Large language models (LLMs) show promising performance on small-scale graph reasoning tasks but fail when handling real-world graphs with complex queries. This phenomenon arises from LLMs' working memory constraints, which result in their inability to retain long-range graph topology over extended contexts while sustaining coherent multi-step reasoning. However, real-world graphs are often structurally complex, such as Web, Transportation, Social, and Citation networks. To address these limitations, we propose GraphCogent, a collaborative agent framework inspired by human Working Memory Model that decomposes graph reasoning into specialized cognitive processes: sense, buffer, and execute. The framework consists of three modules: Sensory Module standardizes diverse graph text representations via subgraph sampling, Buffer Module integrates and indexes graph data across multiple formats, and Execution Module combines tool calling and tool creation for efficient reasoning. We also introduce Graph4real, a comprehensive benchmark that contains four domains of real-world graphs (Web, Transportation, Social, and Citation) to evaluate LLMs' graph reasoning capabilities. Our Graph4real covers 21 different graph reasoning tasks, categorized into three types (Structural Querying, Algorithmic Reasoning, and Predictive Modeling tasks), with graph scales up to 10 times larger than existing benchmarks. Experiments show that Llama3.1-8B based GraphCogent achieves a 50% improvement over massive-scale LLMs like DeepSeek-R1 (671B). Compared to state-of-the-art code-based baseline, our framework outperforms by 20% in accuracy while reducing token usage by 80% for in-toolset tasks and 30% for out-toolset tasks.

Session-based recommendation (SBR) predicts the next interaction of anonymous users based on their session data. In practical scenarios, low-exposure tail items constitute the majority of interactions, leading to the biased recommendation whereby a small set of high-exposure head items is preferentially recommended. Existing solutions focus on promoting those tail items but with a sacrifice in recommendation accuracy. We attribute such a tradeoff to session-irrelevant noise within the tail items, which existing long-tail approaches fail to identify and constrain effectively. To resolve this fundamental conflict, we propose Hybrid Intent-based Dual Constraint framework (HID) , a plug-and-play framework that converts the traditional tradeoff between long-tail performance and accuracy into a synergistic relationship by introducing hybrid intent-based dual constraints. HID incorporates two key innovations: (i) Hybrid Intent Learning , which reformulates intent extraction through attribute-aware spectral clustering and discriminates session-irrelevant noise by assigning target and noise intents to each session and (ii) Intent Constraint Loss , which introduces two constraints—diversity and accuracy—to regulate item and session representation learning. These two objectives are unified into a single training loss through rigorous theoretical derivation. To handle ambiguous user intent in the real-world, we also propose a slack version of HID. Extensive experiments show that HID enhances both long-tail performance and accuracy, establishing new state-of-the-art performance in long-tail recommender systems. The implementation code is available at: https://github.com/jarviswww/Code4TradeSynergy .

Large language models (LLMs) have been demonstrated to possess the capabilities to understand fundamental graph properties and address various graph reasoning tasks. Existing methods fine-tune LLMs to understand and execute graph reasoning tasks by specially designed task instructions. However, these Text-Instruction methods generally exhibit poor performance. Inspired by tool learning, researchers propose Tool-Instruction methods to solve various graph problems by special tool calling (e.g., function, API and model), achieving significant improvements in graph reasoning tasks. Nevertheless, current Tool-Instruction approaches focus on the tool information and ignore the graph structure information, which leads to significantly inferior performance on small-scale LLMs (less than 8B). To tackle this issue, we propose GraphTool-Instruction, an innovative Instruction-tuning approach that decomposes the graph reasoning task into three distinct subtasks (i.e., graph extraction, tool name identification and tool parameter extraction), and design specialized instructions for each subtask. Our GraphTool-Instruction can be used as a plug-and-play prompt for different LLMs without fine-tuning. Moreover, building on GraphTool-Instruction, we develop GTools, a dataset that includes twenty graph reasoning tasks, and create a graph reasoning LLM called GraphForge based on Llama3-8B. We conduct extensive experiments on twenty graph reasoning tasks with different graph types (e.g., graph size or graph direction), and we find that GraphTool-Instruction achieves SOTA compared to Text-Instruction and Tool-Instruction methods. Fine-tuned on GTools, GraphForge gets further improvement of over 30% compared to the Tool-Instruction enhanced GPT-3.5-turbo, and it performs comparably to the high-cost GPT-4o. Our codes and data are available at https://github.com/RongzhengWang/GraphTool-Instruction.

