Pith. sign in

REVIEW 5 major objections 5 minor 5 cited by

Research on Personalized Financial Product Recommendation by Integrating Large Language Models and Graph Neural Networks

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Combining LLM text embeddings with graph-structured interactions improves financial product ranking, reaching NDCG@10 of 0.372 versus 0.301 for the best standalone baseline.

desk verdict The claimed 12.5% gain doesn't survive contact with the paper's own table, and the evaluation is not reproducible; the method is a standard hybrid with a minor twist. read the letter →

arxiv 2506.05873 v1 pith:REXAXORI submitted 2025-06-06 cs.IR cs.AI

Yushang Zhao , Yike Peng , Dannier Li , Yuxin Yang , Chengrui Zhou , Jing Dong This is my paper · ORCID
classification cs.IRcs.AI
keywords largelanguagemodelsgraphneuralnetworkspersonalizedrecommendationfinancialproductshybridmodelpseudo-labelmessagepassingcross-modalfusion
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 argues that financial product recommendation improves when a large language model's text embeddings and a graph neural network's relational structure are trained together rather than used separately. It proposes a three-stage hybrid framework that lets the GNN enrich LLM features, lets both modalities co-train, and lets LLM semantics guide GNN structure, with a dual-stream implementation that fuses text and graph features and adds pseudo-labels as auxiliary supervision. The reported payoff is a ranking model that reaches NDCG@10 of 0.372 on financial data, a 12.5% improvement over the best baseline, while staying close to the GNN-only model in training time and size. A sympathetic reader would take away that cross-modal fusion, not a bigger model, is the source of the gains.

What carries the argument

The carrying mechanism is a dual-stream feature pipeline with a shared prediction head. Stream one encodes each node's text with a pre-trained LLM ($e_i=LLM(t_i)$), turns that embedding into a pseudo-label through a linear layer and sigmoid ($\hat{y}_i=\sigma(W_p^\top e_i+b_p)$), concatenates the two, and refines the result through a multi-layer graph attention network whose attention coefficients $\alpha_{ij}$ weight neighbor messages. Stream two computes LLM text embeddings and graph-convolutional features separately and merges them with a learnable fusion operator $\phi(W_p h_i^{text}+W_g h_i^{graph}+b_f)$. The final node representation concatenates both streams and is trained with a joint loss $\mathcal{L}=\lambda_{rank}\mathcal{L}_{rank}+\lambda_{pseudo}\mathcal{L}_{pseudo}+\lambda_{reg}\|\Theta\|_2^2$, which lets the model adjust text-versus-graph contributions and use pseudo-labels as auxiliary supervision.

What would settle it

Run the hybrid model and the three baselines on a named public financial recommendation dataset under the paper's stated splits and hyperparameters, or release the private dataset, and check whether NDCG@10 reaches 0.372 with a 12.5% margin over the best baseline; materially different numbers would disprove the performance claim.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a dual-stream architecture—one stream concatenating LLM text embeddings with pseudo-labels and passing them through a graph attention network, the other stream fusing LLM text embeddings with graph-convolutional features through a learnable operator—jointly learned with a multi-objective loss outperforms each modality alone. On the test set the hybrid model beats collaborative filtering, a GNN-only model, and an LLM-only model on Hit Rate@10, Precision@10, Recall@10, NDCG@10, and MRR, with the headline NDCG@10 of 0.372 versus 0.301 for the GNN baseline. Ablations attribute the gain to both the text and graph streams and to the pseudo-label loss, and the authors report that removing any component also lowers the model's Gini-coefficient interpretability score.

Load-bearing premise

The load-bearing premise is that the experiments are run on real financial recommendation data with honest splits; the datasets are never named, described, or released, so the reported numbers cannot be checked for leakage or baseline fairness.

