MoE-LLaVA : Mixture of Experts for Large Vision-Language Models

Mixture of Experts (MoE) models have enabled the scaling of Large Language Models (LLMs) and Vision Language Models (VLMs) by achieving massive parameter counts while maintaining computational efficiency. However, MoEs introduce several inference-time challenges, including load imbalance across experts and the additional routing computational overhead. To address these challenges and fully harness the benefits of MoE, a systematic evaluation of hardware acceleration techniques is essential. We present MoE-Inference-Bench, a comprehensive study to evaluate MoE performance across diverse scenarios. We analyze the impact of batch size, sequence length, and critical MoE hyperparameters such as FFN dimensions and number of experts on throughput. We evaluate several optimization techniques on Nvidia H100 GPUs, including pruning, Fused MoE operations, speculative decoding, quantization, and various parallelization strategies. Our evaluation includes MoEs from the Mixtral, DeepSeek, OLMoE and Qwen families. The results reveal performance differences across configurations and provide insights for the efficient deployment of MoEs.

While large language models (LLMs) are endowed with broad knowledge, their task-specific performance is often suboptimal. Fine-tuning LLMs with task-specific data from diverse nodes is necessary, but this data is typically safeguarded and not shared publicly due to privacy concerns. A common solution involves downstream nodes downloading the LLM locally and fine-tuning it with their proprietary data. However, owners often regard pre-trained LLMs as valuable assets and are reluctant to share them. Additionally, the significant computational resources required by LLMs make local fine-tuning impractical for many nodes. To mitigate these problems, this paper proposes CrossLM, a data-free collaborative fine-tuning framework for large and small language models. CrossLM enables resource-constrained nodes to train smaller language models (SLMs) using their private task-specific data. These SLMs are subsequently leveraged to promote the task-specific natural language generation and understanding capabilities of the LLMs. Simultaneously, the SLMs of nodes also benefit from enhancement by the fine-tuned LLMs. In this way, CrossLM avoids sharing private data and proprietary LLMs, and also reduces the resource requirements of nodes. Through extensive experiments across a range of benchmark tasks and popular language models, we demonstrate that CrossLM significantly boosts the task-specific performance of both LLMs and SLMs while preserving the generalization capabilities of LLMs.

How Can <i>IJDS</i> Authors, Reviewers, and Editors Use (and Misuse) Generative AI?

Interactive multimodal applications (IMAs), such as route planning in the Internet of Vehicles, enrich users’ personalized experiences by integrating various forms of data over wireless networks. Recent advances in large language models (LLMs) utilize mixture-of-experts (MoE) mechanisms to empower multiple IMAs, with each LLM trained individually for a specific task that presents different business workflows. In contrast to existing approaches that rely on multiple LLMs for IMAs, this paper presents a novel paradigm that accomplishes various IMAs using a single compositional LLM over wireless networks. The two primary challenges include 1) guiding a single LLM to adapt to diverse IMA objectives and 2) ensuring the flexibility and efficiency of the LLM in resource-constrained mobile environments. To tackle the first challenge, we propose ContextLoRA, a novel method that guides an LLM to learn the rich structured context among IMAs by constructing a task dependency graph. We partition the learnable parameter matrix of neural layers for each IMA to facilitate LLM composition. Then, we develop a step-by-step fine-tuning procedure guided by task relations, including training, freezing, and masking phases. This allows the LLM to learn to reason among tasks for better adaptation, capturing the latent dependencies between tasks. For the second challenge, we introduce ContextGear, a scheduling strategy to optimize the training procedure of ContextLoRA, aiming to minimize computational and communication costs through a strategic grouping mechanism. Experiments on three benchmarks show the superiority of the proposed ContextLoRA and ContextGear. Furthermore, we prototype our proposed paradigm on a real-world wireless testbed, demonstrating its practical applicability for various IMAs. We will release our code to the community.

