Knowledge Graph Question Answering (KGQA) aims to improve factual accuracy by leveraging structured knowledge. However, real-world Knowledge Graphs (KGs) are often incomplete, leading to the problem of Incomplete KGQA (IKGQA). A common solution is to incorporate external data to fill knowledge gaps, but existing methods lack the capacity to adaptively and contextually fuse multiple sources, failing to fully exploit their complementary strengths. To this end, we propose Debate over Mixed-knowledge (DoM), a novel framework that enables dynamic integration of structured and unstructured knowledge for IKGQA. Built upon the Multi-Agent Debate paradigm, DoM assigns specialized agents to perform inference over knowledge graphs and external texts separately, and coordinates their outputs through iterative interaction. It decomposes the input question into sub-questions, retrieves evidence via dual agents (KG and Retrieval-Augmented Generation, RAG), and employs a judge agent to evaluate and aggregate intermediate answers. This collaboration exploits knowledge complementarity and enhances robustness to KG incompleteness. In addition, existing IKGQA datasets simulate incompleteness by randomly removing triples, failing to capture the irregular and unpredictable nature of real-world knowledge incompleteness. To address this, we introduce a new dataset, Incomplete Knowledge Graph WebQuestions, constructed by leveraging real-world knowledge updates. These updates reflect knowledge beyond the static scope of KGs, yielding a more realistic and challenging benchmark. Through extensive experiments, we show that DoM consistently outperforms state-of-the-art baselines.

Naixin Zhai, Pengyang Shao, Binbin Zheng, Yonghui Yang, Fei Shen, Long Bai, Xun Yang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Cognitive Diagnosis (CD) models are constrained by the data quality of students' response logs. Recent advancements in Large Language Model (LLM) based data augmentation show promise for enhancing CD. However, ensuring the reliability and accuracy of LLM-generated annotations remains a significant challenge. In this paper, we propose Multi-Agent based Concept Augmentation for Cognitive Diagnosis (MACA-CD), a novel approach that enhances CD by generating and fusing reliable concept descriptions and relations based solely on concept names. MACA-CD consists of two main components: (1) a Multi-Agent Debate (MAD) based concept augmentation process that generates diverse and reliable concept descriptions and relations, reducing reliance on behavioral data. For concept descriptions, two agents generate outputs that include definitions, core features, and real-world applications, and continue debating until a judge agent determines that consensus has been reached. Concept relations are then identified using a Breadth-First Search approach to efficiently and progressively uncover relationships based on concept descriptions, with each step carried out by MAD. (2) a concept augmentation-enhanced CD model that refines concept embeddings using a graph self-supervised learning fusion layer and a pairwise comparator-based Description Fusion Layer, leading to more reliable and accurate concept embeddings. Experimental results on three real-world datasets show that MACA-CD consistently outperforms existing methods under various real-world scenarios.

Visual Document Retrieval (VDR) requires representations that capture both fine-grained visual details and global document structure to ensure retrieval efficacy while maintaining computational efficiency. Existing VDR models struggle to balance effectiveness and efficiency when processing high-resolution documents: they often either lose fine-grained information or generate an excessive number of visual tokens, resulting in significant indexing overhead and high retrieval latency. In this work, we rethink the visual encoding mechanism and propose a new X-VisEmb paradigm that progresses from multi-resolution sampling and encoding, through cross-granularity feature fusion, to adaptive representation distillation. A preliminary study validates its feasibility and effectiveness in capturing complementary visual cues at varying scales. Building on the insights, we develop MURE , a novel framework that employs VLMs as a hierarchical multi-resolution encoder, integrates resolution-level Matryoshka representation learning (RMRL) for effective feature fusion, and applies a semantic-aware hierarchical clustering mechanism for visual token compression. Experiments on two widely used VDR benchmarks show that our MURE framework consistently beats strong baselines. Furthermore, it significantly outperforms ColPali with only 50% of its visual token budget.