Editorial extensions

If this is right

  • On the reported test set, the hybrid model reaches NDCG@10 of 0.372, Hit Rate@10 of 0.578, Precision@10 of 0.071, Recall@10 of 0.289, and MRR of 0.193, beating collaborative filtering, GNN-only, and LLM-only baselines on every metric.
  • The text stream and the graph stream are both load-bearing: removing text drops NDCG@10 to 0.318 and MRR to 0.168, while removing graph drops NDCG@10 to 0.329 and MRR to 0.174.
  • The pseudo-label loss contributes a smaller but real margin: dropping it lowers NDCG@10 from 0.372 to 0.357 and MRR from 0.193 to 0.186, while speeding up training.
  • Accuracy gains do not come at the cost of efficiency: training time per epoch is 2.8s for the hybrid model, near the GNN's 2.5s and well below the LLM-only model's 8.7s, with 12.3M parameters.
  • Full-model interpretability, measured by a Gini coefficient, reaches 0.68 and falls to 0.54 when any component is removed, supporting the claim that fusion improves explainability.

Reading between the lines

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

  • The datasets behind Table 1 are never named or released, so the exact margin of 12.5% is not independently checkable; a reader should treat the comparative ranking as the claim and the specific numbers as unverified.
  • The ablation design removes whole streams, so it cannot separate the contribution of pseudo-labels from the contribution of text embeddings themselves; an LLM-only model with pseudo-labels would isolate that effect.
  • The proposed fusion operator and pseudo-label mechanism are generic enough to transfer to other cross-modal recommendation domains, such as job or real-estate matching, where text descriptions and interaction graphs coexist.
  • If the conclusion's suggested extension to dynamic graphs and online learning holds, the same text-plus-graph fusion could apply to fraud detection and credit-risk monitoring, where user text and transaction networks arrive continuously.
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

5 major / 5 minor

Summary. The paper proposes a hybrid recommendation framework that combines a pre-trained large language model (LLM) with a graph neural network (GNN) for personalized financial product recommendation. The architecture encodes user/product text via an LLM, builds a heterogeneous user-product graph, fuses the two modalities through parallel streams and a learnable fusion operator, and adds a pseudo-label auxiliary loss. The authors report experiments on 'public and real-world financial datasets' with Hit Rate@10, Precision@10, Recall@10, NDCG@10, and MRR, claiming that the hybrid model outperforms collaborative filtering, GNN-only, and LLM-only baselines, together with ablation and interpretability analyses.

Significance. If the empirical claims were fully supported, the paper would offer a modest incremental contribution: a plausible way to combine text semantics from LLMs with structural signals from GNNs for recommendation, with the pseudo-label mechanism as a distinctive design choice. However, as presented, the central superiority claim is not verifiable because the datasets are never named or described, the reported improvement is numerically inconsistent, and no uncertainty estimates are given. The paper also does not release code or provide baseline implementation details, so the experiments cannot be reproduced or independently checked. The potential significance of the architectural idea is therefore outweighed by the current lack of evidential support.

