Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Research on Low-Latency Inference and Training Efficiency Optimization for Graph Neural Network and Large Language Model-Based Recommendation Systems

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that hybrid GNN-LLM recommender systems, optimized with FPGA acceleration, DeepSpeed, and LoRA fine-tuning, can beat standalone GNN and LLM models on accuracy while meeting real-time latency budgets.

desk verdict An unverifiable engineering report whose central accuracy claim is conceptually implausible as stated. read the letter →

arxiv 2507.01035 v1 pith:KGENO7XM submitted 2025-06-21 cs.LG cs.AIcs.PF

Yushang Zhao , Haotian Lyu , Yike Peng , Aijia Sun , Feng Jiang , Xinyue Han This is my paper · ORCID
classification cs.LGcs.AIcs.PF
keywords GraphNeuralNetworksLargeLanguageModelsRecommendationSystemsInferenceLatencyTrainingEfficiencyLoRAFPGAAccelerationDeepSpeed
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper seeks to show that hybrid recommender systems, which combine a graph neural network (GNN) for user-item structure with a large language model (LLM) for textual semantics, can be optimized to run within real-time latency budgets without giving up accuracy. It reports that adding FPGA hardware acceleration and DeepSpeed to the hybrid model lowers inference latency from 140 ms to about 45 ms, while Low-Rank Adaptation (LoRA) reduces training time from 11.3 hours to 3.8 hours. The best configuration, Hybrid + FPGA + DeepSpeed, reaches NDCG@10 of 0.75, a 13.6% improvement over the unoptimized hybrid baseline and higher than GNN-only or LLM-only models. If these numbers hold, the paper offers a concrete recipe for making LLM-powered recommenders fast enough for production systems.

What carries the argument

The mechanism that carries the argument is the fusion of a GNN encoder over the user-item graph with an LLM encoder over textual data, concatenated into a joint representation that an MLP prediction head scores. The optimizations are: INT8 quantization to shrink weights, knowledge distillation to transfer learning from a larger teacher, LoRA low-rank adapters for parameter-efficient fine-tuning, FPGA acceleration for graph operations such as neighbor sampling and aggregation, and DeepSpeed pipeline parallelism for multi-GPU training. The reported gains depend on the combination: FPGA plus DeepSpeed delivers the latency reduction, LoRA delivers the training-time reduction, and the hybrid architecture delivers the accuracy advantage.

What would settle it

Re-run the same configurations (GNN-only, LLM-only, hybrid, hybrid+LoRA, hybrid+FPGA+DeepSpeed) on a single fixed train/test split of MovieLens 1M or Amazon Books, with identical model sizes and a specified GPU/FPGA stack; if the optimized hybrid fails to beat the GNN-only baseline on NDCG@10, or its latency exceeds 100 ms, the central claim is refuted.

Watch

Extended reading notes

Core claim

The central claim is that a GNN-LLM hybrid recommender, optimized with quantization, knowledge distillation, LoRA, and FPGA/DeepSpeed acceleration, outperforms both standalone GNN and standalone LLM recommenders in accuracy while meeting real-time latency constraints. The paper reports that Hybrid + FPGA + DeepSpeed attains NDCG@10 of 0.75 and Precision@10 of 0.80 at 40–60 ms latency, a 13.6% accuracy gain over the unoptimized hybrid baseline (NDCG@10 of 0.66). LoRA alone cuts training time by 66%, from 11.3 to 3.8 hours. The author interprets these results as evidence that hardware-software co-design and parameter-efficient fine-tuning are the crucial levers that make hybrid architectures practically deployable.

Load-bearing premise

All reported improvements depend on an unstated, consistent experimental protocol—identical data splits, preprocessing, model sizes, and latency measurement conditions across configurations—so the 13.6% accuracy gain and 45 ms latency are only meaningful if those conditions were held fixed.

