EPMORE: Explainable Process Mixture-of-Experts

The mainstream approach of GNN-based recommendation aggregates high-order ID information associated with the node in the user-item graph. The aggregation pattern using ID as signal has two disadvantages: lack of textual semantics and the impact of interaction noise. These disadvantages pose a threat to effectively learn user preferences, especially in capturing intricate user-item semantic relationships. Although large language models (LLMs) allow the integration of rich textual information into recommenders and have had groundbreaking applications in recommender systems, current works need to bridge the gap between different representation spaces. This is because LLM-based methods align the representations of GNN-based models only by using text embedding of LLM, leading to unsatisfactory results. To address this challenge, we propose a denoising alignment framework with LLMs for GNN-based recommenders (DALR) , which aims to align structural representation with textual representation and mitigate the effects of noise. Specifically, we propose a modeling framework that integrates the representation of graph structure with textual information from LLMs to capture intricate user-item interactions. We also suggest an alignment paradigm to enhance representation performance by aligning semantic signals from LLMs and structural features from GNN models. Additionally, we introduce a contrastive learning scheme to relieve the impact of noise and improve model performance. Extensive experiments on public datasets demonstrate that our model consistently outperforms the state-of-the-art methods. DALR achieves improvements ranging from 2.82% to 12.20% in Recall@5 and from 1.04% to 3.48% in NDCG@5 compared to the strongest baseline model, using the Steam dataset as an example.

Recent advancements in pre-trained large language models (LLMs) have significantly influenced various domains. Adapting these models for specific tasks often involves fine-tuning (FT) with private, domain-specific data. However, privacy concerns keep this data undisclosed, and the computational demands for deploying LLMs pose challenges for resource-limited data holders. This has sparked interest in split learning (SL), a Model-as-a-Service (MaaS) paradigm that divides LLMs into smaller segments for distributed training and deployment, transmitting only intermediate activations instead of raw data. SL has garnered substantial interest in both industry and academia as it aims to balance user data privacy, model ownership, and resource challenges in the private fine-tuning of LLMs. Despite its privacy claims, this paper reveals significant vulnerabilities arising from the combination of SL and LLM-FT: the Not-too-far property of fine-tuning and the auto-regressive nature of LLMs. Exploiting these vulnerabilities, we propose Bidirectional Semi-white-box Reconstruction (BiSR), the first data reconstruction attack (DRA) designed to target both the forward and backward propagation processes of SL. BiSR utilizes pre-trained weights as prior knowledge, combining a learning-based attack with a bidirectional optimization-based approach for highly effective data reconstruction. Additionally, it incorporates a Noise-adaptive Mixture of Experts (NaMoE) model to enhance reconstruction performance under perturbation. We conducted systematic experiments on various mainstream LLMs and different setups, empirically demonstrating BiSR's state-of-the-art performance. Furthermore, we thoroughly examined three representative defense mechanisms, showcasing our method's capability to reconstruct private data even in the presence of these defenses.

Recent advancements in sequential recommendation have underscored the potential of Large Language Models (LLMs) for enhancing item embeddings. However, existing approaches face three key limitations: 1) the degradation of the semantic space when high-dimensional language embeddings are mapped to lower-dimensional ID embeddings, 2) the underutilization of language embeddings, and 3) the reliance on additional trainable parameters, such as an adapter, to bridge the gap between the semantic and behavior spaces. In this paper, we introduce AlphaFuse, a simple but effective language-guided learning strategy that addresses these challenges by learning ID embeddings within the null space of language embeddings. Specifically, we decompose the semantic space of language embeddings via Singular Value Decomposition (SVD), distinguishing it into a semantic-rich row space and a semantic-sparse null space. Collaborative signals are then injected into the null space, while preserving the rich semantics of the row space. AlphaFuse prevents degradation of the semantic space, integrates the retained language embeddings into the final item embeddings, and eliminates the need for auxiliary trainable modules, enabling seamless adaptation to any sequential recommendation framework. We validate the effectiveness and flexibility of AlphaFuse through extensive experiments on three benchmark datasets, including cold-start user and long-tail settings, showcasing significant improvements in both discriminative and diffusion-based generative sequential recommenders.