As Large Language Models (LLMs) increasingly shape online content, how to remove targeted information from well-trained LLMs (also known as LLM unlearning) has become increasingly critical for web governance. A key challenge in LLM unlearning lies in the sample-wise imbalance within the forget set: different samples exhibit widely varying unlearning difficulty, leading to asynchronous forgetting speeds where some knowledge remains insufficiently erased while others become over-forgotten. To address this challenge, we propose BalDRO, a novel and efficient framework for balanced LLM unlearning. BalDRO formulates unlearning as a min–sup process, where the inner process identifies a worst-case data distribution that adaptively emphasizes hard-to-unlearn samples, while the outer process updates model parameters based on the worst-case data distribution. We instantiate this formulation through two efficient variants: BalDRO-G, a discrete GroupDRO-based approximation that focuses on high-loss subsets, and BalDRO-DV, a continuous Donsker–Varadhan dual method that enables smooth, adaptive weighting within standard LLM training pipelines. Extensive experiments on the TOFU and MUSE benchmarks demonstrate the effectiveness of our proposed BalDRO, yielding significant improvements in both forgetting quality and model utility over existing methods. For reproducibility, we have released the code for BalDRO. https://github.com/nxZhai/BalDRO.

Recommender Systems (RS) have been shown to be vulnerable to injective attacks, where attackers inject limited fake user profiles to promote the exposure of target items to real users for unethical gains (e.g., economic or political advantages). Since attackers typically lack knowledge of the victim model deployed in the target RS, existing methods resort to using a fixed surrogate model to mimic the potential victim model. Despite considerable progress, we argue that the assumption thatpoisoned data generated for the surrogate model can be used to attack other victim modelsis wishful. When there are significant structural discrepancies between the surrogate and victim models, the attack transferability inevitably suffers. Intuitively, if we can identify the worst-case victim model and iteratively optimize the poisoning effect specifically against it, then the generated poisoned data would be better transferred to other victim models. However, exactly identifying the worst-case victim model during the attack process is challenging due to the large space of victim models. To this end, in this work, we propose a novel attack method called Sharpness-Aware Poisoning (SharpAP). Specifically, it employs the sharpness-aware minimization principle to seek the approximately worst-case victim model and optimizes the poisoned data specifically for this worst-case model. The poisoning attack withSharpAPis formulated as a min-max-min tri-level optimization problem. By integratingSharpAPinto the iterative process for attacks, our method can generate more robust poisoned data which is less sensitive to the shift of model structure, mitigating the overfitting to the surrogate model. Comprehensive experimental comparisons on three real-world datasets demonstrate thatSharpAPcan significantly enhance the attack transferability.

Recommender Systems (RS) have been widely adopted to provide personalized suggestions based on historical user behaviors. However, some users are hesitant to allow RS to learn their preferences at the expense of their privacy information. Therefore, these users prefer to hide their preferences from RS. In this article, we consider this practical yet important question: can privacy-concerned users make RS unavailable to learn their preferences? The challenge lies in achieving this goal while complying with real-world constraints. Normal users still expect accurate recommendations; the scope should target privacy-concerned users. Also, as most companies do not allow users to delete their implicit feedback, the solution cannot rely on data deletion. To this end, we propose a novel ULRec from the perspective of fake interaction generation, a general method for making preferences of privacy-concerned users U n L earnable to personalized Rec ommendation algorithms. First, we formulate the constraints of the data attack based on practical considerations. Then, we define a bi-level optimization process, where the outer loop updates data addition, and the inner loop dynamically updates RS parameters. After that, we propose a loss function that simultaneously considers the requests of both privacy-concerned users and normal users. To ensure the feasible range and model efficiency, we adopt projected gradient descent and automatic differentiation. Finally, extensive experiments on three real-world datasets have demonstrated the effectiveness of our proposed ULRec .

Graph-based Cognitive Diagnosis (CD) has attracted much research interest due to its strong ability on inferring students' proficiency levels on knowledge concepts. While graph-based CD models have demonstrated remarkable performance, we contend that they still cannot achieve optimal performance due to the neglect of edge heterogeneity and uncertainty. Edges involve both correct and incorrect response logs, indicating heterogeneity. Meanwhile, a response log can have uncertain semantic meanings, e.g., a correct log can indicate true mastery or fortunate guessing, and a wrong log can indicate a lack of understanding or a careless mistake. In this paper, we propose an Informative Semantic-aware Graph-based Cognitive Diagnosis model (ISG-CD), which focuses on how to utilize the heterogeneous graph in CD and minimize effects of uncertain edges. Specifically, to explore heterogeneity, we propose a semantic-aware graph neural networks based CD model. To minimize effects of edge uncertainty, we propose an Informative Edge Differentiation layer from an information bottleneck perspective, which suggests keeping a minimal yet sufficient reliable graph for CD in an unsupervised way. We formulate this process as maximizing mutual information between the reliable graph and response logs, while minimizing mutual information between the reliable graph and the original graph. After that, we prove that mutual information maximization can be theoretically converted to the classic binary cross entropy loss function, while minimizing mutual information can be realized by the Hilbert-Schmidt Independence Criterion.Finally, we adopt an alternating training strategy for optimizing learnable parameters of both the semantic-aware graph neural networks based CD model and the edge differentiation layer. Extensive experiments on three real-world datasets have demonstrated the effectiveness of ISG-CD.