Editorial extensions

If this is right

  • Real-time recommender deployments can include LLM semantic signals without falling back to lightweight GNN-only models.
  • LoRA-style fine-tuning makes periodic model updates feasible, with training under four hours on this setup.
  • FPGA and DeepSpeed co-design pushes the latency-accuracy Pareto frontier toward lower latency and higher accuracy.
  • The 13.6% NDCG gain over the unoptimized baseline quantifies the value of a fully optimized hybrid pipeline.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The reported numbers come from a single, under-specified experimental protocol; an independent replication with fixed splits, model sizes, and hardware details is needed before treating the 13.6% gain and 45 ms latency as reliable.
  • If the speed-ups generalize, the same co-design strategy (parameter-efficient fine-tuning plus hardware acceleration) could apply to other hybrid LLM-and-graph systems, such as graph-based retrieval for question answering.
  • The accuracy-latency Pareto improvement suggests that 50 ms latency budgets are within reach for semantic recommender systems, which would enable LLM-based personalization in interactive products.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes a hybrid GNN+LLM recommendation architecture and reports experiments combining architectural optimization strategies (quantization, knowledge distillation, LoRA) with hardware acceleration (FPGA, DeepSpeed). The headline results are a 13.6% NDCG@10 improvement (from 0.66 to 0.75) at 40-60 ms inference latency for the 'Hybrid + FPGA + DeepSpeed' configuration and a 66% training-time reduction (from 11.3 or 11.4 hours to 3.8 hours) with LoRA, using MovieLens 1M, Amazon Books, and Yelp datasets. The manuscript contains no code, no data, no experimental protocol, and no statistical analysis; the reported tables contain multiple internal inconsistencies.

Significance. If the reported numbers were substantiated, the result would be practically significant: it would show that a fused GNN-LLM recommender can beat both standalone GNN and standalone LLM baselines while meeting real-time latency budgets, and that parameter-efficient fine-tuning plus FPGA acceleration can close the training/inference cost gap. The paper's conceptual taxonomy of optimization axes (quantization, distillation, LoRA, FPGA, DeepSpeed) is sensible, and the emphasis on hardware-software co-design is timely. However, the manuscript provides no reproducible artifacts, no machine-checked proofs, no parameter-free derivations, and no falsifiable prediction beyond post-hoc reported measurements. The central quantitative claim is unsupported as written, and the internal inconsistencies in the tables prevent verification. The significance is therefore potential rather than demonstrated.

major comments (3)
  1. [Table 4 and §V] The claimed 13.6% NDCG@10 improvement (0.66 to 0.75) for 'Hybrid + FPGA + DeepSpeed' cannot be caused by the listed optimizations. FPGA acceleration and DeepSpeed change the compute backend and parallelization, not the model weights, data, or evaluation; if the same trained model were measured across backends, NDCG should be unchanged up to numerical precision. The monotonic accuracy progression across rows (0.66, 0.70, 0.71, 0.72, 0.75) therefore implies each row is a separately fitted model or a different evaluation protocol. No training details, seeds, data splits, or error bars are reported, so the headline accuracy gain is unverifiable and internally implausible as stated.
  2. [§V, Table 4, §VI, Table 2, §IV] The manuscript contains mutually inconsistent numbers that indicate no fixed experimental protocol: the abstract and §VI state NDCG@10=0.75 while §V states the highest score is 0.74 (Fig. 6); §VI reports training time reduced from 11.4 to 3.8 hours while Table 2 lists the unoptimized baseline as 11.3 hours; §IV reports 'Precision@10 (75–100%)' while Table 4 reports scalar Precision@10 values between 0.65 and 0.80. These inconsistencies make it impossible to know which numbers are authoritative and undermine the quantitative claims of the paper.
  3. [§III.D, §III.E, §IV] No experimental protocol is provided. The datasets in Table 1 are named but no train/validation/test split, preprocessing, model architecture details (GNN layers, LLM backbone, embedding dimension), training hyperparameters, batch size, inference stack (GPU model, FPGA board, batch size, framework version), or number of repeated runs are given. Without this information, Tables 2–4 cannot be interpreted as controlled comparisons; the 13.6% accuracy gain and the 40–60 ms latency figure are not reproducible and cannot be attributed to the named optimizations.
minor comments (6)
  1. [§III.A, Equations (1)-(4)] Equations (1)-(4) contain garbled or under-specified notation, such as the unreadable propagation rule in Eq. (1) and unexplained symbols in Eqs. (3)-(4); these should be rewritten with standard mathematical formatting and explicit definitions.
  2. [Abstract and §III.C] The environment is described as 'R 4.4.2', but DeepSpeed, LoRA, and typical GNN/LLM tooling are Python-based; please clarify the actual software stack used for the experiments.
  3. [§IV-V, Figures 2-7] Figures 2-7 are referenced in the text but are not present in the manuscript; please include the figures or remove the references.
  4. [Table 1] The header 'Form at' appears to be a typo, and the dataset links should be formatted consistently in a bibliography style.
  5. [References [1]-[36]] Many cited works are unrelated to the topic (e.g., temperature prediction, tunnel leakage, COVID-19 analysis) and several in-text citations (e.g., [4], [5], [6]) do not support the claims they are attached to; the related-work section needs to be re-grounded in the actual GNN/LLM recommendation literature.
  6. [Table 4] The final row of Table 4 is split across lines ('Hybrid + FPGA + 0.8 0.65 0.75 DeepSpeed') and should be merged for readability.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the headline results are reported measurements rather than derived predictions, and the paper's self-citations are contextual and non-load-bearing.

