Attributing terrorism-related incidents to responsible armed groups remains challenging because incident narratives contain ambiguous descriptions, inconsistent organization naming, heterogeneous contextual information, and substantial class imbalance. This study proposes a unified transformer-based framework for two complementary tasks: armed-group named entity recognition (NER) and attack attribution using the Global Terrorism Database (GTD). A multi-modal preprocessing pipeline transforms 209,706 incidents with 135 heterogeneous variables into an 80-feature analysis-ready representation and supports structured-information supplementation when relevant incident information is absent from the narrative. This supplementation provides limited benefit for information-rich narratives but improves attribution accuracy by up to 13.43 percentage points under severe narrative sparsity, demonstrating its value as an information-completion mechanism. For armed-group extraction, a RoBERTa–BiLSTM–CRF architecture augmented with gazetteer features is developed and evaluated on a silver-standard NER corpus of 67,386 incidents, achieving 99.23
Hybrid EEG–fNIRS brain–computer interfaces combine rapid electrophysiological activity with hemodynamic information, but neurovascular latency and entangled multimodal features hinder effective fusion. We propose the Neurovascular Coupling-Aware Delay Alignment and Subspace Disentanglement Network (NVC-DSDNet), which couples delay-aware attention with shared–private neural subspaces. EEG responses serve as temporal anchors that retrieve task-relevant fNIRS features over candidate delays, while Neural Feature Subspace Decomposition separates shared neural intent from modality-specific information. An adaptive multi-head decision module then weights unimodal and fused predictions at the sample level. In subject-independent five-fold cross-validation, NVC-DSDNet achieved accuracies of 72.67
Agglomeration detection in scanning electron microscopy (SEM) images of high-performance fiber ultrastructure is challenged by cross-scale feature shifts and complex textures. To address these issues, we propose the Cross-Scale Fiber SEM Detection Transformer (CFS-DETR), a specialized detection framework that is designed for high-performance fiber cross-scale ultrastructure. The framework effectively addresses key challenges for ultrastructural image detection by designing a Cross-Scale Invariant Module (CIM) and a Fiber SEM Edge Enhancement Network (FSE-Net). Specifically, the CIM employs adversarial training mechanisms to guide the backbone network in learning cross-scale invariant feature representations, thereby mitigating scale-related feature shift issues. The FSE-Net achieves selective enhancement of edge features, improving the model’s capability for representing ultrastructural edge information. Experiments are conducted on the DHU-OD-696 dataset, which contains 696 images and 3705 agglomeration instances collected across scales from 300 nm to 500 µm. On this dataset, CFS-DETR achieves an average precision of 88.2 at an intersection-over-union threshold of 0.50 ( AP_50 ) and a precision of 87.6, representing improvements of 2.2 and 5.8
Transformer-based large language models construct rich contextualized token representations across multiple layers, yet redundancy within their intermediate hidden-state spaces remains insufficiently explored as an inference-time representation property. This paper introduces token representation reuse (TRR), an inference-time framework that identifies highly similar contextual token representations and replaces selected hidden states with preceding anchor representations under similarity and reuse-budget constraints. Unlike token pruning or merging, TRR preserves the original sequence length and positional structure and requires neither model retraining nor parameter modification. We further provide an end-to-end complexity analysis of representation discovery and characterize the local conditions under which identical hidden states can, in principle, permit projection reuse, while explicitly showing that such equality does not generally persist across subsequent transformer layers. Experiments on BoolQ, AGNews, SST-2, and TruthfulQA using Qwen2.5-1.5B-Instruct and Llama-3-8B demonstrate representation reuse ratios of up to 12.40
Compact models on edge devices need knowledge distillation (KD) to close their cross-domain accuracy gap, but the choice of teacher decides whether distillation helps or harms. We characterise that failure at a scale that tests selection rules rather than illustrates them: 180 teacher–student pairings on ePillID over three training seeds, and 180 more on an independently derived CURE benchmark. Harmful transfer is common (21.8