Incorporating explicit personas into dialogue models is critical for generating responses that fulfill specific user needs and preferences, creating a more personalized and engaging interaction. Early works on persona-based dialogue generation directly concatenate the persona descriptions and dialogue history into relatively small pre-trained language models (PLMs) for response generation, which leads to uninformative and inferior results due to the sparse persona information and the limited model generation capabilities. Recently, large language models (LLMs) have shown their surprising capabilities in language generation. Prompting the LLMs with the persona descriptions for role-playing dialogue generation has also achieved promising results. However, deploying LLMs is challenging for practical applications due to their large scale, spurring efforts to distill the generation capabilities into more concise and compact models through teacher-student learning. In this article, we propose an efficient compact K nowledge-grounded P ersona-based D ialogue model enhanced by LLM D istillation (KPDD). Specifically, first, we propose to enrich the annotated persona descriptions by integrating external knowledge graphs (KGs) with a mixed encoding network, coupled with a mixture of experts (MoE) module for both informative and diverse response generation. The mixed encoding network contains multiple layers of modality interaction operations, enabling information from both modalities propagates to the other. Second, to fully exploit the generation capabilities of LLMs, we turn to the distillation technique to improve the generation capabilities of our model, facilitated by a natural language inference (NLI)-based filtering mechanism to extract high-quality information from LLMs. In addition, we employ a curriculum learning strategy to train our model on the high-quality filtered distilled data and progressively on the relatively noisy original data, enhancing its adaptability and performance. Extensive experiments show that KPDD outperforms state-of-the-art baselines in terms of both automatic and human evaluation.

Large language models’ exceptional all-purpose abilities have made human-computer conversations normal, but for particular industries and verticals, they fall short of enhancing the expertise of knowledge and the timeliness of information. In order to give current information, and provide improved search capabilities, large language models need to increasingly incorporate specialist resources and databases. In this research, a model for intelligent assisted decision-making was proposed that the model incorporates knowledge from domain-specific databases and real-time data and uses large language models to offer expert tax guidance. The research proposed to overcome the limits of general-purpose language models and deliver specialized advise for tax-related inquiries by complementing large language models with domain-specific information.The results we achieve demonstrate that by offering tax advice tailored to a given situation, and the model we proposed goes beyond the validity of general large language language models. Our contribution is that not only exploring the combination of tax area and large language model, but also proposing a new effective model for government tax department to use in real life. This study highlights the potential of big language models for use in real-world professional domains and advances the field of domain-specific human-computer interaction.

Parameter-efficient fine-tuning (PEFT) is crucial for customizing Large Language Models (LLMs) with constrained resources.Although there have been various PEFT methods for dense-architecture LLMs, PEFT for sparsearchitecture LLMs is still underexplored.In this work, we study the PEFT method for LLMs with the Mixture-of-Experts (MoE) architecture and the contents of this work are mainly threefold: (1) We investigate the dispersion degree of the activated experts in customized tasks, and found that the routing distribution for a specific task tends to be highly concentrated, while the distribution of activated experts varies significantly across different tasks.(2) We propose Expert-Specialized Fine-Tuning, or ESFT, which tunes the experts most relevant to downstream tasks while freezing the other experts and modules; experimental results demonstrate that our method not only improves the tuning efficiency, but also matches or even surpasses the performance of fullparameter fine-tuning.(3) We further analyze the impact of the MoE architecture on expertspecialized fine-tuning.We find that MoE models with finer-grained experts are more advantageous in selecting the combination of experts that are most relevant to downstream tasks, thereby enhancing both the training efficiency and effectiveness.Our code is available at https://github.com/deepseek-ai/ESFT.

Mixture-of-Experts (MoE) large language models (LLM) have memory requirements that often exceed the GPU memory capacity, requiring costly parameter movement from secondary memories to the GPU for expert computation. In this work, we present Mixture of Near-Data Experts (MoNDE), a near-data computing solution that efficiently enables MoE LLM inference. MoNDE reduces the volume of MoE parameter movement by transferring only the hot experts to the GPU, while computing the remaining cold experts inside the host memory device. By replacing the transfers of massive expert parameters with the ones of small activations, MoNDE enables far more communication-efficient MoE inference, thereby resulting in substantial speedups over the existing parameter of-floading frameworks for both encoder and decoder operations.

