Class imbalance and class overlap in tabular data cause deep neural networks to exhibit severe memorization-induced overfitting during later training stages. Existing methods fail to explicitly connect overlap phenomena with the learning-then-memorization dynamics of neural networks, limiting their ability to regulate model behavior and improve generalization. This paper proposes MemGrad, an imbalanced classification method for learning generalizable boundaries via dual gradient regulation guided by dynamic memory perception. A sample importance score is constructed by integrating the memorization score with local overlap degrees. Based on dynamic thresholds during optimization, samples are partitioned into high-importance and low-importance groups, respectively representing overlap samples prone to memorization and safe samples providing stable learnable features. A gradient direction convergence mechanism geometrically rotates high-importance sample gradients toward safe directions defined by low-importance samples, injects structured noise perpendicular to the safe direction for implicit regularization, and performs conflict resolution to ensure optimization stability, thereby suppressing specialized decision boundaries and mitigating memorization. Simultaneously, a gradient magnitude balancing mechanism equalizes the total gradient norms of the majority and minority classes to alleviate gradient inundation and decision boundary bias. Experimental results on 40 imbalanced tabular datasets demonstrate that MemGrad outperforms 29 representative baseline methods in F1-score and G-mean metrics, particularly on 22 datasets exhibiting severe class overlap.
Large Language Models (LLMs) are increasingly deployed in high-stakes domains, yet they remain prone to hallucinations. Existing mitigation methods face a fundamental trade-off: suppressing factuality hallucinations—where the output contradicts world knowledge—often worsens faithfulness hallucinations, where the output deviates from the given context, and vice versa. Our prior work SPACE (Wang et al., 2025) first addressed this by editing shared activation subspaces, but could not distinguish error-carrying pathways from useful ones within a subspace. In this paper, we show theoretically that this limitation stems from polysemantic neurons, and propose Lancet, a framework that shifts the intervention from subspaces to pathways. Lancet locates source neurons via contrastive gradient analysis, maps their error-propagation pathways by partitioning a cross-layer graph using Structural Entropy minimization, and applies graded suppression that decays with distance from the source. Experiments on four benchmarks across three models show that Lancet consistently improves over SPACE and all other baselines on both factuality and faithfulness.
Corrosion represents a formidable challenge for the marine or military equipment operating under the harsh “three highs” service environments (i.e. temperature, humidity and salinity). Corroded metallic materials will lose their mechanical strength as critical structural components or fail to perform the functions as designed, and thus causing not only additional economic costs from frequent maintenance and even premature retirement but also serious safety hazards. Anticorrosion coatings, applied to the surface of metallic substrates and serving as a protective barrier, are regarded as a viable solution to address this persistent challenge. MXenes have emerged as promising candidates for anticorrosion coating materials, owing to their unique layered configuration, excellent barrier properties and tunable surface chemistry. Herein, detailed density functional theory (DFT) calculations were performed to investigate the potential applications of the undoped and heteroatom-doped MXene monolayer complexes. Upon screening the optimal dopant atoms, the structural integrity remains well preserved with substitutional dopants incorporated. The undoped MXenes exhibit decent binding capability toward corrosive agent, and the binding energies and migration energy barriers of Cl-ion can be further enhanced through low-concentration substitutional doping. This study elucidates the anticorrosion performance of the layered Ti3C2 MXene materials and the influence of heteroatom dopants.
White balance (WB) correction for sRGB images captured under multiple illuminations remains challenging due to spatially varying and nonlinear color casts. Existing sRGB-domain WB methods, relying on global adjustment or content-dependent priors, struggle in such scenes because they lack spatial adaptivity correction and explicit illumination guidance. Further, both types of methods lack the consideration of multi-illumination characteristics due to only single-illumination data being available for supervision. To address these limitations, we propose AdaWBNet, an illumination-guided spatially adaptive WB framework formulated under a Maximum A Posteriori estimation paradigm. Within this framework, learnable illumination prototypes serve as priors that capture representative single-illumination patterns, and a K-means cross-attention module enhances these prototypes by assigning local features of multi-illumination inputs to similar prototypes. Then, guided by the illumination prototypes, the likelihood term employs query-adaptive layers to achieve illumination-aware, spatially adaptive color correction. This design enables AdaWBNet to achieve both spatially adaptive color correction and illumination guidance using only single-illumination supervision, thereby allowing it to effectively handle complex multi-illumination scenes. Extensive experiments demonstrate that AdaWBNet outperforms state-of-the-art methods on multi-illumination benchmarks, achieving spatially consistent and perceptually faithful results. The code is available at here.
Underwater objects often suffer from severe visual degradation, including strong background noise induced by water impurities, blurred boundaries caused by optical scattering, and small object suppression in complex environments. These factors collectively pose significant challenges to underwater object detection (UOD). To address these issues, we propose the Multi-Scale Inverted Pyramid Network (MIP-Net) tailored for UOD. Moving beyond standard feature fusion paradigms, MIP-Net introduces explicit conceptual shifts through two key components: the Local Adaptive Contrast module (LAC) and the Multi-Scale Inverted Feature Pyramid Network (MSIFPN). Unlike conventional global attention mechanisms, LAC selectively calibrates intra-feature contrast layer-by-layer, establishing a theoretical basis for dynamic feature modulation that prevents high-frequency detail dilution. Simultaneously, MSIFPN transcends traditional FPN architectures by coupling a dual-pyramid bidirectional flow with a strict mathematical foreground-background separation strategy, effectively isolating target semantics from ambiguous water impurities. Experiments on the official DUO benchmark demonstrate that MIP-Net achieves an AP of 70.1%, surpassing state-of-the-art single-stage, two-stage, and Transformer-based methods, while maintaining a highly competitive trade-off between computational efficiency (Params/FLOPs) and accuracy. Furthermore, evaluations on the terrestrial COCO dataset yield an AP of 45.6%, confirming the strong cross-domain generalization capability of our framework. The code is publicly available at: https://github.com/YitengGuo/MIP-Net
