Currently, Tibetan sentiment analysis faces two key challenges, one is the inadequate semantic encoding capability of pre-trained language models, another one is the performance bottleneck caused by the scarcity of fine-tuning data. To address the scarcity of fine-tuned data, the paper constructs a Tibetan tri-classifiedified sentiment dataset containing 24,853 samples. To address the inadequate semantic encoding capability of pre-trained language models, the paper proposes a contrastive distillation knowledge method based on cross-language sentence vector, which contains two key phases. Firstly, the paper proposes Tibetan-Chinese contrastive learning based on continuous cueing templates, which avoids the complexity of manually designing templates in a Tibetan-Chinese bilingual environment and achieves cross-language vector-space alignment to enhance the encoding ability of pre-trained language models for Tibetan. At the same time, it also introduces internal contrastive learning for Tibetan to balance the cross-language migration with the preservation of linguistic features. Secondly, introducing a learnable linear mapping layer and comparative learning strategy achieves the alignment of the Tibetan student model with the Chinese teacher model in the feature vector space and semantic space, which effectively mitigates the problem of semantic division in cross-language scenarios. And the method also adopts an orthogonal linear mapping mechanism based on Frobenius' paradigm constraints for cross-language knowledge distillation, which is able to achieve effective knowledge transfer from the Chinese teacher model to the Tibetan student model while maintaining the structure of the vector space. The experimental results show that the method achieves accuracy of 87.1% and F1 value of 0.868 on the constructed Tibetan tri-classifiedified sentiment dataset. The method outperforms the full-volume fine-tuning baseline as well as mainstream parameter-efficient fine-tuning methods(e.g., LoRA, PET and P-tuning), and it also outperforms model variants with more complex structures(e.g., CINO+BiLSTM and CINO+ Adapter).
In this work, we propose a parallel low-depth flagged syndrome extraction protocol for fault-tolerant quantum error correction. The protocol combines parallel measurement of different stabilizers with dynamic selection of a minimal stabilizer set, guided by flag outcomes. Circuit-level optimizations, including CNOT reordering, are integrated while rigorously preserving fault-tolerance guarantees. Implementations for several distance-3 and distance-5 codes achieve substantial reductions in circuit depth and idle locations. Numerical simulations under a circuit-level depolarizing noise model demonstrate consistent pseudo-threshold improvements, such as 12.56% and 4.43% over prior schemes for [[5, 1, 3]], 16.2% for [[7, 1, 3]], and 14.55% to 30.13% for distance-5 codes, confirming the effectiveness of the co-optimized approach. This co-optimized framework offers a practical pathway towards more resource-efficient fault-tolerant quantum computation.
Power Distribution Networks (PDNs) face increasingly stealthy and causally complex cyber threats that challenge traditional detection approaches. While Large Language Models (LLMs) offer promising capabilities to reason about such threats, directly feeding fragmented, noisy, and heterogeneous machine logs to them often overwhelms their reasoning capabilities, obscuring critical causal chains. This paper presents MCCD, a Multi-Layer Collaborative Continuous Defense framework driven by an LLM-interactive Trusted Intelligence Agent (TIA). The TIA operates inside a TEE to ensure the authenticity and provenance of endpoint intelligence. A lightweight Temporal-Causal-Hierarchical Encoding (TCHE) transforms heterogeneous logs into structured Semantic Security Intelligence (SSI), enabling the cloud-based LLM to perform coherent causal reasoning. Through an adaptive closed-loop mechanism, the LLM iteratively refines the collected evidence and converges to a reliable final threat determination, rather than relying on one-shot inference. Evaluations on PDN-style datasets and DARPA Transparent Computing traces demonstrate that MCCD provides substantial gains in industrial power-system scenarios. These results demonstrate that combining Trusted Execution Environment (TEE), semantic intelligence, and iterative reasoning provides an effective and practical path toward robust threat determination in modern PDNs.
Dialogue summarization in informal conversational settings presents unique challenges, including fragmented syntax, colloquial expressions and dynamically shifting speaker roles. Existing summarisation models, largely adapted from monologue-oriented text, often fail to preserve these informal cues and neglect the discourse structures that underpin conversational coherence. To address these gaps, this study proposes DISCO-DIAL, a discourse and informality-aware neural framework for dialogue summarization. The framework introduces three novel components: (i) INFO-PRESERVE, an informality preserving preprocessing layer that retains colloquial elements, contractions and emotive signals such as emoji's, while adaptively filtering noise, (ii) Modified Bayesian Relevance Vector Machine (MBRVM), a discourse-aware graph neural encoder that models utterances as graph nodes connected through adjacency, speaker role and topic shift edges, thereby capturing both semantic salience and interactional dependencies and (iii) REL-AttnRank, a relevance guided attention ranking network that integrates hierarchical dual attention with Contrastive Relevance Loss and Semantic Fidelity Loss to ensure precise, semantically aligned sentence selection. The framework operates as an extractive summarizer, selecting salient utterances while preserving informal cues and discourse structure. Extensive experiments on the SAMsum benchmark dataset demonstrate that DISCO-Dial achieves ROUGE-1, ROUGE-2 and ROUGE-L scores of 87.5, 70.67 and 87.50, respectively, alongside a BERTScore of 99, significantly outperforming state-of-the art baselines. Qualitative analysis further reveals that the framework excels in maintaining dialogue intent, speaker relationships and contextual coherence, even in the presence of noisy, informal constructs. These contributions establish DISCO-Dial as a robust and scalable solution for real-world dialogue summarization with promising applications in conversational AI, customer support systems and meeting transcription.
Most existing studies on modulation recognition for orthogonal frequency-division multiplexing (OFDM) signals operate under the idealized assumption of perfect synchronization, thereby overlooking the practical challenges posed by real-world wireless environments. Such assumptions neglect the detrimental effects of timing and carrier frequency offsets, which inevitably arise in practice and can severely degrade recognition accuracy. In this work, we address this gap by proposing a novel modulation recognition framework tailored for coded OFDM signals operating under imperfect synchronization conditions over unknown channel coefficients. The core of the proposed approach is a maximum-likelihood (ML) formulation that jointly estimates the transmitted modulation format, the underlying channel response, and the synchronization parameters-namely, the timing and carrier frequency offsets. To efficiently solve this inherently coupled estimation problem, we adopt an expectation-maximization (EM) algorithm, which alternates between soft-data detection and parameter refinement. Unlike conventional schemes that rely exclusively on known pilot symbols for parameter estimation, our method leverages the soft outputs of a bit-interleaved coded modulation with iterative decoding (BICM-ID) receiver as virtual pilots. These decoder-generated reliability values are recursively fed back into the EM process, enabling continuous improvement of modulation recognition, synchronization, and channel estimates without requiring additional pilot overhead. Extensive simulation results demonstrate that the proposed scheme achieves superior modulation recognition accuracy compared to traditional synchronization-agnostic methods, particularly in scenarios with moderate-to-severe timing and frequency misalignments.