Large language models (LLMs), with their remarkable generative capacities, have greatly impacted a range of fields, but they face scalability challenges due to their large parameter counts, which result in high costs for training and inference. The trend of increasing model sizes is exacerbating these challenges, particularly in terms of memory footprint, latency and energy consumption. Here we explore the deployment of 'mixture of experts' (MoEs) networks-networks that use conditional computing to keep computational demands low despite having many parameters-on three-dimensional (3D) non-volatile memory (NVM)-based analog in-memory computing (AIMC) hardware. When combined with the MoE architecture, this hardware, utilizing stacked NVM devices arranged in a crossbar array, offers a solution to the parameter-fetching bottleneck typical in traditional models deployed on conventional von-Neumann-based architectures. By simulating the deployment of MoEs on an abstract 3D AIMC system, we demonstrate that, due to their conditional compute mechanism, MoEs are inherently better suited to this hardware than conventional, dense model architectures. Our findings suggest that MoEs, in conjunction with emerging 3D NVM-based AIMC, can substantially reduce the inference costs of state-of-the-art LLMs, making them more accessible and energy-efficient.

Aim/Purpose: The study investigates the factors influencing the acceptance and utilisation of large language models (LLMs) (predictor variables of LLM usage), such as ChatGPT, in Learning design by instructional designers and university-teaching academics from various countries. Background: Large language models (LLMs) have exploded onto the scene, transforming the landscape of learning design. Instructional designers and university teaching academics have been overburdened with content creation for their teaching programmes, and the arrival of LLM models will help in this regard by developing more interactive content that drives student engagement and, in turn, contributes to student success. Since LLMs are a relatively new phenomenon, little is known about the factors influencing their acceptance in learning design; therefore, this research is needed, as learning design principles are the bedrock of student engagement and success. Methodology: A cross-sectional correlational quantitative study was employed. Data was collected using an online questionnaire posted on social media, including LinkedIn, from 203 instructional designers and university teaching academics. Purposive and snowball sampling methods were used to target instructional designers and university teaching academics at colleges and universities worldwide. Participants were asked to share the survey link with fellow instructional designers and university-teaching academics in their communities. The factor structure of the data was determined using exploratory factor analysis. Nonetheless, the factor structure derived from the LLMs did not entirely reflect the original configuration of the Unified Theory of Acceptance and Use of Technology (UTAUT3), as certain predictors appeared to coalesce, indicating LLMs’ unique nature in learning design. Confirmatory factor analysis was used to verify the fit of the data on the measurement model. First-order and second-order structural modelling were used to identify the structural relationships among the variables. Contribution: The study determines significant factors for the acceptance of LLMs by instructional designers and academic teaching staff in learning design, enabling possible opportunities for best practices in the field through interventions to optimize LLM usage. The study applies the technology acceptance model to the emerging LLM technology and extends the technology acceptance model by adding the trust construct as a predictor variable. Findings: The structural analysis results indicated that the ingrained LLM practices, LLM peer-driven expectations, innovative propensity towards LLM adoption, reliability and provider trust in LLMs, and ease of use and support influenced perceived LLM benefits and usage, but community standards and infrastructure had no influence. The second-order structural equation modelling indicated that perceived LLM benefits and usage and ingrained LLM habits contributed most to the learning design. Recommendations for Practitioners: Teaching academics and instructional designers must use LLMs in designing content, assessments, and interactive learning activities, and attend LLM training workshops on prompting and best practices in integrating LLMs into learning and teaching to see their benefits; hence, regular use of LLMs will then lead to trust and innovation in LLMs usage, enhancing learning design and improving student learning outcomes. Recommendation for Researchers: Researchers must use mixed methods approaches to have a deeper understanding of the factors influencing LLMs. Since habit and perceived LLM benefits and usage contributed the most variance to learning design, researchers must investigate strategies that optimise these factors in learning design, such as effective intervention strategies that can help form positive LLM habits. In addition, the findings provide researchers with a starting point for future research. Further researchers must investigate interventions that optimise the influence of personal innovativeness and trust that contributed the least variance to learning design, hence unlocking the potential of LLMs in learning design through innovation, responsible, and ethical use. Impact on Society: The use of LLMs in learning design has a high possibility of transforming education, specifically the learning design landscape. Using LLMs will free up more time for teaching academics and instructional designers so that they spend more time on higher-order thinking skill demands. Consequently, the students will be exposed to more engaging and interactive content, resulting in improved learning outcomes. Future Research: Future research must include context-derived external variables in technology acceptance models, such as levels of prompting competencies, to provide a deeper understanding of LLMs. In addition, future research must be based on the application and impact of LLMs on student engagement and success, and their attainment of 21st-century skills.