full rationale

The paper contains no derivation chain that could be circular: its headline claims (the 13.6% NDCG@10 improvement for Hybrid + FPGA + DeepSpeed, the 66% training-time reduction from LoRA, and the 40-60ms latency) are reported measurements in Tables 2-4, not predictions derived from a model or from fitted parameters. Equations (1)-(4) define standard GNN message passing, LLM encoding, feature concatenation, and an MLP prediction head, but none of these equations is used to derive the experimental numbers, and no parameter is fitted to a subset of data and then 'predicted' on a closely related quantity. The accuracy climb across the configuration rows, although implausible if only the hardware/software backend changed, is a protocol-validity concern (uncontrolled training/evaluation variation, missing splits, seeds, model sizes, and error bars), not a circular reduction, so it belongs under correctness risk rather than circularity. The paper does cite prior work by its own authors (e.g., references [16], [17], and [22] include author Haotian Lyu; reference [34] includes authors Yushang Zhao and Yike Peng), but these citations support only background context ('joint optimization remains an underexplored yet transformative direction, which this study aims to address[21-24]') and the Pareto-frontier discussion [31-34]; they are not load-bearing for the central empirical claims, which are the paper's own measurements. No uniqueness theorem, imported ansatz, or renaming step is present, and the paper is self-contained against its own baselines (the GNN-only and LLM-only rows in Tables 3-4). The internal inconsistencies (NDCG 0.75 vs 0.74 in Section V; training time 11.3 vs 11.4 hours; 'Precision 75-100%' in Section IV vs a scalar Precision@10 in Table 4) undermine verifiability but do not constitute circularity. Net finding: no significant circularity; the score of 1 reflects only the presence of minor, non-load-bearing self-citations.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper introduces no free parameters in its equations; its empirical claims depend on unstated data splits, model choices, and hardware settings. The architecture relies on standard GNN message passing and concatenation-based fusion without justification or ablation. No new theoretical entities are introduced.

assumptions (4)
  • domain assumption The GNN message-passing rule in Eq. (1) with neighbor normalization is an effective encoder for user-item structure.
    The paper does not compare GNN variants or justify this update rule; the architecture's performance claims depend on it.
  • domain assumption Concatenating GNN and LLM embeddings (Eq. 3) and feeding them to an MLP head (Eq. 4) is a sufficient fusion mechanism.
    No ablation of fusion methods is provided; the reported improvement assumes this design choice.
  • domain assumption The MovieLens, Amazon, and Yelp datasets can be transformed into the user-item graph and text-feature representation described, and evaluation is on a held-out split.
    Preprocessing is described only at a high level; the reported metrics depend on an unstated data split and text encoding.
  • ad hoc to paper The post hoc selected 'optimal' Hybrid + FPGA + DeepSpeed configuration is representative of the method's performance rather than the best random seed.
    The paper does not report variance across seeds, so the single best result is used as the headline claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Research on Low-Latency Inference and Training Efficiency Optimization for Graph Neural Network and Large Language Model-Based Recommendation Systems." pith.science (2026). https://pith.science/paper/KGENO7XM

@misc{pith2026250701035,
  author       = {Pith},
  title        = {Pith review of: Research on Low-Latency Inference and Training Efficiency Optimization for Graph Neural Network and Large Language Model-Based Recommendation Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KGENO7XM}},
  note         = {Machine review of arXiv:2507.01035}
}
read the original abstract