In recent years, the field of natural language processing has witnessed remarkable advancements due to the success of large language models. These models leverage the Transformer architecture and pre-training techniques to achieve impressive results. In this paper, we draw inspiration from large language models and apply these techniques into the task of named entity recognition in the domain of power grids, which is critical for building power grid knowledge graphs and question-answering systems. Specifically, we propose a BERT-CNN-BIGRU-CRF deep learning model for named entity recognition. This model effectively harnesses the semantic modeling capabilities and pre-training knowledge of BERT, which is based on the Transformer architecture. By incorporating CNN and BIGRU, the model captures and models both local and global features, respectively. The CRF layer is employed for label classification. This combination of components ensures a high level of recognition accuracy. To evaluate the performance of the proposed model, we train our model on annotated maintenance plan data. We compare its results with those of other commonly used models. The evaluation metrics include recall, precision, and F1 score, which are widely employed in named entity recognition tasks. Our proposed model achieves optimal performance across all three metrics, demonstrating its superiority over other models.

Researchers face a critical choice: how to use -- or not use -- large language models in their work. Using them well requires understanding the mechanisms that shape what LLMs can and cannot do. This chapter makes LLMs comprehensible without requiring technical expertise, breaking down six essential components: pre-training data, tokenization and embeddings, transformer architecture, probabilistic generation, alignment, and agentic capabilities. Each component is analyzed through both technical foundations and research implications, identifying specific affordances and limitations. Rather than prescriptive guidance, the chapter develops a framework for reasoning critically about whether and how LLMs fit specific research needs, finally illustrated through an extended case study on simulating social media dynamics with LLM-based agents.

Dense video captioning aims to identify events within a video and generate natural language descriptions for each event. Most existing approaches adhere to a two-stage framework consisting of an event proposal module and a caption generation module. Previous methodologies have predominantly employed convolutional neural networks and sequential models to describe individual events in isolation. However, these methods limit the influence of neighboring events when generating captions for a specific segment, often resulting in descriptions that lack coherence with the broader storyline of the video. To address this limitation, we propose a captioning module that leverages both Transformer architecture and a Large Language Model (LLM). A convolutional and LSTM-based proposal module is used to detect and localize events within the video. An encoder-decoder-based Transformer model generates an initial caption for each proposed event. Additionally, we introduce a Large Language Model (LLM) that takes the set of individually generated event captions as input and produces a coherent, multi-sentence summary. This summary captures cross-event dependencies and provides a contextually unified and narratively rich description of the entire video. Extensive experiments on the ActivityNet dataset demonstrate that the proposed model, Transformer-LLM based Dense Video Captioning (TL-DVC), achieves a 9.22% improvement over state-of-the-art models, increasing the Meteor score from 11.28 to 12.32.

Although traditional statistical techniques and machine learning methods have contributed significantly to genetics and, in particular, inherited disease diagnosis, they often struggle with complex, high-dimensional data, a challenge now addressed by state-of-the-art deep learning models. Large language models (LLMs), based on transformer architectures, have excelled in tasks requiring contextual comprehension of unstructured medical data. This systematic review examines the role of generative Artificial Intelligence (AI) methods in human medical genomics, focusing on the genetic research and diagnostics of both rare and common diseases. Automated keyword-based search in PubMed, bioRxiv, medRxiv, and arXiv was conducted, targeting studies on LLM applications in diagnostics and education within genetics and removing irrelevant or outdated models. A total of 195 studies were analyzed, highlighting the prospects of their applications in knowledge navigation, analysis of clinical and genetic data, and interaction with patients and medical professionals. Key findings indicate that while transformer-based models perform well across a diverse range of tasks (such as identification of tentative molecular diagnosis from clinical data or genetic variant interpretation), major challenges persist in integrating multimodal data (genomic sequences, imaging, and clinical records) into unified and clinically robust pipelines, facing limitations in generalizability and practical implementation in clinical settings. This review provides a comprehensive classification and assessment of the current capabilities and limitations of LLMs in transforming hereditary disease diagnostics and supporting genetic education, serving as a guide to navigate this rapidly evolving field, while outlining application use cases, implementation guidance, and forward-looking research directions.

Explainable Artificial Intelligence (XAI) constitutes a foundational pillar of responsible AI deployment, addressing the opacity of data-driven decision systems by making their internal logic transparent, auditable, and contestable. This chapter provides a systematic account of explainability and interpretability across the full spectrum of AI models — from classical supervised learning algorithms such as decision trees, support vector machines, and gradient-boosted ensembles to contemporary Large Language Models (LLMs) built on transformer architectures. Beginning with a structured taxonomy of XAI properties — scope, stage, model-dependency, and output type — the chapter critically examines canonical explanation methods including SHAP, LIME, attention visualization, mechanistic circuit analysis, and counterfactual generation.

