Recent studies have raised significant concerns regarding the reliability of current mathematical benchmarks, highlighting key limitations such as simplistic design and potential data contamination that undermine evaluation accuracy. Consequently, developing a reliable benchmark that effectively evaluates large language models' (LLMs) genuine capabilities in mathematical reasoning remains a critical challenge. To address these concerns, we propose RV-Bench, a novel evaluation methodology for Benchmarking LLMs with Random Variables in mathematical reasoning. Specifically, we develop question-generating functions to produce random variable questions (RVQs), whose background content mirrors the original benchmark problems, but with randomized variable combinations, rendering them "unseen" to LLMs. Models must completely understand the inherent question pattern to correctly answer RVQs with diverse variable combinations. Thus, an LLMs' genuine reasoning capability is reflected through its accuracy and robustness on RV-Bench. We conducted extensive experiments on over 30 representative LLMs across more than 1,000 RVQs. Our findings reveal that LLMs exhibit a proficiency imbalance between encountered and "unseen" data distributions. Furthermore, RV-Bench reveals that proficiency generalization across similar mathematical reasoning tasks is limited, but we verified that it can still be effectively elicited through test-time scaling.

Chuang Zhou, Junnan Dong, Yilin Xiao, Shengyuan Chen, Su Dong, di Yin, Xing Sun, Zhaozhuo Xu, Xiao Huang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Chuang Zhou, Zheng Yuan, Linhao Luo, Zhaozhuo Xu, Yilin Xiao, Junnan Dong, Siyu An, di Yin, Xing Sun, Xiao Huang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Recent advancements in large language models (LLMs) have shown promise in bridging the gap between natural language queries and database management systems, enabling users to interact with databases without the background of SQL. However, LLMs often struggle to fully exploit and comprehend the user intention and complex structures of databases. Decomposition-based methods have been proposed to enhance the performance of LLMs on complex tasks, but decomposing SQL generation into subtasks is non-trivial due to the declarative structure of SQL syntax and the intricate connections between query concepts and database elements. In this paper, we propose a novel Structure GUided text-to-SQL framework ( SGU-SQL) that incorporates syntax-based prompting to enhance the SQL generation capabilities of LLMs. Specifically, SGU-SQL establishes structure-aware links between user queries and database schema and recursively decomposes the complex generation task using syntax-based prompting to guide LLMs in incrementally constructing target SQLs. Extensive experiments on two benchmark datasets demonstrate that SGU-SQL consistently outperforms state-of-the-art text-to-SQL baselines.

Knowledge-based visual question answering (KVQA) has been extensively studied to answer visual questions with external knowledge, e.g., knowledge graphs (KGs).While several attempts have been proposed to leverage large language models (LLMs) as an implicit knowledge source, it remains challenging since LLMs may generate hallucinations.Moreover, multiple knowledge sources, e.g., images, KGs and LLMs, cannot be readily aligned for complex scenarios.To tackle these, we present a novel modality-aware integration with LLMs for KVQA (MAIL).It carefully leverages multimodal knowledge for both image understanding and knowledge reasoning.Specifically, (i) we propose a two-stage prompting strategy with LLMs to densely embody the image into a scene graph with detailed visual features; (ii) We construct a coupled concept graph by linking the mentioned entities with external facts.(iii) A tailored pseudo-siamese graph medium fusion is designed for sufficient multimodal fusion.We utilize the shared mentioned entities in two graphs as mediums to bridge a tight intermodal exchange, while maximally preserving insightful intra-modal learning by constraining the fusion within mediums.Extensive experiments show the superiority of MAIL.

Inductive knowledge graph completion (KGC) aims to infer the missing relation for a set of newly-coming entities that never appeared in the training set. Such a setting is more in line with reality, as real-world KGs are constantly evolving and introducing new knowledge. Recent studies have shown promising results using message passing over subgraphs to embed newly-coming entities for inductive KGC. However, the inductive capability of these methods is usually limited by two key issues. (i) KGC always suffers from data sparsity, and the situation is even exacerbated in inductive KGC where new entities often have few or no connections to the original KG. (ii) Cold-start problem. It is over coarse-grained for accurate KG reasoning to generate representations for new entities by gathering the local information from few neighbors. To this end, we propose a novel iNfOmax RelAtion Network, namely NORAN, for inductive KG completion. It aims to mine latent relation patterns for inductive KG completion. Specifically, by centering on relations, NORAN provides a hyper view towards KG modeling, where the correlations between relations can be naturally captured as entity-independent logical evidence to conduct inductive KGC. Extensive experiment results on five benchmarks show that our framework substantially outperforms the state-of-the-art KGC methods.