Recent advancements have shown that the Mixture of Experts (MoE) approach significantly enhances the capacity of large language models (LLMs) and improves performance on downstream tasks. Building on these promising results, multi-modal large language models (MLLMs) have increasingly adopted MoE techniques. However, existing multi-modal MoE tuning methods typically face two key challenges: expert uniformity and router rigidity. Expert uniformity occurs because MoE experts are often initialized by simply replicating the FFN parameters from LLMs, leading to homogenized expert functions and weakening the intended diversification of the MoE architecture. Meanwhile, router rigidity stems from the prevalent use of static linear routers for expert selection, which fail to distinguish between visual and textual tokens, resulting in similar expert distributions for image and text. To address these limitations, we propose EvoMoE, an innovative MoE tuning framework. EvoMoE introduces a meticulously designed expert initialization strategy that progressively evolves multiple robust experts from a single trainable expert, a process termed expert evolution that specifically targets severe expert homogenization. Furthermore, we introduce the Dynamic Token-aware Router (DTR), a novel routing mechanism that allocates input tokens to appropriate experts based on their modality and intrinsic token values. This dynamic routing is facilitated by hypernetworks, which dynamically generate routing weights tailored for each individual token. Extensive experiments demonstrate that EvoMoE significantly outperforms other sparse MLLMs across a variety of multi-modal benchmarks, including MME, MMBench, TextVQA, and POPE. Our results highlight the effectiveness of EvoMoE in enhancing the performance of MLLMs by addressing the critical issues of expert uniformity and router rigidity.

Large language models (LLMs) are transforming healthcare by enabling the analysis of clinical texts, supporting diagnostics, and facilitating decision-making. This systematic review examines the evolution of LLMs from recurrent neural networks (RNNs) to transformer-based and multimodal architectures (e.g., BioBERT, MedPaLM), with a focus on their application in medical practice and challenges in Russia. Based on 40 peer-reviewed articles from Scopus, PubMed, and other reliable sources (2019-2025), LLMs demonstrate high performance (e.g., Med-PaLM: F1-score 0.88 for binary pneumonia classification on MIMIC-CXR; Flamingo-CXR: 77.7% preference for in/outpatient X-ray re-ports). However, limitations include data scarcity, interpretability challenges, and privacy concerns. An adaptation of the Mixture of Experts (MoE) architecture for rare disease diagnostics and automated radiology report generation achieved promising results on synthetic datasets. Challenges in Russia include limited annotated data and compliance with Federal Law No. 152-FZ. LLMs enhance clinical workflows by automating routine tasks, such as report generation and patient triage, with advanced models like KARGEN improving radiology report quality. Russia’s focus on AI-driven healthcare aligns with global trends, yet linguistic and infrastructural barriers necessitate tailored solutions. Developing robust validation frameworks for LLMs will ensure their reliability in diverse clinical scenarios. Collaborative efforts with international AI research communities could accelerate Russia’s adoption of advanced medical AI technologies, particularly in radiology automation. Prospects involve integrating LLMs with healthcare systems and developing specialized models for Russian medical contexts. This study provides a foundation for advancing AI-driven healthcare in Russia.