Large Language Models (LLMs), built on Transformer architectures, exhibit remarkable generalization across a wide range of tasks. However, fine-tuning these models for specific tasks remains resource-intensive due to their extensive parameterization. In this paper, we explore two remarkable phenomena related to the attention mechanism during the fine-tuning of LLMs (where Wq, Wk, and Wv denote the weights of the query, key, and value layers, respectively). The first phenomenon, termed “Unequal Importance of Attention Matrices”, highlights the impact of fine-tuning different weight matrices. It shows that optimizing the Wv matrix yields significantly better performance than optimizing the Wk matrix. Fine-tuning only the Wq and Wv matrices is computationally efficient while delivering results comparable to, or even better than fine-tuning all three matrices (Wq, Wk, and Wv). The second phenomenon, “Attention Matrices with Customized Learning Rate Lead to Better Convergence”, emphasizes the importance of assigning distinct learning rates to these matrices. Specifically, a higher learning rate for the Wv matrix compared to Wq and Wk accelerates convergence and improves performance. Building on these insights, we propose a new strategy that improves fine-tuning efficiency in terms of both storage and time. Experimental results on benchmark datasets validate the effectiveness of this approach, supporting our theoretical findings. Our analysis lays the theoretical groundwork for configuring and improving algorithms in LLMs fine-tuning.

An interesting behavior in large language models (LLMs) is prompt sensitivity. When provided with different but semantically equivalent versions of the same prompt, models may produce very different distributions of answers. This suggests that the uncertainty reflected in a model's output distribution for one prompt may not reflect the model's uncertainty about the meaning of the prompt. We model prompt sensitivity as a type of generalization error, and show that sampling across the semantic concept space with paraphrasing perturbations improves uncertainty calibration without compromising accuracy. Additionally, we introduce a new metric for uncertainty decomposition in black-box LLMs that improves upon entropy-based decomposition by modeling semantic continuities in natural language generation. We show that this decomposition metric can be used to quantify how much LLM uncertainty is attributed to prompt sensitivity. Our work introduces a new way to improve uncertainty calibration in prompt-sensitive language models, and provides evidence that some LLMs fail to exhibit consistent general reasoning about the meanings of their inputs.

Large Language Models (LLMs) demonstrate potential in the field of scientific idea generation. However, the generated results often lack controllable academic context and traceable inspiration pathways. To bridge this gap, this paper proposes a scientific idea generation system called GYWI, which combines author knowledge graphs with retrieval-augmented generation (RAG) to form an external knowledge base to provide controllable context and trace of inspiration path for LLMs to generate new scientific ideas. We first propose an author-centered knowledge graph construction method and inspiration source sampling algorithms to construct external knowledge base. Then, we propose a hybrid retrieval mechanism that is composed of both RAG and GraphRAG to retrieve content with both depth and breadth knowledge. It forms a hybrid context. Thirdly, we propose a Prompt optimization strategy incorporating reinforcement learning principles to automatically guide LLMs optimizing the results based on the hybrid context. To evaluate the proposed approaches, we constructed an evaluation dataset based on arXiv (2018-2023). This paper also develops a comprehensive evaluation method including empirical automatic assessment in multiple-choice question task, LLM-based scoring, human evaluation, and semantic space visualization analysis. The generated ideas are evaluated from the following five dimensions: novelty, feasibility, clarity, relevance, and significance. We conducted experiments on different LLMs including GPT-4o, DeepSeek-V3, Qwen3-8B, and Gemini 2.5. Experimental results show that GYWI significantly outperforms mainstream LLMs in multiple metrics such as novelty, reliability, and relevance.