major comments (5)
  1. [§4.1] The central empirical claim rests on experiments performed on datasets described only as 'public and real-world financial datasets' (Section 4.1). No dataset names, sizes, interaction counts, user-item statistics, or train/validation/test split details are provided. Without this information, the reported numbers cannot be checked for split correctness, label leakage, or baseline fairness. This is a load-bearing omission because the paper's main contribution is an empirical superiority claim.
  2. [§4.2, Table 1, §5.1] The reported improvement is internally inconsistent. Section 4.2 states that the hybrid model achieves NDCG@10 of 0.372, 'a 12.5% improvement over the best baseline,' but the best baseline in Table 1 is the GNN at 0.301, which gives (0.372 - 0.301) / 0.301 = 23.6%. Against the LLM baseline (0.289) the improvement is 28.7%, and against CF (0.265) it is 40.4%. Section 5.1 separately reports '6.6%–23.6% gains over GNN,' which is also irreconcilable with the 12.5% figure. The claimed improvement cannot be reproduced from the paper's own table.
  3. [§4.2, §5.2] All experimental results are reported as single numbers without error bars, confidence intervals, or significance tests. The manuscript states that experiments were 'repeated three times and averaged' (Section 4.2), but no standard deviations are given for Table 1, Figure 6, or Figure 7. Consequently, the reported differences between the hybrid model and the baselines, and between the full model and the ablation variants, cannot be distinguished from random variation.
  4. [§3.2, Eq. (7)] The pseudo-label mechanism is not convincingly justified. In Eq. (7), the second loss term enforces consistency between pseudo-labels ŷ and ground-truth labels y, while the pseudo-labels themselves are produced from LLM embeddings of the same nodes (Eqs. (1)-(2)). Since the pseudo-label loss is supervised by the same ground-truth labels used in the main loss, it is unclear what independent signal the pseudo-labels provide. The ablation in §5.2 shows NDCG@10 of 0.357 without the pseudo-label loss versus 0.372 with it, but without variance or dataset details this difference is not meaningful support.
  5. [§5.1] The comparison against baselines is not sufficient to support the claim of state-of-the-art or even competitive performance. The baselines are simple collaborative filtering, a generic GNN, and an LLM-only model, with no implementation details, hyperparameter settings for baselines, or stronger reference methods such as LightGCN, NGCF, or recent LLM-based recommender systems. The statement that the hybrid model 'outperformed all baselines' is therefore limited to these elementary baselines and does not establish practical significance.
minor comments (5)
  1. [§3.2] The mathematical notation is frequently garbled: symbols such as 'wB', 'a!C', 'h!DEFD', 'h!BGH&I', 'h!JKL"MN', and 'λLK$' appear in place of proper subscripts and superscripts. The formulas need to be typeset correctly before the manuscript can be evaluated for technical soundness.
  2. [Table 2] The table caption contains a typographical inconsistency: 'Table 2：Key hyperparameter settings' uses a full-width colon, unlike other captions. This is a minor formatting issue.
  3. [§6] The conclusion cites Shen et al. [51] and Wang et al. [52] as sources of inspiration, but these references are about temperature prediction and network log anomaly detection, respectively. The connection to the proposed pseudo-label mechanism is not explained in the methodology, and the citations feel disconnected from the technical content.
  4. [References] The reference list contains many entries that appear unrelated to the topics of LLM-based recommendation, GNN-based recommendation, or financial product recommendation (e.g., warehouse robot path planning, CT image reconstruction, baseball teaching evaluation). The authors should either connect these references to the text or remove them, as the current list is not a reliable scholarly apparatus.
  5. [§4.2] The statement that hyperparameter tuning was performed 'within the same search space' is not backed by any details about the search space or the tuning procedure for the baselines. Reporting at least the range of hyperparameters and the selection criterion would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's central claim is an empirical comparison against external baselines, not a derivation that reduces to its own inputs.

full rationale

The central claim is an empirical superiority result (Section 4.2, Table 1) evaluated against three external baselines on a test set, and the hybrid model could in principle have lost, so the claim is not forced by construction. The architectural equations (Eqs. 1-7) define an embedding and training objective: e_i=LLM(t_i) is a text embedding, y-hat is a sigmoid projection of that embedding, h is a concatenation, the GAT update is standard message passing, and the joint loss combines a task loss, a pseudo-label loss, and regularization. The pseudo-label term is unusual because it is trained against ground-truth labels y rather than self-generated targets (Eq. 7), but that is a supervision choice, not a self-definitional or fitted-input-called-prediction circularity. No load-bearing argument relies on self-citations; references [51] and [52] are invoked only as loose inspiration in the conclusion, and the other citations are external prior work. The internal inconsistency in the reported 12.5% improvement (Table 1 implies a 23.6% gain over the best baseline GNN) and the unnamed datasets are correctness and reproducibility concerns, not circular reasoning. Therefore the circularity score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The model relies on pre-trained LLM embeddings and graph structure, both treated as given. The only paper-specific invention is the pseudo-label auxiliary loss, which is a modeling choice rather than a new entity. Six hyperparameters are tuned by grid search and are needed to reproduce the results.

