Multimodal Knowledge Graph Completion (MMKGC) is critical for reasoning over sparse relational data enriched with visual and textual features. However, existing methods struggle to generalize in real-world scenarios due to two pervasive challenges: (1) inter-entity heterogeneity, where modality availability is highly inconsistent (e.g., missing images or textual descriptions); and (2) intra-entity imbalance, where the semantic informativeness of modalities varies substantially, often causing models to rely on spurious statistical correlations. To address these limitations, we propose CoMu, a novel counterfactual-based multimodal fusion framework. CoMu first alleviates modality sparsity through a lightweight prompt-guided alignment strategy that recovers missing modalities at the raw feature level. Crucially, rather than relying on static fusion weights, CoMu introduces a counterfactual intervention mechanism that selectively removes modalities and measures the resulting causal deviation in prediction confidence, enabling adaptive feature integration based on their true causal contributions. Furthermore, we introduce a consistency-aware causal regularization to maintain semantic faithfulness during the intervention process. Extensive experiments on three real-world benchmark datasets demonstrate that CoMu effectively mitigates modality imbalance and spurious statistical correlations, consistently outperforming 19 state-of-the-art MMKGC baselines.
Recent work in continual learning has highlighted the stability gap – a temporary performance drop on previously learned tasks when new ones are introduced. This phenomenon reflects a mismatch between rapid adaptation and strong retention at task boundaries, underscoring the need for optimization mechanisms that balance plasticity and stability over abrupt distribution changes. While optimizers such as momentum-SGD and Adam introduce implicit multi-timescale behavior, they still exhibit pronounced stability gaps. Importantly, these gaps persist even under ideal joint training, making it crucial to study them in this setting to isolate their causes from other sources of forgetting. Motivated by how noradrenergic (neuromodulatory) bursts transiently increase neuronal gain under uncertainty, we introduce a dynamic gain scaling mechanism as a two-timescale optimization technique that balances adaptation and retention by transiently increasing the effective update magnitude while dynamically reparameterizing the weights governing the forward pass, thereby empirically mitigating transition-induced curvature amplification. Across domain- and class-incremental MNIST, CIFAR, and mini-ImageNet benchmarks under task-agnostic joint training, dynamic gain scaling effectively attenuates stability gaps while maintaining competitive accuracy, improving robustness at task transitions.
Molecular docking is a key technique in drug discovery, but existing deep learning–based methods still struggle to generate physically valid binding poses consistently. Post-processing with empirical force fields can partially improve physical plausibility, yet it breaks end-to-end modeling, increases computational cost, and often weakens ligand–protein interactions and leads to pose deviations. Diffusion-based docking methods have recently shown improved accuracy and physical realism, but they typically require many sampling steps, leading to low inference efficiency. Approaches based on internal coordinates reduce sampling steps but suffer from complex pipelines and high per-step computational cost. In addition, many docking models rely heavily on extensive sampling and external scoring functions to identify reliable poses, further increasing inference time. To address these challenges, we propose an equivariant flow matching approach for molecular docking guided by ligand–protein interactions in Cartesian space (IFMDock). IFMDock guides the flow matching model to generate physically plausible docking poses with favorable repeated-sampling behavior, while substantially reducing the number of inference steps and improving docking accuracy and physical validity. Experiments on the PoseBusters and DeepDockingDare benchmarks demonstrate its effectiveness and superior performance.
LiDAR semantic segmentation is a fundamental task in autonomous driving and robot perception. Although the existing voxel-based methods using cylindrical partition can adapt to the distribution characteristics of point clouds, they are limited by the modeling within a single cylinder and have difficulty effectively capturing the horizontal context information across cylinders, resulting in insufficient discriminative ability for sparse and minority classes. To address this limitation, we propose SDCBNet, a sparse-to-dense Cylinder-BEV network that transforms sparse cylindrical features into dense BEV representations. This method introduces a BEV branch on the basis of Cylinder3D and establishes an explicit injection mechanism between cylindrical features and the BEV space in multiple stages of encoding-decoding to enhance the modeling ability of horizontal context. Further, this paper designs the sparse-to-dense feature aggregation module SDFA, which generates a robust BEV representation through a dual-branch strategy of maximum pooling and attention aggregation; at the same time, it proposes the conditional routing feature selection module CRFS, which selectively injects cross features to suppress irrelevant context interference. Experimental results show that this method achieves 69.5% mIoU on the SemanticKITTI validation set, a 3.4% improvement compared to the Baseline, and achieves significant gains in multiple context-sensitive categories. Ablation experiments further verify the effectiveness of each proposed module.
