Large language models (LLMs) face critical limitations in dynamically integrating external knowledge, mitigating hallucinations, and addressing domain-specific expertise. While retrieval-augmented generation (RAG) mitigates these issues, traditional chunk-based methods suffer from retrieval noise, multi-hop reasoning failures, and poor interpretability. Graph-based RAG (GraphRAG) has improved RAG in terms of context retrieval and relational reasoning, but there are still shortcomings in jointly processing these two types of tasks. To bridge this gap, we propose TriG-RAG, a unified framework that synergizes Triple-Granularity graph indexing with an adaptive retrieval strategy. TriG-RAG constructs a heterogeneous graph with a document-chunk-entity hierarchical structure and a weight allocation strategy, enabling flexible handling of both context-dependent and relation-dependent queries. Extensive experiments on six benchmarks demonstrate the superiority of TriG-RAG in terms of retrieval capability, flexibility and interpretability.
This paper aims to address the challenge of correcting semantic, syntactic, and knowledge-related errors in Chinese government texts. The existing methods often overlook content-level inaccuracies and knowledge inconsistencies. To address the above issue, we propose a knowledge-aware rewriting method with large language models (LLMs) for Chinese government text correction (in short KARTC). The proposed KARTC comprehensively integrates structured knowledge bases (KBs) with LLMs with a three-stage hierarchical task-chain framework. To demonstrate the effectiveness of the proposed KARTC, we conduct the experiments on NLPCC 2025 shared task 5. The results show that KARTC achieves the 79.64
Temporal Knowledge Graph Reasoning (TKGR) aims to predict future events or relationships by leveraging historical facts within temporal knowledge graphs. By integrating the structured semantics of knowledge graphs with temporal dynamics, TKGR is well-suited for tasks involving time-dependent inference, such as temporal relation extraction and event prediction. In recent years, large language models (LLMs) have been increasingly applied to TKGR. However, a single LLM agent often struggles to handle complex reasoning tasks effectively, primarily due to its relatively narrow reasoning perspective and shallow reasoning depth. To address these challenges, this paper proposes a Multi-agent Multi-round Debate framework for Temporal Knowledge Graph Reasoning (MMD-TKGR). The framework optimizes the construction of query-specific historical contexts by introducing a relation-temporal cooperative retrieval mechanism, and it dynamically calibrates prediction results and enhances their interpretability through a multi-round debate structure based on multi-agent collaboration. Experimental results demonstrate that the proposed method significantly outperforms existing approaches on several benchmark datasets.
Supervised Fine-Tuning (SFT) has become a critical step in adapting pre-trained base models to downstream tasks. However, treating samples equally, without considering the inherent heterogeneity in data quality and task characteristics, can lead to poor data efficiency and affect model performance. In this paper, we propose a mutual information-based adaptive fine-tuning method, which measures the information richness of data across multiple granularities, and tune the weight of smoothing accordingly to improve model alignment. We apply less label smoothing to data with more information richness, ensuring that the informative content is fully learned, while for data with less information richness, we retain their potential variability. Our method is model-agnostic, thus offering excellent generalizability and cost-effectiveness. Comprehensive experiments across multiple domains demonstrate that our method achieves consistent and robust improvements over standard Supervised Fine-Tuning, particularly +6.37
Multimodal sentiment analysis aims to integrate information from textual, acoustic, and visual modalities to improve emotional understanding. However, existing methods still face challenges in handling feature redundancy and achieving effective cross-modal representation learning. In this paper, we propose MAGI (Modality-Aligned Geometry-aware Integration), a novel framework that explicitly disentangles multimodal inputs into modality-shared and modality-specific representations for more robust modeling. MAGI employs a shared encoder and modality-specific encoders to perform representation disentanglement, guided by geometry-aware regularization losses including reconstruction, specificity, and soft orthogonality constraints. To align the shared representations across modalities, we introduce a volume-based contrastive loss that encourages geometric consistency in the latent space. Meanwhile, the modality-specific representations are enhanced via a cross-modal attention mechanism and then fused using a gating module that adaptively integrates modality-aware information. Finally, MAGI incorporates a triple-space prediction architecture that jointly utilizes shared, specific, and fused representations to improve prediction accuracy. Experimental results on CMU-MOSI and CMU-MOSEI demonstrate that MAGI achieves competitive performance across multiple evaluation metrics. Further ablation studies confirm the effectiveness of each core component.