free parameters (6)
  • learning_rate = 3 x 10^-4
    Selected by grid search (Table 2); affects optimization and final results.
  • embedding_dimension = 256
    Selected by grid search; dimensionality of text and graph feature vectors (Table 2).
  • num_gnn_layers = 3
    Selected by grid search; number of graph attention layers (Table 2).
  • batch_size = 128
    Selected by grid search; training batch size (Table 2).
  • dropout_rate = 0.2
    Selected by grid search; dropout in GNN and fusion layers (Table 2).
  • pseudo_label_loss_weight = 0.3
    Selected by grid search; weight of the pseudo-label loss in the total loss (Table 2, Eq. 7).
assumptions (4)
  • domain assumption Pre-trained LLM embeddings capture implicit user preferences relevant to financial product choices.
    Section 3.2, Eq. 1: the whole pipeline relies on LLM text features being informative for the recommendation task.
  • domain assumption User-product interactions and social ties form a heterogeneous graph whose structure is compatible with the text features.
    Section 3.1: the graph and text streams are assumed to be complementary and jointly optimizable.
  • ad hoc to paper Pseudo-labels generated from LLM embeddings are a useful training signal even when supervised by the same ground-truth labels.
    Section 3.2, Eq. 7: the second loss term enforces consistency between pseudo-labels and ground-truth y, so the pseudo-label branch is trained with task labels.
  • standard math Standard deep learning assumptions such as differentiability and i.i.d. sampling hold.
    Implicit in the use of gradient-based training and the reported loss functions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Research on Personalized Financial Product Recommendation by Integrating Large Language Models and Graph Neural Networks." pith.science (2026). https://pith.science/paper/REXAXORI

@misc{pith2026250605873,
  author       = {Pith},
  title        = {Pith review of: Research on Personalized Financial Product Recommendation by Integrating Large Language Models and Graph Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/REXAXORI}},
  note         = {Machine review of arXiv:2506.05873}
}
read the original abstract

With the rapid growth of fintech, personalized financial product recommendations have become increasingly important. Traditional methods like collaborative filtering or content-based models often fail to capture users' latent preferences and complex relationships. We propose a hybrid framework integrating large language models (LLMs) and graph neural networks (GNNs). A pre-trained LLM encodes text data (e.g., user reviews) into rich feature vectors, while a heterogeneous user-product graph models interactions and social ties. Through a tailored message-passing mechanism, text and graph information are fused within the GNN to jointly optimize embeddings. Experiments on public and real-world financial datasets show our model outperforms standalone LLM or GNN in accuracy, recall, and NDCG, with strong interpretability. This work offers new insights for personalized financial recommendations and cross-modal fusion in broader recommendation tasks.

Figures

Figures reproduced from arXiv: 2506.05873 by the authors.

Figure 4
Figure 4. Process of text feature extraction and graph structure fusion [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 6
Figure 6. Comprehensive comparison of recommendation accuracy and training [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Ablation study results and interpretability analysis [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 5 Pith papers

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

  1. Multimodal Foundation Model-Driven User Interest Modeling and Behavior Analysis on Short Video Platforms

    cs.IR 2025-09 reject novelty 3.0 of 10

    A standard attention-fusion plus Transformer sequence model is applied to short-video recommendation, with claimed gains over weak baselines and no reproducible artifacts.

  2. Meta-Learning for Cold-Start Personalization in Prompt-Tuned LLMs

    cs.LG 2025-07 reject novelty 3.0 of 10

    A meta-learned prompt-tuning method for cold-start LLM recommendations reports better Hit@10 and nDCG@10 on MovieLens-1M, but with no code, no error bars, and no shown results for Amazon or Recbole.

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

    cs.LG 2025-06 reject novelty 3.0 of 10

    A hybrid GNN-LLM recommender with FPGA, DeepSpeed, and LoRA reportedly reaches NDCG@10 of 0.75 at 40-60ms latency while cutting training time by 66%, but the supporting artifacts are absent.