Microbial biosynthetic diversity holds immense potential for discovering natural products with therapeutic applications, yet a substantial quantity of natural products derived from uncultivated microorganisms remains uncharacterized. The intricate nature of biosynthetic enzymes poses a major challenge in accurately predicting the chemical structures of secondary metabolites solely based on genome sequences using current rule-based methods. Here, we present DeepSeMS, a large language model designed to predict the chemical structures of secondary metabolites from various microbial biosynthetic gene clusters. Built on the Transformer architecture, DeepSeMS innovatively identifies sequence features using functional domains of biosynthetic enzymes, and incorporates feature-aligned chemical structure enumeration for training data augmentation. External evaluation results show that DeepSeMS predicts more accurate chemical structures of secondary metabolites with a Tanimoto coefficient up to 0.6 compared with the ground truth, significantly outperforming antiSMASH and PRISM with coefficients of only 0.14 and 0.45 respectively. Moreover, DeepSeMS successfully predicted secondary metabolites for 96.60% of cryptic biosynthetic gene clusters, surpassing existing methods with success rates less than 50%. Leveraging DeepSeMS, we characterized over 65,000 novel secondary metabolites from the global ocean microbiome with previously undocumented structural types, ecological distribution, and biomedical applications especially antibiotics. A login-free and user-friendly web server for DeepSeMS (https://biochemai.cstspace.cn/deepsems/) has been launched, featuring an integrated global ocean microbial secondary metabolites repository to expedite the discovery of novel natural products. Collectively, this study underscores the great capacity of a large language model-driven method in revealing hidden biosynthetic potential of the global ocean microbiome.

Large language models (LLMs) based on transformer architecture have shown outstanding performance across numerous real-world tasks. However, the autoregressive nature of these models makes the inference process slow and costly. Speculative decoding has emerged as a promising solution, leveraging a smaller auxiliary model to draft future tokens, which are then validated simultaneously by the larger model, achieving a speed-up of 1-2x. Although speculative decoding matches the same distribution as multinomial sampling, multinomial sampling itself is prone to suboptimal outputs, where as beam sampling is widely recognized for producing higher-quality results by maintaining multiple candidate sequences at each step. This paper explores the novel integration of speculative decoding with beam sampling. However, there are four key challenges: (1) how to generate multiple sequences from the larger model's distribution given drafts sequences from the small model; (2) how to dynamically optimize the number of beams to balance efficiency and accuracy; (3) how to efficiently verify the multiple drafts in parallel; and (4) how to address the extra memory costs inherent in beam sampling. To address these challenges, we propose dynamic-width speculative beam decoding (DSBD). Specifically, we first introduce a novel draft and verification scheme that generates multiple sequences following the large model's distribution based on beam sampling trajectories from the small model. Then, we introduce an adaptive mechanism to dynamically tune the number of beams based on the context, optimizing efficiency and effectiveness. Besides, we extend tree-based parallel verification to handle multiple trees simultaneously, accelerating the verification process. Finally, we illustrate a simple modification to our algorithm to mitigate the memory overhead of beam sampling. Experimental results show that our approach achieves a 1.5-1.9x speed-up and1.8-2.5x lower energy consumption compared to beam sampling, with no loss in downstream performance. Moreover, it can produce significantly higher-quality outputs than speculative decoding, while maintaining similar time, memory, and energy costs. In summary, our method offers a more efficient and effective inference process for LLMs.

Generative artificial intelligence (AI) has greatly advanced the code translation process, particularly through large language models (LLMs), which translate source code from one programming language into another. This translation has historically been error-prone, labor-intensive, and highly dependent on manual intervention. Although traditional tools such as compilers and transpilers have restrictions in managing complex programming paradigms, recent generative AI models most importantly those based on transformer architectures, have shown promise. This systematic mapping study intends to evaluate and compile studies on Generative AI applications in code translation released between 2020 and 2025. Using five main criteria, the publication year and channel, research type, publication type, empirical study type, and AI models. A total of 53 relevant articles were chosen and examined. The results show that conferences and journals are the most often used publishing venues. Although historical-based evaluations and case studies were the empirical methodologies most often used, researchers have primarily focused on implementing transformer-based artificial intelligence models.

The crystallization of modeling methods around the Transformer architecture has been a boon for practitioners. Simple, well-motivated architectural variations that transfer across tasks and scale, increasing the impact of modeling research. However, with the emergence of state-of-the-art 100B+ parameters models, large language models are increasingly expensive to accurately design and train. Notably, it can be difficult to evaluate how modeling decisions may impact emergent capabilities, given that these capabilities arise mainly from sheer scale.Targeting a multilingual language model in the 100B+ parameters scale, our goal is to identify an architecture and training setup that makes the best use of our 1,000,000 A100-GPU-hours budget. Specifically, we perform an ablation study comparing different modeling practices and their impact on zero-shot generalization. We perform all our experiments on 1.3B models, providing a compromise between compute costs and the likelihood that our conclusions will hold for the target 100B+ model. In addition, we study the impact of various popular pretraining corpora on zero-shot generalization. We also study the performance of a multilingual model and how it compares to the English-only one. Finally, we consider the scaling behaviour of Transformers to chose the target model size, shape, and training setup.