Graph-based social recommender systems have demonstrated great potential in alleviating data sparsity by leveraging high-order user influence embedded in social networks.However, most existing methods rely heavily on the observed social graph, which is often noisy and includes spurious or task-irrelevant connections that can mislead user preference learning.Identifying and removing these noisy relations is crucial but challenging due to the lack of ground-truth annotations.In this paper, we approach the social denoising problem from the perspective of graph invariant learning and propose a novel approach, Social Graph Invariant Learning(SGIL).Specifically, SGIL aims to uncover stable user preferences within the input social graph, thereby enhancing the robustness of

Recently, the user-side fairness issue in Collaborative Filtering (CF) algorithms has gained considerable attention, arguing that results should not discriminate an individual or a sub-user group based on users’ sensitive attributes (e.g., gender). Researchers have proposed fairness-aware CF models by decreasing statistical associations between predictions and sensitive attributes. A more natural idea is to achieve model fairness from a causal perspective. The remaining challenge is that we have no access to interventions, i.e., the counterfactual world that produces recommendations when each user has changed the sensitive attribute value. To this end, we first borrow the Rubin-Neyman potential outcome framework to define average causal effects of sensitive attributes. Next, we show that removing causal effects of sensitive attributes is equal to average counterfactual fairness in CF. Then, we use the propensity re-weighting paradigm to estimate the average causal effects of sensitive attributes and formulate the estimated causal effects as an additional regularization term. To the best of our knowledge, we are one of the first few attempts to achieve counterfactual fairness from the causal effect estimation perspective in CF, which frees us from building sophisticated causal graphs. Finally, experiments on three real-world datasets show the superiority of our proposed model.

Recommender systems have been widely used in recent years. By exploiting historical user-item interactions, recommender systems can model personalized potential interests of users and have been widely applied to a wide range of scenarios. Despite their impressive performance, most of them may be subject to unwanted biases related to sensitive attributes (e.g., race and gender), leading to unfairness. An intuitive idea to alleviate this problem is to ensure that there is no mutual information between recommendation results and sensitive attributes. However, keeping independence conditions solely achieves fairness improvement while causing an obvious degradation of recommendation accuracy, which is not a desired result. To this end, in this paper, we re-define recommendation fairness with a novel two-fold mutual information objective. In concerned details, we define fairness as mutual information minimization between embeddings and sensitive information, and mutual information maximization between embeddings and non-sensitive information. Then, a flexible Fair Mutual Information (FairMI) framework is designed to achieve this goal. FairMI first employs a sensitive attribute encoder to capture sensitive information in the data. Then, based on results from the sensitive attribute encoder, an interest encoder is developed to generate sensitive-free embeddings, which are expected to contain rich non-sensitive information of input data. Moreover, we propose novel mutual information (upper/lower) bounds with contrastive information estimation for model optimization. Extensive experiments over two real-world datasets demonstrate the effectiveness of our proposed FairMI in reducing unfairness and improving recommendation accuracy simultaneously.

As a key application of artificial intelligence, recommender systems are among the most pervasive computer aided systems to help users find potential items of interests. Recently, researchers paid considerable attention to fairness issues for artificial intelligence applications. Most of these approaches assumed independence of instances, and designed sophisticated models to eliminate the sensitive information to facilitate fairness. However, recommender systems differ greatly from these approaches as users and items naturally form a user-item bipartite graph, and are collaboratively correlated in the graph structure. In this paper, we propose a novel graph based technique for ensuring fairness of any recommendation models. Here, the fairness requirements refer to not exposing sensitive feature set in the user modeling process. Specifically, given the original embeddings from any recommendation models, we learn a composition of filters that transform each user’s and each item’s original embeddings into a filtered embedding space based on the sensitive feature set. For each user, this transformation is achieved under the adversarial learning of a user-centric graph, in order to obfuscate each sensitive feature between both the filtered user embedding and the sub graph structures of this user. Finally, extensive experimental results clearly show the effectiveness of our proposed model for fair recommendation. We publish the source code at https://github.com/newlei/FairGo.