  4. LLM-Augmented Symptom Analysis for Cardiovascular Disease Risk Prediction: A Clinical NLP

    cs.CL 2025-07 reject novelty 2.0 of 10

    A small synthetic study reports that Bio_ClinicalBERT embeddings with Random Forest classify CVD risk in about 20 hand-written symptom texts, but the claims of MIMIC-III and CARDIO-NLP evaluation are unsupported.

  5. Research on Model Parallelism and Data Parallelism Optimization Methods in Large Language Model-Based Recommendation Systems

    cs.DC 2025-06 reject novelty 2.0 of 10

    A hybrid model-plus-data parallel scheme is reported to boost training throughput and GPU utilization for LLM-based recommenders, but the supporting experiments are not reproducible from the paper.

Reference graph

Works this paper leans on

43 extracted references · 13 canonical work pages · cited by 5 Pith papers

  1. [1]

    C., Tan, S

    Su, P. C., Tan, S. Y., Liu, Z., & Yeh, W. C. A mixed-heuristic quantum-inspired simplified swarm optimization algorithm for scheduling of real-time tasks in the multiprocessor system. Applied Soft Computing, 2022, 131: 109807

    work page 2022

  2. [3]

    An ensemble approach to stock price prediction using deep learning and time series models

    Sui M, Zhang C, Zhou L, et al. An ensemble approach to stock price prediction using deep learning and time series models. In 2024 IEEE 6th International Conference on Power, Intelligent Computing and Systems (ICPICS). IEEE, 2024: 793-797

    work page 2024

  3. [4]

    Research on Effectiveness Evaluation and Optimization of Baseball Teaching Method Based on Machine Learning

    Sun S, Yuan J, Yang Y. Research on Effectiveness Evaluation and Optimization of Baseball Teaching Method Based on Machine Learning. arXiv preprint arXiv:2411.15721,

    arXiv

  4. [5]

    Real-Time Prediction for Athletes' Psychological States Using BERT-XGBoost: Enhancing Human-Computer Interaction

    Duan, C., Shu, Z., Zhang, J., & Xue, F. Real-Time Prediction for Athletes' Psychological States Using BERT-XGBoost: Enhancing Human-Computer Interaction. arXiv preprint arXiv:2412.05816,

    arXiv

  5. [6]

    COVID-19 localization and recognition on chest radiographs based on Yolov5 and EfficientNet

    Zhang T, Zhang B, Zhao F, et al. COVID-19 localization and recognition on chest radiographs based on Yolov5 and EfficientNet. In 2022 7th International Conference on Intelligent Computing and Signal Processing (ICSP). IEEE, 2022: 1827-1830

    work page 2022

  6. [7]

    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. arXiv preprint arXiv:2501.04996,

    arXiv

  7. [9]

    Multidimensional precipitation index prediction based on CNN-LSTM hybrid framework

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

    arXiv

  8. [10]

    CCi-YOLOv8n: Enhanced Fire Detection with CARAFE and Context-Guided Modules

    Lv K. CCi-YOLOv8n: Enhanced Fire Detection with CARAFE and Context-Guided Modules. arXiv preprint arXiv:2411.11011,

    arXiv

Show all 43 references
  1. [11]

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

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

    arXiv

  2. [12]

    CTLformer: A Hybrid Denoising Model Combining Convolutional Layers and Self-Attention for Enhanced CT Image Reconstruction

    Zheng Z, Wu S, Ding W. CTLformer: A Hybrid Denoising Model Combining Convolutional Layers and Self-Attention for Enhanced CT Image Reconstruction. arXiv preprint arXiv:2505.12203,

    arXiv

  3. [13]

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

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

    2024

  4. [14]

    Automated Parking Trajectory Generation Using Deep Reinforcement Learning

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

    arXiv

  5. [15]

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

    Zhao H, Ma Z, Liu L, et al. Optimized path planning for logistics robots using ant colony algorithm under multiple constraints. arXiv preprint arXiv:2504.05339,

    arXiv

  6. [16]

    Contextual Bandits for Unbounded Context Distributions

    Zhao, P., Fan, R., Wang, S., Shen, L., Zhang, Q., Ke, Z., & Zheng, T. Contextual Bandits for Unbounded Context Distributions. arXiv preprint arXiv:2408.09655,

    arXiv

  7. [17]

