrained loss networks. Experiments on multiple image translation tasks show that our framework almost performs superior to other methods. And the results of experiments illustrate that our framework can generate more realistic and more natural photos.