Entity alignment (EA) aims to merge two knowledge graphs (KGs) by identifying equivalent entity pairs. While existing methods heavily rely on human-generated labels, it is prohibitively expensive to incorporate cross-domain experts for annotation in real-world scenarios. The advent of Large Language Models (LLMs) presents new avenues for automating EA with annotations, inspired by their comprehensive capability to process semantic information. However, it is nontrivial to directly apply LLMs for EA since the annotation space in real-world KGs is large. LLMs could also generate noisy labels that may mislead the alignment. To this end, we propose a unified framework, LLM4EA, to effectively leverage LLMs for EA. Specifically, we design a novel active learning policy to significantly reduce the annotation space by prioritizing the most valuable entities based on the entire inter-KG and intra-KG structure. Moreover, we introduce an unsupervised label refiner to continuously enhance label accuracy through in-depth probabilistic reasoning. We iteratively optimize the policy based on the feedback from a base EA model. Extensive experiments demonstrate the advantages of LLM4EA on four benchmark datasets in terms of effectiveness, robustness, and efficiency.

Knowledge-based question answering (KBQA) is widely used in many scenarios that necessitate domain knowledge. Large language models (LLMs) bring opportunities to KBQA, while their costs are significantly higher and absence of domain-specific knowledge during pre-training. We are motivated to combine LLMs and prior small models on knowledge graphs (KGMs) for both inferential accuracy and cost saving. However, it remains challenging since accuracy and cost are not readily combined in the optimization as two distinct metrics. It is also laborious for model selection since different models excel in diverse knowledge. To this end, we propose Coke, a novel cost-efficient strategy for KBQA with LLMs, modeled as a tailored multi-armed bandit problem to minimize calls to LLMs within limited budgets. We first formulate the accuracy expectation with a cluster-level Thompson Sampling for either KGMs or LLMs. A context-aware policy is optimized to further distinguish the expert model subject to the question semantics. The overall decision is bounded by the cost regret according to historical expenditure on failures. Extensive experiments showcase the superior performance of Coke, which moves the Pareto frontier with up to 20.89% saving of GPT-4 fees while achieving a 2.74% higher accuracy on the benchmark datasets.

Large Language Models (LLMs) have demonstrated remarkable capabilities in many real-world applications. Nonetheless, LLMs are often criticized for their tendency to produce hallucinations, wherein the models fabricate incorrect statements on tasks beyond their knowledge and perception. To alleviate this issue, graph retrieval-augmented generation (GraphRAG) has been extensively explored which leverages the factual knowledge in knowledge graphs (KGs) to ground the LLM's responses in established facts and principles. However, most state-of-the-art LLMs are closed-source, making it challenging to develop a prompting framework that can efficiently and effectively integrate KGs into LLMs with hard prompts only. Generally, existing KG-enhanced LLMs usually suffer from three critical issues, including huge search space, high API costs, and laborious prompt engineering, that impede their widespread application in practice. To this end, we introduce a novel **Know**ledge **Gr**aph based **P**romp**T**ing framework, namely **KnowGPT**, to enhance LLMs with domain knowledge. KnowGPT contains a knowledge extraction module to extract the most informative knowledge from KGs, and a context-aware prompt construction module to automatically convert extracted knowledge into effective prompts. Experiments on three benchmarks demonstrate that KnowGPT significantly outperforms all competitors. Notably, KnowGPT achieves a 92.6% accuracy on OpenbookQA leaderboard, comparable to human-level performance.

Knowledge graphs (KGs) can structurally organize large-scale information in the form of triples and significantly support many real-world applications. While most KG embedding algorithms hold the assumption that all triples are correct, considerable errors were inevitably injected during the construction process. It is urgent to develop effective error-aware KG embedding, since errors in KGs would lead to significant performance degradation in downstream applications. To this end, we propose a novel framework named Attributed Error-aware Knowledge Embedding (AEKE). It leverages the semantics contained in entity attributes to guide the KG embedding model learning against the impact of erroneous triples. We design two triple-level hypergraphs to model the topological structures of the KG and its attributes, respectively. The confidence score of each triple is jointly calculated based on self-contradictory within the triple, consistency between local and global structures, and homogeneity between structures and attributes. We leverage confidence scores to adaptively update the weighted aggregation in the multi-view graph learning framework and margin loss in KG embedding, such that potential errors will contribute little to KG learning. Experiments on three real-world KGs demonstrate that AEKE outperforms state-of-the-art KG embedding and error detection algorithms.