    Learn from Global Correlations: Enhancing Evolutionary Algorithm via Spectral GNN

    Ouyang, K., Fu, S., Ke, Z., Guan, R., Liang, K., & Hu, D. Learn from Global Correlations: Enhancing Evolutionary Algorithm via Spectral GNN. arXiv preprint arXiv:2412.17629,

  8. [18]

    AI-driven prognostics for state of health prediction in Li-ion batteries: A comprehensive analysis with validation

    Ding T, Xiang D, Sun T, et al. AI-driven prognostics for state of health prediction in Li-ion batteries: A comprehensive analysis with validation. arXiv preprint arXiv:2504.05728,

    arXiv

  9. [19]

    NeRF-Based defect detection

    Xiang D, Qi Y, Yang Z, et al. NeRF-Based defect detection. arXiv preprint arXiv:2504.00270,

    arXiv

  10. [21]

    Cloud-Based AI Systems: Leveraging Large Language Models for Intelligent Fault Detection and Autonomous Self-Healing

    Ji C, Luo H. Cloud-Based AI Systems: Leveraging Large Language Models for Intelligent Fault Detection and Autonomous Self-Healing. arXiv preprint arXiv:2505.11743,

    arXiv

  11. [22]

    Data Augmentation Through Random Style Replacement

    Yang Q, Ji C, Luo H, et al. Data Augmentation Through Random Style Replacement. arXiv preprint arXiv:2504.10563,

  12. [23]

    Application of an ANN and LSTM-based Ensemble Model for Stock Market Prediction

    Liu F, Guo S, Xing Q, et al. Application of an ANN and LSTM-based Ensemble Model for Stock Market Prediction. In 2024 IEEE 7th International Conference on Information Systems and Computer Aided Education (ICISCAE). IEEE, 2024: 390-395

    2024

  13. [24]

    Optimized Credit Score Prediction via an Ensemble Model and SMOTEENN Integration

    Cheng Y, Wang L, Sha X, et al. Optimized Credit Score Prediction via an Ensemble Model and SMOTEENN Integration. In 2024 IEEE 7th International Conference on Information Systems and Computer Aided Education (ICISCAE). IEEE, 2024: 355-361

    2024

  14. [25]

    Enhancing Code LLMs with Reinforcement Learning in Code Generation

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

    arXiv

  15. [26]

    Machine learning optimizes the efficiency of picking and packing in automated warehouse robot systems

    Yu D, Liu L, Wu S, et al. Machine learning optimizes the efficiency of picking and packing in automated warehouse robot systems. In 2025 IEEE International Conference on Electronics, Energy Systems and Power Engineering (EESPE). IEEE, 2025: 1325-1332

    2025

  16. [27]

    Research on reinforcement learning based warehouse robot navigation algorithm in complex warehouse layout

    Li K, Liu L, Chen J, et al. Research on reinforcement learning based warehouse robot navigation algorithm in complex warehouse layout. In 2024 6th International Conference on Artificial Intelligence and Computer Applications (ICAICA). IEEE, 2024: 296-301

    2024

  17. [29]

    Research and design on intelligent recognition of unordered targets for robots based on reinforcement learning

    Mao Y, Tao D, Zhang S, et al. Research and design on intelligent recognition of unordered targets for robots based on reinforcement learning. arXiv preprint arXiv:2503.07340,

    arXiv

  18. [30]

    Score: Story coherence and retrieval enhancement for ai narratives

    Yi Q, He Y, Wang J, et al. Score: Story coherence and retrieval enhancement for ai narratives. arXiv preprint arXiv:2503.23512,

  19. [31]

    Psychological health knowledge-enhanced LLM-based social network crisis intervention text transfer recognition method

    Wu S, Huang X, Lu D. Psychological health knowledge-enhanced LLM-based social network crisis intervention text transfer recognition method. arXiv preprint arXiv:2504.07983,

    arXiv

  20. [35]

