Logit Knowledge Distillation has gained substantial research interest in recent years due to its simplicity and lack of requirement for intermediate feature alignment; however, it suffers from limited interpretability in its decision-making process. To address this, we propose implicit Clustering Distillation (iCD): a simple and effective method that mines and transfers interpretable structural knowledge from logits, without requiring ground-truth labels or feature-space alignment. iCD leverages Gram matrices over decoupled local logit representations to enable student models to learn latent semantic structural patterns. Extensive experiments on benchmark datasets demonstrate the effectiveness of iCD across diverse teacher-student architectures, with particularly strong performance in fine-grained classification tasks---achieving a peak improvement of +5.08% over the baseline.

Digitizing historical Mongolian newspapers is challenging due to severe degradation and unique script properties. Elongated words and artifacts, like ink bleeding cause geometric distortion. These issues limit traditional OCR performance. To address these issues, we propose MGL-LMM, a Generative OCR Paradigm specifically designed for degraded agglutinative scripts. Our framework integrates a topology-preserving visual normalization strategy. This maintains structural integrity under anisotropic distortion. To stabilize training, we employ a decoupled curriculum learning scheme that transitions from word-level glyph morphologies to line-level contextualization. By projecting high-fidelity visual features directly into the semantic space of a 1.5B-scale Large Language Model, MGL-LMM performs end-to-end generative reasoning. This allows the model to leverage contextual language priors to effectively reduce visual ambiguities and provide semantic restoration for obscured characters. Extensive experiments demonstrate that our method achieves state-of-the-art performance, reducing the Character Error Rate (CER) from 27% to 11%. This significant improvement demonstrates that generative decoding with language priors is fundamentally more effective than purely visual recognition under severe degradation.

In recent years, deep learning is widely applied in surface defect detection. However, there are still challenges with low detection efficiency, often leading to false positives and false negatives. To address these issues, we propose a multi-scale convolution and multi-dimensional feature fusion detection model (MAF-DETR) based on RT-DETR. We designed the Dilated Gated Extraction Module (DGEM) to enhance multi-scale feature extraction and global information awareness. Additionally, we developed a Dynamic Feature Aggregation Module (DFAM) to improve the correlation between semantic and feature information, enhancing the ability to detect micro-texture variations. Furthermore, we optimized RepC3 with the C3K2 module for better feature fusion. Experimental results demonstrate that our model achieves mAP50and mAP50−95of 82% and 37.4% respectively on the RoboFlow wire rope defect dataset, surpassing the baseline model by 4.4% and 4.5%. On the NEU-DET dataset, the model attains mAP50and mAP50−95of 74.1% and 41.6%, demonstrating improvements of 5% and 2.8% over the baseline model.

With the advancement of deep learning, surface defect detection has achieved remarkable progress. However, traditional detection methods employ single feature extraction strategies, leading to feature aliasing, information redundancy, and interference that fails to model complex non-linear feature dependencies. This reduces detection accuracy and increases error rates. We propose a Wavelet Domain Enhancement-Based Graph Guided Fusion Model for Defect Detection (WG-DETR). Our Anisotropic Direction Enhancement Unit (ADEU) enhances spatial features through optimized multi-directional filters, improving sensitivity to linear defects and robustness against non-homogeneous textures. The Wavelet Bidirectional Frequency Enhancement Unit (WBFU) uses Haar wavelets for frequency domain decomposition, creating bidirectional feature interaction channels that effectively isolate and enhance different frequency domain representations. Our Feature-similarity Graph-guided Unit (FGU) constructs graph topological structures for receptive field decoupling and joint modeling of local and global feature dependencies. Experimental results show WG-DETR achieves 82.4% and 35.7% on mAP50and mAP50−95metrics, improvements of 4.8% and 2.8% over the baseline.

Conventional Neural Machine Translation (NMT) relies on previous tokens and the hidden state of the target for the inference of the target tokens in the decoding phase, and this left-to-right decoding approach loses the context of the target sequence from right-to-left. In addition, the vanilla attention mechanism lacks interactivity in the bilingual training phase, where the computation of the attention weights is independent at each step, which leads the decoder to disregard whether or not the current token has already been translated. In this paper, we proposed a novel interactive synchronized bi-directional inference method that uses past and future contexts to synchronously predict target sequences and incorporates Neural Turing Machine (NTM) ideas to detect historical attentional information, which relies on a read − write mechanism to update the source hidden state. We evaluate the proposed model in the WMT14 German-English translation task and the LDC Chinese-English translation task, and the experimental results show that our method improves 2.18 and 4.07 BLEU scores over the Transformer, respectively, which fully demonstrates the effectiveness of the proposed method.