The graph neural network-based collaborative filtering (CF) models user-item interactions as a bipartite graph and performs iterative aggregation to enhance performance. Unfortunately, the aggregation process may amplify the popularity bias, which impedes user engagement with niche (unpopular) items. While some efforts have studied the popularity bias in CF, they often focus on modifying loss functions, which can not fully address the popularity bias in GNN-based CF models. This is because the debiasing loss can be falsely backpropagated to non-target nodes during the backward pass of the aggregation.

Knowledge graphs (KGs) could effectively integrate a large number of real-world assertions, and improve the performance of various applications, such as recommendation and search. KG error detection has been intensively studied since real-world KGs inevitably contain erroneous triples. While existing studies focus on developing a novel algorithm dedicated to one or a few data characteristics, we explore advancing KG error detection by assembling a set of state-of-the-art (SOTA) KG error detectors. However, it is nontrivial to develop a practical ensemble learning framework for KG error detection. Existing ensemble learning models heavily rely on labels, while it is expensive to acquire labeled errors in KGs. Also, KG error detection itself is challenging since triples contain rich semantic information and might be false because of various reasons. To this end, we propose to leverage active learning to minimize human efforts. Our proposed framework - KAEL, could effectively assemble a set of off-the-shelf error detection algorithms, by actively using a limited number of manual annotations. It adaptively updates the ensemble learning policy in each iteration based on active queries, i.e., the answers from experts. After all annotation budget is used, KAEL utilizes the trained policy to identify remaining suspicious triples. Experiments on real-world KGs demonstrate that we can achieve significant improvement when applying KAEL to assemble SOTA error detectors. KAEL also outperforms SOTA ensemble learning baselines significantly.

Knowledge graphs (KGs) have been widely used to enhance complex question answering (QA). To understand complex questions, existing studies employ language models (LMs) to encode contexts. Despite the simplicity, they neglect the latent relational information among question concepts and answers in KGs. While question concepts ubiquitously present hyponymy at the semantic level, e.g., mammals and animals, this feature is identically reflected in the hierarchical relations in KGs, e.g., a_type_of. Therefore, we are motivated to explore comprehensive reasoning by the hierarchical structures in KGs to help understand questions. However, it is non-trivial to reason over tree-like structures compared with chained paths. Moreover, identifying appropriate hierarchies relies on expertise. To this end, we propose HamQA, a novel Hierarchy-aware multi-hop Question Answering framework on knowledge graphs, to effectively align the mutual hierarchical information between question contexts and KGs. The entire learning is conducted in Hyperbolic space, inspired by its advantages of embedding hierarchical structures. Specifically, (i) we design a context-aware graph attentive network to capture context information. (ii) Hierarchical structures are continuously preserved in KGs by minimizing the Hyperbolic geodesic distances. The comprehensive reasoning is conducted to jointly train both components and provide a top-ranked candidate as an optimal answer. We achieve a higher ranking than the state-of-the-art multi-hop baselines on the official OpenBookQA leaderboard with an accuracy of 85%.

Knowledge Graph (KG) errors introduce non-negligible noise, severely affecting KG-related downstream tasks. Detecting errors in KGs is challenging since the patterns of errors are unknown and diverse, while ground-truth labels are rare or even unavailable. A traditional solution is to construct logical rules to verify triples, but it is not generalizable since different KGs have distinct rules with domain knowledge involved. Recent studies focus on designing tailored detectors or ranking triples based on KG embedding loss. However, they all rely on negative samples for training, which are generated by randomly replacing the head or tail entity of existing triples. Such a negative sampling strategy is not enough for prototyping practical KG errors, e.g., (Bruce_Lee, place_of_birth, China), in which the three elements are often relevant, although mismatched. We desire a more effective unsupervised learning mechanism tailored for KG error detection. To this end, we propose a novel framework - ContrAstive knowledge Graph Error Detection (CAGED). It introduces contrastive learning into KG learning and provides a novel way of modeling KG. Instead of following the traditional setting, i.e., considering entities as nodes and relations as semantic edges, CAGED augments a KG into different hyper-views, by regarding each relational triple as a node. After joint training with KG embedding and contrastive learning loss, CAGED assesses the trustworthiness of each triple based on two learning signals, i.e., the consistency of triple representations across multi-views and the self-consistency within the triple. Extensive experiments on three real-world KGs show that CAGED outperforms state-of-the-art methods in KG error detection. Our codes and datasets are available at https://github.com/Qing145/CAGED.git.