Xueying Ding, Xingyue Huang, Mingxuan Ju, Liam Collins, Yozen Liu, Leman Akoglu, Neil Shah, Tong Zhao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Softmax attention struggles with long contexts due to structural limitations: the strict sum-toone constraint forces attention sinks on irrelevant tokens, and probability mass disperses as sequence lengths increase.We tackle these problems with Threshold Differential Attention (TDA), a sink-free attention mechanism that achieves ultra-sparsity and improved robustness at longer sequence lengths without the computational overhead of projection methods or the performance degradation caused by noise accumulation of standard rectified attention.TDA applies row-wise extreme-value thresholding with a length-dependent gate, retaining only exceedances.Inspired by the differential transformer, TDA also subtracts an inhibitory view to enhance expressivity.Theoretically, we prove that TDA controls the expected number of spurious survivors per row to O(1) and that consensus spurious matches across independent views vanish as context grows.Empirically, TDA produces > 99% exact zeros and eliminates attention sinks while maintaining competitive performance on standard and long-context benchmarks.

Knowledge Graph Foundation Models (KGFMs) are at the frontier for deep learning on knowledge graphs (KGs), as they can generalize to completely novel knowledge graphs with different relational vocabularies. Despite their empirical success, our theoretical understanding of KGFMs remains very limited. In this paper, we conduct a rigorous study of the expressive power of KGFMs. Specifically, we show that the expressive power of KGFMs directly depends on the *motifs* that are used to learn the relation representations. We then observe that the most typical motifs used in the existing literature are *binary*, as the representations are learned based on how pairs of relations interact, which limits the model's expressiveness. As part of our study, we design more expressive KGFMs using richer motifs, which necessitate learning relation representations based on, e.g., how triples of relations interact with each other. Finally, we empirically validate our theoretical findings, showing that the use of richer motifs results in better performance on a wide range of datasets drawn from different domains.

Graph neural networks are popular architectures for graph machine learning, based on iterative computation of node representations of an input graph through a series of invariant transformations. A large class of graph neural networks follow a standard message-passing paradigm: at every layer, each node state is updated based on an aggregate of messages from its neighborhood. In this work, we propose a novel framework for training graph neural networks, where every node is viewed as a player that can choose to either listen, broadcast, listen and broadcast, or to isolate. The standard message propagation scheme can then be viewed as a special case of this framework where every node listens and broadcasts to all neighbors. Our approach offers a more flexible and dynamic message-passing paradigm, where each node can determine its own strategy based on their state, effectively exploring the graph topology while learning. We provide a theoretical analysis of the new message-passing scheme which is further supported by an extensive empirical analysis on a synthetic and real-world datasets.

The development of data sensing technology has generated a vast amount of high-dimensional data, posing great challenges for machine learning models. Over the past decades, despite demonstrating its effectiveness in data classification, genetic programming (GP) has still encountered three major challenges when dealing with high-dimensional data: 1) solution diversity; 2) multiclass imbalance; and 3) large feature space. In this article, we have developed a problem-specific multiobjective GP framework (PS-MOGP) for handling classification tasks with high-dimensional data. To reduce the large solution space caused by high dimensionality, we incorporate the recursive feature elimination strategy based on mining the archive of evolved GP solutions. A progressive domination Pareto archive evolution strategy (PD-PAES), which optimizes the objectives in a specific order according to their objectives, is proposed to evaluate the GP individuals and maintain a better diversity of solutions. Besides, to address the seriously imbalanced class issue caused by traditional binary decomposition (BD) one versus rest (OVR) for multiclass classification problems, we design a method named BD with a similar positive and negative class size (BD-SPNCS) to generate a set of auxiliary classifiers. Experimental results on benchmark and real-world datasets demonstrate that our proposed PS-MOGP outperforms state-of-the-art traditional and evolutionary classification methods in the context of high-dimensional data classification.

Graph neural networks are prominent models for representation learning over graph-structured data. While the capabilities and limitations of these models are well-understood for simple graphs, our understanding remains incomplete in the context of knowledge graphs. Our goal is to provide a systematic understanding of the landscape of graph neural networks for knowledge graphs pertaining to the prominent task of link prediction. Our analysis entails a unifying perspective on seemingly unrelated models and unlocks a series of other models. The expressive power of various models is characterized via a corresponding relational Weisfeiler-Leman algorithm. This analysis is extended to provide a precise logical characterization of the class of functions captured by a class of graph neural networks. The theoretical findings presented in this paper explain the benefits of some widely employed practical design choices, which are validated empirically.

For high-dimensional data classification, feature selection has played a significant role in removing the redundant or irrelevant features and improving the performances of classifiers. Particle Swarm Optimization (PSO), an evolutionary computational tool, has been widely and successfully applied in the field of feature selection. A recent pioneer work, Potential PSO (PPSO) which utilizes feature discretization process based on cut-points, provides better overall performance in classification and lower amount of selected features compared with some existing approaches. In this paper, we proposed an adaptive PPSO (APPSO) to reduce the number of features further and increase the classification accuracy. In APPSO, the pre-filtering method, ReliefF, is incorporated into the discretization process to help screen the potential cut-points effectively, and a new updating strategy is proposed to strengthen the searching scope of particles and avoid being trapped into local optima. Also, APPSO follows an adaptive way to alter the random level of the selection of cutpoints. The random level depends on the mean instance number per class of each dataset. Experiments on 10 datasets are carried out, and APPSO is proven to select less than 2% of the number of features while maintaining a very competitive accuracy compared with some state-of-the-art methods.