Large language models (LLMs) and large visual-language models (LVLMs) have exhibited near-human levels of knowledge, image comprehension, and reasoning abilities, and their performance has undergone evaluation in some healthcare domains. However, a systematic evaluation of their capabilities in cervical cytology screening has yet to be conducted. Here, we constructed CCBench, a benchmark dataset dedicated to the evaluation of LLMs and LVLMs in cervical cytology screening, and developed a GPT-based semi-automatic evaluation pipeline to assess the performance of six LLMs (GPT-4, Bard, Claude-2.0, LLaMa-2, Qwen-Max, and ERNIE-Bot-4.0) and five LVLMs (GPT-4V, Gemini, LLaVA, Qwen-VL, and ViLT) on this dataset. CCBench comprises 773 question-answer (QA) pairs and 420 visual-question-answer (VQA) triplets, making it the first dataset in cervical cytology to include both QA and VQA data. We found that LLMs and LVLMs demonstrate promising accuracy and specialization in cervical cytology screening. GPT-4 achieved the best performance on the QA dataset, with an accuracy of 70.5% for close-ended questions and average expert evaluation score of 6.9/10 for open-ended questions. On the VQA dataset, Gemini achieved the highest accuracy for close-ended questions at 67.8%, while GPT-4V attained the highest expert evaluation score of 6.1/10 for open-ended questions. Besides, LLMs and LVLMs revealed varying abilities in answering questions across different topics and difficulty levels. However, their performance remains inferior to the expertise exhibited by cytopathology professionals, and the risk of generating misinformation could lead to potential harm. Therefore, substantial improvements are required before these models can be reliably deployed in clinical practice.

Purpose This study aims to explore the integration of large language models (LLMs) and vision-language models (VLMs) in robotics, highlighting their potential benefits and the safety challenges they introduce, including robustness issues, adversarial vulnerabilities, privacy concerns and ethical implications. Design/methodology/approach This survey conducts a comprehensive analysis of the safety risks associated with LLM- and VLM-powered robotic systems. The authors review existing literature, analyze key challenges, evaluate current mitigation strategies and propose future research directions. Findings The study identifies that ensuring the safety of LLM-/VLM-driven robots requires a multi-faceted approach. While current mitigation strategies address certain risks, gaps remain in real-time monitoring, adversarial robustness and ethical safeguards. Originality/value This study offers a structured and comprehensive overview of the safety challenges in LLM-/VLM-driven robotics. It contributes to ongoing discussions by integrating technical, ethical and regulatory perspectives to guide future advancements in safe and responsible artificial intelligence-driven robotics.

To evaluate the accuracy and reasoning capabilities of large multimodal language models compared with those of neuroradiology subspecialty-trained radiologists in neuroradiology case interpretation. This experimental study used custom-made 401 radiologic quizzes derived from articles published in RadioGraphics covering neuroradiology and head and neck topics (October 2020 to February 2024). We prompted the GPT-4 Turbo with Vision (GPT-4V), GPT-4 Omni, Gemini Flash, and Claude models to provide the top three differential diagnoses with a rationale and describe examination characteristics such as imaging modality, sequence, use of contrast, image plane, and body part. The temperature was adjusted to 0 and 1 (T1). Two neuroradiologists answered the same questions. The accuracies of the large language models (LLMs) and the neuroradiologists were compared using generalized estimating equations. Three neuroradiologists assessed the rationale provided by the LLMs for their differential diagnoses using four-point scales, separately for specific lesion locations and imaging findings, and evaluated the presence of hallucinations and the overall acceptability of the responses. Top-3 accuracy (i.e., correct answers present among top-3 differential diagnoses) of LLMs ranged from 29.9% (120 of 401) to 49.4% (198 of 401, obtained with GPT-4V in the T1 setting), while radiologists achieved 80.3% (322 of 401) and 68.3% (274 of 401), respectively (P < 0.001). Regarding the rationale for differential diagnoses, GPT-4V (T1) accurately identified both the specific lesion location and imaging findings in 30.7% (123 of 401) and 12.9% (16 of 124) of cases without textual clinical history. Hallucinations occurred in 4.5% (18 of 401), and only 29.4% (118 of 401) of the LLM-generated analyses were deemed acceptable. GPT-4V (T1) demonstrated high accuracy in identifying the imaging modality (97.4% [800 of 821]) and scanned body parts (92.2% [756 of 820]). LLMs remarkably underperformed compared with neuroradiologists and showed unsatisfactory reasoning for their differential diagnoses, with performance declining further in cases without textual input of clinical history. These findings highlight the limitations of current multimodal LLMs in neuroradiological interpretation and their reliance on text input.