The incessant advent of online services demands high speed and efficient recommender systems (ReS) that can maintain real-time performance along with processing very complex user-item interactions. The present study, therefore, considers computational bottlenecks involved in hybrid Graph Neural Network (GNN) and Large Language Model (LLM)-based ReS with the aim optimizing their inference latency and training efficiency. An extensive methodology was used: hybrid GNN-LLM integrated architecture-optimization strategies(quantization, LoRA, distillation)-hardware acceleration (FPGA, DeepSpeed)-all under R 4.4.2. Experimental improvements were significant, with the optimal Hybrid + FPGA + DeepSpeed configuration reaching 13.6% more accuracy (NDCG@10: 0.75) at 40-60ms of latency, while LoRA brought down training time by 66% (3.8 hours) in comparison to the non-optimized baseline. Irrespective of domain, such as accuracy or efficiency, it can be established that hardware-software co-design and parameter-efficient tuning permit hybrid models to outperform GNN or LLM approaches implemented independently. It recommends the use of FPGA as well as LoRA for real-time deployment. Future work should involve federated learning along with advanced fusion architectures for better scalability and privacy preservation. Thus, this research marks the fundamental groundwork concerning next-generation ReS balancing low-latency response with cutting-edge personalization.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. A Multimodal RAG Framework for Housing Damage Assessment: Collaborative Optimization of Image Encoding and Policy Vector Retrieval

    cs.CV 2025-09 conditional novelty 4.0 of 10

    A multimodal retrieval-augmented generation framework jointly encodes disaster images and insurance policies, reporting higher damage classification and retrieval accuracy than unimodal baselines on a self-constructed...

Reference graph

Works this paper leans on

36 extracted references · 10 canonical work pages · cited by 1 Pith paper

  1. [1]

    A Mixed-Heuristic Quantum-Inspired Simplified Swarm Optimization Algorithm for scheduling of real-time tasks in the multiprocessorsystem

    Su, Pei-Chiang, et al. "A Mixed-Heuristic Quantum-Inspired Simplified Swarm Optimization Algorithm for scheduling of real-time tasks in the multiprocessorsystem."AppliedSoftComputing131(2022):109807

    work page 2022

  2. [2]

    Real-TimePredictionforAthletes'Psychological States Using BERT-XGBoost: Enhancing Human-Computer Interaction

    Duan,Chenming,etal."Real-TimePredictionforAthletes'Psychological States Using BERT-XGBoost: Enhancing Human-Computer Interaction."arXivpreprintarXiv:2412.05816(2024)

    arXiv 2024

  3. [3]

    Accurate Prediction of Temperature Indicators in Eastern China Using a Multi-Scale CNN-LSTM-Attention model[J]

    Shen J, Wu W, Xu Q. Accurate Prediction of Temperature Indicators in Eastern China Using a Multi-Scale CNN-LSTM-Attention model[J]. arXivpreprintarXiv:2412.07997,2024

    arXiv 2024

  4. [4]

    II-NVM: Enhancing Map Accuracy and Consistency with Normal Vector-Assisted Mapping[J]

    Zhao C, Li Y, Jian Y, et al. II-NVM: Enhancing Map Accuracy and Consistency with Normal Vector-Assisted Mapping[J]. IEEE Robotics andAutomationLetters,2025

    2025

  5. [5]

    TanC,LiX,WangX,etal.Real-timeVideoTargetTrackingAlgorithm Utilizing Convolutional Neural Networks (CNN)[C]//2024 4th International Conference on Electronic Information Engineering and Computer(EIECT).IEEE,2024:847-851

    work page 2024

  6. [6]

    Construction and Analysis of Collaborative Educational Networks based on Student Concept Maps[J]

    Freedman H, Young N, Schaefer D, et al. Construction and Analysis of Collaborative Educational Networks based on Student Concept Maps[J]. Proceedings of the ACM on Human-Computer Interaction, 2024, 8(CSCW1):1-22

    work page 2024

  7. [7]

    AvailableatSSRN5179192

    WuC,HuangH,NiYQ,etal.EvaluationofTunnelRockMassIntegrity UsingMulti-ModalDataandGenerativeLargeModels:Tunnelrip-Gpt[J]. AvailableatSSRN5179192

    work page

  8. [8]

    Wu S,HuangX, Lu D.Psychologicalhealthknowledge-enhanced LLM- based social network crisis intervention text transfer recognition method[J].arXivpreprintarXiv:2504.07983,2025

    arXiv 2025

Show all 36 references
  1. [9]

    Optimized path planning for logistics robots using ant colony algorithm under multiple constraints[J]

    ZhaoH, Ma Z, Liu L, et al. Optimized path planning for logistics robots using ant colony algorithm under multiple constraints[J]. arXiv preprint arXiv:2504.05339,2025

    2025 arXiv

  2. [10]