Recently, prompt tuning has garnered considerable attention due to its success across various Vision-Language (VL) tasks. However, unimodal prompts, coupled prompts, and joint prompts in these models often lead to suboptimal performance due to differences in information density and complexity between modalities. Particularly, in scenarios with missing modalities, these prompt-based approaches tend to exacerbate 'Channel Bias'-a phonomenon where models overly rely on specific feature (such as unmissing-modal feature) channels from the base tasks, thereby undermining the model's ability to capture crucial shared knowledge applicable to new tasks and affecting its generalizability. To address this challenge, we propose RobustPT, a dynamic disentanglement prompt tuning model designed to enhance the robustness of VL models under modality missing conditions. RobustPT utilizes a multi-channel prompting mechanism to dynamically disentangle and align prompts. Specifically, RobustPT is divided into single-channel tuning and alignment-channel tuning, where prompts for each modality run independently in sequence to delve deeply into their intrinsic characteristics, followed by an integration through a non-strong coupling strategy to effectively balance information contributions and enhance overall performance. Extensive experiments demonstrate that our RobustPT achieve significant improvements over the current state-of-the-art across all benchmark datasets. Our codes are available at https://github.com/Trae1ounG/RobustPT.

Abstract Session-based recommendation (SBR) focuses on predicting the next potential item for anonymous users based on short-click sessions. However, these interaction sessions often contain noise items, which arise from misclicks or shifts in user interests. Existing denoising methods typically presume a strong exclusionary relationship between noise items and the recommendation target, assuming that reducing noise can enhance recommendation accuracy. In contrast, our observations reveal a nuanced phenomenon: as the length of the interaction session shortens, the effect of noise removal on recommendation performance gradually transitions from positive to negative. This finding suggests that in short sessions with insufficient contextual information, relying solely on the exclusion of noise items within the session may fail to improve and could even hinder-the recommendation performance. Such complexities have been largely overlooked in prior research. To bridge this gap, we propose two solutions: (i) expanding the view of denoising from a single session to multiple sessions (i.e., from local to global), and (ii) introducing relevant contextual information into each session by employing enhancement strategies. Therefore, we design the Hybrid Prototype-based In-and-Out Flow Network (HyPro), which employs both denoising and enhancing processes for each session based on our proposed hybrid prototypes. Specifically, for each item, HyPro first learns the hybrid prototype by aggregating information from the item’s semantic and topological neighbors across all sessions. Then, based on the hybrid prototypes, HyPro employs an in-and-out flow network comprising two components: (i) the out-flow channel, which targets the removal of irrelevant information at both the data and feature levels, and (ii) the in-flow channel, which integrates global information for each session at the item and session levels. Extensive experiments conducted on three real-world datasets demonstrate that HyPro outperforms the state-of-the-art baselines. The implementation code is available at: https://github.com/jarviswww/Code4HyPro .

Predicting missing facts for temporal knowledge graphs (TKGs) is a fundamental task, called temporal knowledge graph completion (TKGC). One key challenge in this task is the imbalance in data distribution, where facts are unevenly spread across entities and timestamps. This imbalance can lead to poor completion performance for long-tail entities and timestamps, and unstable training due to the introduction of false negative samples. Unfortunately, few previous studies have investigated how to mitigate these effects. Moreover, for the first time, we found that existing methods suffer from model preferences, revealing that entities with specific properties (e.g., recently active) are favored by different models. Such preferences will lead to error accumulation and further exacerbate the effects of imbalanced data distribution. To alleviate the impacts of imbalanced data and model preferences, we introduce Booster , the first data augmentation strategy for TKGs. The unique requirements here lie in generating new samples that fit the complex semantic and temporal patterns within TKGs, and identifying hard-learning samples specific to models. Therefore, we propose a hierarchical scoring algorithm based on triadic closures within TKGs. By incorporating both global semantic patterns and local time-aware structures, the algorithm enables pattern-aware validation for new samples. Meanwhile, we propose a two-stage training approach to identify samples that deviate from the model's preferred patterns. With a frequency-based filtering strategy, this approach also helps to avoid the misleading of false negatives. Experiments justify that Booster can seamlessly adapt to existing TKGC models and achieve on average 4.5% performance improvement.

