Knowledge Graph (KG)-based Retrieval-Augmented Generation (RAG) shifts the contents of retrieval from narrative text to a relational knowledge network, empowering large language models (LLMs) to harness structured relationships between entities. However, conventional KG-RAG approaches are resource-intensive, requiring either query decomposition with multiple LLM rounds or parameterized static knowledge injection to update the model. Although subgraph reasoning aims to address these issues, most current methods are based on heuristic shortest path and multi-hop graph traversal algorithms. The retrieved subgraphs suffer from incompleteness and semantic drift, and neglect the interaction between subgraph and LLMs in terms of fine-grained structural semantics. We propose a dual-constraint subgraph optimization for KG-RAG (DCTR). It improves subgraph retrieval and generates high-quality subgraphs with structural integrity and information salience for LLMs. Specifically, it formulates subgraph generation as a two-stage graph-theoretic constrained optimization problem to create compact and complete pseudo-labels. Since these pseudo-labels are discrete, a smooth approximation is employed to convert them into a differentiable representation, thereby optimizing the retriever to highlight key information while extracting subgraphs. On two benchmark datasets, DCTR significantly enhances subgraph quality, achieving state-of-the-art performance in LLM reasoning.

Multi-Hop Question Answering (MHQA) requires step-by-step reasoning across multiple pieces of information to answer complex questions. The cache-aided Retrieval-Augmented Generation (RAG) can accelerate the process of external knowledge retrieval at each reasoning step for MHQA. However, existing methods focus on the internal structure and ignore the misalignment between the queries’ arrival order and cache hit order. To tackle this, we propose Mnemosyne, a cache hit order fitting method designed to accelerate the RAG progress for MHQA. Specifically, our cache-aware order fitting strategy adjusts the order of queries arrival via graph reordering to better align with the cache hit order, thereby reducing the likelihood of failed or unproductive retrieval attempts. The multi-granularity caching storage mechanism is designed to loosen the strict hit condition to multiple similar semantic matching modes, facilitating that relevant documents can still be retrieved. Experiments conducted on four multi-hop QA datasets demonstrate that Mnemosyne effectively reduces retrieval latency while enhancing task answer F1 score, achieving a superior trade-off between efficiency and effectiveness.

Audio-Visual Question Answering (AVQA) aims to answer textual questions by perceiving and understanding audio-visual information, which requires not only effective question-guided multimodal reasoning, but also fine-grained spatio-temporal modeling to capture dynamic changes. However, existing methods mainly operate at the patch level with strict partitioning, overlooking object-level semantic features. Furthermore, direct fusion of question features in most studies restricts attention to question-specific details. To address these challenges, we propose a unified object-centric spatio-temporal graph reasoning framework. It begins with a Motion-guided Object Tracking (MOT) module that identifies and localizes fine-grained visual objects with significant motion, which are more likely to relate to sounding behaviors and questions. Next, we design a Sound-guided Object Tracking (SOT) module that associates visual objects with corresponding audio signals, enabling the model to highlight key objects that are relevant to sound information. Then, we devise a Question-guided Spatio-Temporal Graph (QSTG) module, which first constructs a unified graph to jointly model visual object nodes across frames and audio object nodes. Based on this graph, QSTG continuously captures the dynamic evolution of objects across spatial and temporal dimensions, while adaptively modulating information propagation weights throughout the graph under question guidance, thereby enabling refined multimodal inference. Extensive experiments on the MUSIC-AVQA and MUSIC-AVQA-R datasets demonstrate that our framework is effective and achieves state-of-the-art performance across a variety of complex reasoning tasks.

Temporal knowledge graph (TKG) reasoning that infers future missing facts is an essential and challenging task. Predicting future events typically relies on closely related historical facts, yielding more accurate results for repetitive or periodic events. However, for future events with sparse historical interactions, the effectiveness of this method, which focuses on leveraging high-frequency historical information, diminishes. Recently, the capabilities of diffusion models in image generation have opened new opportunities for TKG reasoning. Therefore, we propose a graph node diffusion model with dual-domain periodic contrastive learning (DPCL-Diff). Graph node diffusion model (GNDiff) introduces noise into sparsely related events to simulate new events, generating high-quality data that better conforms to the actual distribution. This generative mechanism significantly enhances the model's ability to reason about new events. Additionally, the dual-domain periodic contrastive learning (DPCL) maps periodic and non-periodic event entities to Poincaré and Euclidean spaces, leveraging their characteristics to distinguish similar periodic events effectively. Experimental results on four public datasets demonstrate that DPCL-Diff significantly outperforms state-of-the-art TKG models in event prediction, demonstrating our approach's effectiveness. This study also investigates the combined effectiveness of GNDiff and DPCL in TKG tasks.