    Research on Personalized Medical Intervention Strategy Generation System based on Group Relative Policy Optimization and Time-Series Data Fusion

    Lu D, Wu S, Huang X. Research on Personalized Medical Intervention Strategy Generation System based on Group Relative Policy Optimization and Time-Series Data Fusion. arXiv preprint arXiv:2504.18631,

    arXiv

  21. [36]

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

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

    arXiv

  22. [38]

    Intelligent road crack detection and analysis based on improved YOLOv8

    Zuo H, Li Z, Gong J, et al. Intelligent road crack detection and analysis based on improved YOLOv8. arXiv preprint arXiv:2504.13208,

    arXiv

  23. [39]

    Large language models as master key: unlocking the secrets of materials science with GPT

    Xie T, Wan Y, Huang W, et al. Large language models as master key: unlocking the secrets of materials science with GPT. arXiv preprint arXiv:2304.02213,

    arXiv

  24. [40]

    Darwin series: Domain specific large language models for natural science

    Xie T, Wan Y, Huang W, et al. Darwin series: Domain specific large language models for natural science. arXiv preprint arXiv:2308.13565,

    arXiv

  25. [41]

    Creation of a structured solar cell material dataset and performance prediction using large language models

    Xie T, Wan Y, Zhou Y, et al. Creation of a structured solar cell material dataset and performance prediction using large language models. Patterns, 2024, 5(5)

    2024

  26. [42]

    Generating Multimodal Images with GAN: Integrating Text, Image, and Style

    Tan C, Zhang W, Qi Z, et al. Generating Multimodal Images with GAN: Integrating Text, Image, and Style. arXiv preprint arXiv:2501.02167,

    arXiv

  27. [43]

    Real-time Video Target Tracking Algorithm Utilizing Convolutional Neural Networks (CNN)

    Tan C, Li X, Wang X, et al. Real-time Video Target Tracking Algorithm Utilizing Convolutional Neural Networks (CNN). In 2024 4th International Conference on Electronic Information Engineering and Computer (EIECT). IEEE, 2024: 847-851

    2024

  28. [44]

    A novel Bayesian Pay-As-You-Drive insurance model with risk prediction and causal mapping

    Wang B, Chen Y, Li Z. A novel Bayesian Pay-As-You-Drive insurance model with risk prediction and causal mapping. Decision Analytics Journal, 2024, 13: 100522

    2024

  29. [46]

    Research on splicing image detection algorithms based on natural image statistical characteristics

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

    arXiv

  30. [47]

    A multimodal fusion network for student emotion recognition based on transformer and tensor product

    Xiang, A., Qi, Z., Wang, H., Yang, Q., & Ma, D. A multimodal fusion network for student emotion recognition based on transformer and tensor product. In 2024 IEEE 2nd International Conference on Sensors, Electronics and Computer Engineering (ICSECE). IEEE, 2024: 1-4

    2024

  31. [48]

    User Behavior Analysis in Privacy Protection with Large Language Models: A Study on Privacy Preferences with Limited Data

    Yang H, Lu Q, Wang Y, et al. User Behavior Analysis in Privacy Protection with Large Language Models: A Study on Privacy Preferences with Limited Data. arXiv preprint arXiv:2505.06305,

    arXiv

  32. [49]

    Research on the Design of a Short Video Recommendation System Based on Multimodal Information and Differential Privacy

    Yang H, Fu L, Lu Q, et al. Research on the Design of a Short Video Recommendation System Based on Multimodal Information and Differential Privacy. arXiv preprint arXiv:2504.08751,

    arXiv

  33. [50]

    Enhanced Recommendation Combining Collaborative Filtering and Large Language Models

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

    arXiv

  34. [51]

    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]. arXiv preprint arXiv:2412.07997,

    arXiv

  35. [52]

    Deep learning-based anomaly detection and log analysis for computer networks[J]

    Wang S, Jiang R, Wang Z, et al. Deep learning-based anomaly detection and log analysis for computer networks[J]. arXiv preprint arXiv:2407.05639,

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.