Commonsense question answering has demonstrated considerable potential across various applications like assistants and social robots.Although fully fine-tuned Pre-trained Language Model(PLM) has achieved remarkable performance in commonsense reasoning, their tendency to excessively prioritize textual information hampers the precise transfer of structural knowledge and undermines interpretability.Some studies have explored combining Language Models (LM) with Knowledge Graphs (KGs) by coarsely fusing the two modalities to perform Graph Neural Network (GNN)-based reasoning that lacks a profound interaction between heterogeneous modalities.In this paper, we propose a novel Graph-based Structure-Aware Prompt Learning Model for commonsense reasoning, named G-SAP, aiming to maintain a balance between heterogeneous knowledge and enhance the cross-modal interaction within the LM+GNNs model.In particular, an evidence graph is constructed by integrating multiple knowledge sources, i.e.ConceptNet, Wikipedia, and Cambridge Dictionary to boost the performance.Afterward, a structure-aware frozen PLM is employed to fully incorporate the structured and textual information from the evidence graph, where the generation of prompts is driven

The representation of semantic information pertaining to the real world has been active research for some time now. Among the available methods, knowledge graphs have emerged as a widely accepted approach. Meanwhile, graph neural networks (GNNs) have demonstrated excellent performance in embedding graph-based information. Given the natural graph structure of knowledge graphs, employing GNNs to embed them is expected to yield a more interpretable and trustworthy representation of the learned knowledge. In this paper, we propose three customized GNNs for different scenarios of knowledge graph representation, including traditional, multimodal, and uncertain knowledge graphs. In the traditional knowledge graph scenario, we present a graph self-supervised learning method, named deep relation graph infomax (DRGI), which incorporates both the complete graph structure information and semantic information. In the multimodal knowledge graph scenario, we introduce a novel network, named hyper-node relational graph attention network (HRGAT), which combines different modal information with graph structure information for a more precise representation of multimodal knowledge graphs. In the uncertain knowledge graph scenario, we define a novel message-passing paradigm with box embedding, named box graph neural network (BGNN). BGNN leverages both the graph structure information of uncertain knowledge graphs and the probabilistic semantics of box embedding. To validate the effectiveness of our proposed methods, we conduct a series of experiments and report the results. We also discuss possible future work in GNN-based knowledge graph embedding.

Knowledge graph (KG) enhanced recommendation, which aims to solve the cold start and explainability in recommender systems, has attracted considerable research interest recently. Existing recommender systems usually focus on implicit feedback such as purchase history without negative feedback. Most of them apply the negative sampling strategy to deal with the implicit feedback data, which may ignore the latent positive user-item interaction. Some other works adopt the non-sampling strategy that treats all non-observed interactions as negative samples and assigns a weight for each negative sample to represent the probability that this sample is a positive sample. However, they use a simple and intuitive weight assignment strategy and cannot catch the latent relationship from all interaction data. To address these problems, we consider graph structure information of both user-item interaction and knowledge graph, and propose a Graph-based Non-Sampling strategy to achieve efficient performance in Knowledge graph enhanced Recommendation (GNSKR). GNSKR utilizes node centrality to significantly improve recommendation performance with low computation cost. Meanwhile, we combine knowledge graph embedding and recommendation task with a local aggregation block, which efficiently catches the high-order connection information in KG enhanced recommendation. Experiments on three datasets show that GNSKR embraces the state-of-the-art with competitive efficiency.

Recently, many knowledge graph embedding models for knowledge graph completion have been proposed, ranging from the initial translation-based models such as TransE to recent convolutional neural network (CNN) models such as ConvE. However, these models only focus on semantic information of knowledge graph and neglect the natural graph structure information. Although graph convolutional network (GCN)-based models for knowledge graph embedding have been introduced to address this issue, they still suffer from fact incompleteness, resulting in the unconnectedness of knowledge graph. To solve this problem, we propose a novel model called deep relational graph infomax (DRGI) with mutual information (MI) maximization which takes the benefit of complete structure information and semantic information together. Specifically, the proposed DRGI consists of two encoders which are two identical adaptive relational graph attention networks (ARGATs), corresponding to catching semantic information and complete structure information respectively. Our method establishes new state-of-the-art on the standard datasets for knowledge graph completion. In addition, by exploring the complete structure information, DRGI embraces the merits of faster convergence speed over existing methods and better predictive performance for entities with small indegree.

Recently, many knowledge graph embedding models for knowledge graph completion have been proposed, ranging from the initial translation-based models such as TransE to recent convolutional neural network (CNN) models such as ConvE. However, these models only focus on semantic information of knowledge graph and neglect the natural graph structure information. Although graph convolutional network (GCN)-based models for knowledge graph embedding have been introduced to address this issue, they still suffer from fact incompleteness, resulting in the unconnectedness of knowledge graph. To solve this problem, we propose a novel model called deep relational graph infomax (DRGI) with mutual information (MI) maximization which takes the benefit of complete structure information and semantic information together. Specifically, the proposed DRGI consists of two encoders which are two identical adaptive relational graph attention networks (ARGATs), corresponding to catching semantic information and complete structure information respectively. Our method establishes new state-of-the-art on the standard datasets for knowledge graph completion.