Emotion recognition in conversation (ERC) aims to identify the emotion expressed in each utterance within a multi-turn dialogue. In recent years, the widespread adoption of language models (LM) has spurred the development of various prompting paradigms as effective adaptation strategies, aligning LM training objectives with the specific requirements of ERC. However, while the hard prompt-based paradigm offers high human interpretability, it suffers from limited task adaptability due to its discrete nature. In contrast, the soft prompt-based paradigm sacrifices interpretability in favor of improved adaptability by optimizing continuous embedding vectors. Both paradigms generally adopt an invariant prompt across utterances, which restricts their ability to model contextual diversity and results in suboptimal adaptability of the instance. To address these issues, we propose a comprehensive prompting strategy that balances interpretability and adaptability, consisting of two components: reinforced prompt exploration for hard prompts and feature integration for soft prompts. In reinforced prompt exploration, a policy network is trained via reinforcement learning to explore the discrete prompt space under cold-start conditions, efficiently optimizing hard prompts and improving task-level adaptability while preserving interpretability. In feature integration for soft prompts, we incorporate rich semantic features to form contextually relevant soft prompts, assigning each utterance a distinct offset subspace to improve instance-level adaptability. Experiments on three datasets demonstrate that our prompting strategy achieves state-of-the-art performance in ERC.

Emotion Recognition in Conversation (ERC) is a challenging task in conversational AI systems. However, current ERC methods that rely on language models (LM) often use fixed prompts, which lack dynamic adaptability and fail to adequately incorporate external knowledge influencing conversational emotions. To address these limitations, this paper proposes the Dual-Intervention Prompt Framework for emotion recognition in conversation (DIPF). DIPF introduces two types of prompts: explicit prompts and implicit prompts. Explicit prompts leverage external knowledge via knowledge graphs, allowing the model to better simulate conversational flow and emotional dynamics. Implicit prompts are integrated into the hidden layers of the LM, becoming part of the model’s representation and participating in the gradient descent process. Additionally, DIPF manipulates the model representations corresponding to implicit prompts using a low-rank projection matrix, guiding the model’s behavior during inference and improving the optimization of implicit prompts. Extensive experiments on three benchmark datasets demonstrate that DIPF achieves relative F1 score improvements of 1.27% on IEMOCAP, 2.26% on MELD, and 0.72% on EmoryNLP over the strongest baseline (EACL), with an average relative improvement of 1.52%. These results confirm the effectiveness of DIPF in enhancing emotional understanding through dynamic and knowledge-aware prompting strategies.

With the increasing use of interactive applications, the importance of Emotion recognition in conversation (ERC) is growing. Current research in the ERC domain mainly emphasizes the extraction of contextual information. However, challenges arise due to multi-turn conversation scenarios and the natural transformation of emotions, particularly in identifying subtle emotion transfers. Moreover, emotions exhibit nonlinear characteristics in semantic spaces, leading to potential confusion when discerning similar semantic emotions in the Euclidean semantic space. To address these issues, this study proposes a Poincaré-based dual-strategy emotion tracker for emotion recognition in conversation (PeTracker), which introduces the hyperbolic space representation in the ERC domain. Based on the spatial properties of the hyperbolic space representation to capture the nonlinear relationships among features, PeTracker encompasses two learning strategies. Poincaré emotional geometry curriculum learning (PGCL) and Poincaré emotional stratification contrastive learning (PSCL). In PGCL, the similarity of emotion labels is effectively discerned using the Poincaré distance, quantifying emotion transfer distances and facilitating the identification of subtle emotion transfers in utterance. In PSCL, PeTracker extracts and adapts multi-level features, mapping them to the Poincaré ball space to build emotion prototype-based contrastive learning. This process enhances the model’s ability to distinguish between similar emotion labels. while alleviating potential label confusion issues. Experimental results on three general datasets demonstrate that PeTracker achieves optimal or near-optimal performance. Furthermore, the study investigates the role and impact of the Poincaré ball in differentiating similar emotions.

In few-shot knowledge graph completion (FKGC), inferring new relationships from limited data is a key challenge. Current FKGC frameworks often struggle to detect subtle differences among entity types and handle noise in sparse environments, impacting their performance. Aligning relational representations with entity and relationship data, while managing prediction uncertainties, is another significant hurdle. We propose the NAEP model, which integrates a hierarchical relational gating graph attention network to discern complex relational dynamics and differentiate entity types. It employs neighborhood multihead attention for path encoding to reduce noise and incorporates an Evidential Conditional Neural Processes encoder to handle predictive uncertainties. Preliminary results show that NAEP performs robustly across various datasets, from large ones like NELL and WIKI to smaller ones like FB15K-237, demonstrating its versatility and superior performance.