    Xiang,A.,Zhang,J.,Yang,Q., Wang,L.,&Cheng,Y.(2024).Research on splicing image detection algorithms based on natural image statistical characteristics.arXivpreprintarXiv:2404.16296

    2024 arXiv

  3. [11]

    Ad Placement Optimization Algorithm Combined with MachineLearninginInternetE-Commerce[J].2025

    Feng H, Gao Y. Ad Placement Optimization Algorithm Combined with MachineLearninginInternetE-Commerce[J].2025

    2025

  4. [12]

    ZhangT,ZhangB,ZhaoF,etal.COVID-19localizationandrecognition on chest radiographs based on Yolov5 and EfficientNet[C]//2022 7th InternationalConferenceonIntelligentComputingandSignalProcessing (ICSP).IEEE,2022:1827-1830

    2022

  5. [13]

    Gao Z, Tian Y, Lin S C, et al. A ct image classification network framework for lung tumors based on pre-trained mobilenetv2 model and transfer learning, and its application and market analysis in the medical field[J].arXivpreprintarXiv:2501.04996,2025

    2025 arXiv

  6. [14]

    Multidimensional precipitation index prediction based on CNN-LSTM hybrid framework[J]

    Wang Y, Jia P, Shu Z, et al. Multidimensional precipitation index prediction based on CNN-LSTM hybrid framework[J]. arXiv preprint arXiv:2504.20442,2025

    2025 arXiv

  7. [15]

    CCi-YOLOv8n: Enhanced Fire Detection with CARAFE and Context-GuidedModules[J].arXivpreprintarXiv:2411.11011,2024

    Lv K. CCi-YOLOv8n: Enhanced Fire Detection with CARAFE and Context-GuidedModules[J].arXivpreprintarXiv:2411.11011,2024

    2024 arXiv

  8. [16]

    LLM-Driven E-Commerce Marketing Content Optimization: Balancing Creativity and Conversion[J]

    Yang H, Lyu H, Zhang T, et al. LLM-Driven E-Commerce Marketing Content Optimization: Balancing Creativity and Conversion[J]. arXiv preprintarXiv:2505.23809,2025

    2025 arXiv

  9. [17]

    Research on E-Commerce Long-Tail Product Recommendation Mechanism Based on Large-Scale Language Models[J].arXivpreprintarXiv:2506.06336,2025

    Lu Q, Lyu H, Zheng J, et al. Research on E-Commerce Long-Tail Product Recommendation Mechanism Based on Large-Scale Language Models[J].arXivpreprintarXiv:2506.06336,2025

    2025 arXiv

  10. [18]

    Avocado Price Prediction Using a Hybrid Deep Learning Model: TCN-MLP-Attention Architecture[J]

    Zhang L, Liang R. Avocado Price Prediction Using a Hybrid Deep Learning Model: TCN-MLP-Attention Architecture[J]. arXiv preprint arXiv:2505.09907,2025

    2025 arXiv

  11. [19]

    CTLformer: A Hybrid Denoising Model Combining Convolutional Layers and Self-Attention for Enhanced CT ImageReconstruction[J].arXivpreprintarXiv:2505.12203,2025

    Zheng Z, Wu S, Ding W. CTLformer: A Hybrid Denoising Model Combining Convolutional Layers and Self-Attention for Enhanced CT ImageReconstruction[J].arXivpreprintarXiv:2505.12203,2025

    2025 arXiv

  12. [20]

    Analysis of collective response reveals that covid-19-related activities start from the end of 2019 in mainland china[J].medRxiv,2020:2020.10.14.20202531

    Liu J, Huang T, Xiong H, et al. Analysis of collective response reveals that covid-19-related activities start from the end of 2019 in mainland china[J].medRxiv,2020:2020.10.14.20202531

    2019

  13. [21]

    Applications and Effect Evaluation of Generative Adversarial Networks in Semi-Supervised Learning[J]

    Hu J, Zeng H, Tian Z. Applications and Effect Evaluation of Generative Adversarial Networks in Semi-Supervised Learning[J]. arXiv preprint arXiv:2505.19522,2025

    2025 arXiv

  14. [22]

    User Behavior Analysis in Privacy ProtectionwithLargeLanguageModels:AStudyonPrivacyPreferences withLimitedData[J].arXivpreprintarXiv:2505.06305,2025

    Yang H, Lu Q, Wang Y, et al. User Behavior Analysis in Privacy ProtectionwithLargeLanguageModels:AStudyonPrivacyPreferences withLimitedData[J].arXivpreprintarXiv:2505.06305,2025

    2025 arXiv

  15. [23]

    Research on feature fusion and multimodal patent text based on graph attention network[J]

    Song Z, Liu Z, Li H. Research on feature fusion and multimodal patent text based on graph attention network[J]. arXiv preprint arXiv:2505.20188,2025

    2025 arXiv

  16. [24]

    FanP,LiuK, QiZ.MaterialFlowPredictionTaskBasedOnTCN-GRU DeepFusionModel[J].2025

    2025

  17. [25]

    Towards automated 3D evaluation of water leakage on a tunnel face via improved GAN and self-attention DL model[J]

    Wu C, Huang H, Zhang L, et al. Towards automated 3D evaluation of water leakage on a tunnel face via improved GAN and self-attention DL model[J]. Tunnelling and Underground Space Technology, 2023, 142: 105432

    2023

  18. [26]

    Application ofAI inReal-time Credit Risk Detection[J].2025

    WangZ, Zhang Q, ChengZ. Application ofAI inReal-time Credit Risk Detection[J].2025

    2025

  19. [27]

    arXiv preprint arXiv:2501.02167, 2025

    TanC,ZhangW, QiZ, et al.GeneratingMultimodalImages withGAN: Integrating Text, Image, and Style[J]. arXiv preprint arXiv:2501.02167, 2025

    2025 arXiv

  20. [28]

    Personalized Risks and Regulatory Strategies of Large Language Models in Digital Advertising[J]

    Feng H, Dai Y, Gao Y. Personalized Risks and Regulatory Strategies of Large Language Models in Digital Advertising[J]. arXiv preprint arXiv:2505.04665,2025

    2025 arXiv

  21. [29]

    Enhancing Code LLMs with Reinforcement Learning in Code Generation[J]

    Wang J, Zhang Z, He Y, et al. Enhancing Code LLMs with Reinforcement Learning in Code Generation[J]. arXiv preprint arXiv:2412.20367,2024

    2024 arXiv

  22. [30]

    Automated Parking Trajectory Generation Using Deep Reinforcement Learning[J]

    Zhang Z, Luo Y, Chen Y, et al. Automated Parking Trajectory Generation Using Deep Reinforcement Learning[J]. arXiv preprint arXiv:2504.21071,2025

    2025 arXiv

  23. [31]

    (2024, August)

    Xiang, A., Qi, Z., Wang, H., Yang, Q., & Ma, D. (2024, August). A multimodal fusion network for student emotion recognition based on transformer and tensor product. In 2024 IEEE 2nd International ConferenceonSensors,ElectronicsandComputerEngineering(ICSECE) (pp.1-4).IEEE

    2024

  24. [32]

    Research on the Design of a Short Video Recommendation System Based on Multimodal Information and DifferentialPrivacy[J].arXivpreprintarXiv:2504.08751,2025

    Yang H, Fu L, Lu Q, et al. Research on the Design of a Short Video Recommendation System Based on Multimodal Information and DifferentialPrivacy[J].arXivpreprintarXiv:2504.08751,2025

    2025 arXiv

  25. [33]

    Enhanced Recommendation Combining Collaborative Filtering and Large Language Models[J]

    Lin X, Cheng Z, Yun L, et al. Enhanced Recommendation Combining Collaborative Filtering and Large Language Models[J]. arXiv preprint arXiv:2412.18713,2024

    2024 arXiv

  26. [34]

    ZhaoY,PengY,LiD,etal.ResearchonPersonalizedFinancialProduct Recommendation by Integrating Large Language Models and Graph NeuralNetworks[J].arXivpreprintarXiv:2506.05873,2025

    2025 arXiv

  27. [35]

    Financial analysis: Intelligent financial data analysis system based on llm-rag[J]

    Wang J, Ding W, Zhu X. Financial analysis: Intelligent financial data analysis system based on llm-rag[J]. arXiv preprint arXiv:2504.06279, 2025

    2025 arXiv

  28. [36]

    Research on Multi-Modal Retrieval System of E-Commerce Platform Based on Pre-Training Model

    Zhang, Bingbing, Yi Han, and Xiaofei Han. "Research on Multi-Modal Retrieval System of E-Commerce Platform Based on Pre-Training Model."ArtificialIntelligenceTechnologyResearch2.9(2025)

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.