Pith. sign in

REVIEW 5 major objections 4 minor 1 cited by

Deep Learning Model Acceleration and Optimization Strategies for Real-Time Recommendation Systems

T0 review · 5 major / 4 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 A staged pipeline of pruning, quantization, distillation, and inference acceleration cuts recommendation latency to 41% of baseline and more than doubles throughput, with accuracy loss under 1%.

desk verdict The paper's headline latency claim is contradicted by its own Table I — best case is 41% of baseline, not under 30% — and the accuracy claim is similarly overstated. read the letter →

arxiv 2506.11421 v3 pith:CIVTCQ65 submitted 2025-06-13 cs.IR cs.AIcs.LG

Junli Shao , Jing Dong , Dingzhou Wang , Kowei Shih , Dannier Li , Chengrui Zhou This is my paper
classification cs.IRcs.AIcs.LG
keywords real-timerecommendationsystemsmodelaccelerationstructuredpruningweightquantizationknowledgedistillationheterogeneouscomputinginferencelatencythroughputoptimization
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 tries to establish that deep recommendation models can be made fast enough for real-time serving by combining model-level compression with system-level acceleration rather than choosing one or the other. On the model side, it uses lightweight attention projections, dynamic-threshold pruning, 8-bit weight quantization, and knowledge distillation to shrink parameters and compute. On the system side, it relies on heterogeneous hardware backends, FP16 inference, elastic scheduling, and load balancing. The authors report that the fully optimized configuration cuts GPU latency from 52.4 ms to 21.5 ms (41% of baseline) and raises throughput from 190 to 460 requests per second, while keeping Hit Rate and NDCG within roughly one percent of the original model. A sympathetic reader would care because the recipe uses standard components that could be applied to existing recommender stacks without retraining from scratch.

What carries the argument

The cost model is the anchor: with hidden width $h$, depth $L$, and candidate-set size $m$, parameter count scales as $O(Lh^2)$ and per-batch FLOPs as $O(mLh^2)$, so shrinking any of these factors pays off multiplicatively. The paper's named machinery is the dynamic-threshold pruning loop (binary masks, iterative fine-tuning, then dynamic-range quantization with quantization-aware training), plus grouped/depthwise projections and low-rank head factorization to cut self-attention cost from $O(L^2d)$ toward $O(Lwd + L\log L\,d)$, and KL-divergence distillation from teacher to student. This set of mechanisms is what converts the asymptotic complexity into measured latency and throughput gains.

What would settle it

Reproduce Table I on the full dataset with the same candidate set and then deploy the two best variants in a live A/B test measuring end-to-end p99 latency and click-through rate; if the optimized models do not approach 21.5 ms versus 52.4 ms baseline at 460 requests per second, or if online accuracy drops by more than the reported one percent, the central claim fails.

Watch

Extended reading notes

Core claim

The central discovery, presented on the paper's own terms, is that the inference bottleneck of a deep recommender can be attacked at two levels whose gains multiply. At the model level, replacing dense self-attention projections with grouped and low-rank forms, pruning roughly 40% of parameters with dynamic thresholds, quantizing weights and activations to 8-bit, and distilling teacher attention maps into a smaller student cut parameter count and multiply-accumulate operations. At the system level, exporting the student model to a portable graph, running it on accelerated backends with FP16, and adding elastic scheduling and load balancing convert the smaller model into higher serving throughput. The reported best variant uses 6.4 million parameters (12.8 MB), reaches 21.5 ms GPU latency versus 52.4 ms for the baseline, and serves 460 requests per second versus 190, with Hit Rate and NDCG loss below 0.6% and MRR loss below 0.8%.

Load-bearing premise

The paper's real-time claims rest on an offline evaluation of a truncated version of a public e-commerce dataset with no live serving and no A/B test; if that offline setup does not reflect production traffic, the reported latency and throughput gains do not transfer.

Editorial extensions

If this is right

  • If the reported gains hold, a recommender serving tens of thousands of queries per second can use the pruned, quantized, distilled model on standard GPU/CPU hardware instead of a much larger cluster.
  • The combined pruning-quantization pipeline shrinks model storage to roughly 10-15% of the FP32 baseline, making edge and mobile deployment of deep recommenders more plausible.
  • Elastic scheduling and load balancing are claimed to absorb traffic spikes by routing to least-loaded backends and bypassing batching for high-priority requests, so SLO violations should decrease under bursty load.
  • Because accuracy loss stays below one percent in the offline benchmark, the paper argues that the optimized models can replace the baseline in production without a visible quality regression.

Reading between the lines

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

  • Inference: the paper's own numbers imply that model-level compression alone (pruning plus quantization) accounts for most of the latency gain, with distillation plus FP16 acceleration contributing the remaining step; a reader could test each stage's marginal contribution by ablation.
  • Inference: the deployed recipe described in the system-level section depends on infrastructure that is heavier than the model math, so the same latency gains would likely shrink when measured end-to-end through a live microservice stack.
  • Inference: a natural extension is to apply the same pipeline to a model with longer user-history sequences or larger candidate sets, where the quadratic attention and linear candidate-set terms would become more dominant and the relative gains may be larger or smaller.
  • Inference: the accuracy guarantee is stated for offline ranking metrics; online metrics such as click-through rate or revenue per request are a stricter test that the paper does not run.
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 / 4 minor

Summary. The paper proposes a combined model- and system-level optimization framework for real-time recommendation systems. Model-level techniques include lightweight attention (grouped projections, low-rank heads, distillation, sparse attention), dynamic-threshold pruning, and dynamic-range quantization; system-level techniques include heterogeneous platform integration, elastic scheduling, and load balancing. Experiments on the Taobao dataset compare five variants (baseline, quantized, pruned, pruned+quantized, distilled+TensorRT) on V100 and CPU. The abstract claims latency below 30% of baseline, throughput more than double, and accuracy preservation.

Significance. If the claims held, the paper would provide a useful engineering case study on combining standard acceleration techniques for recommendation models, with a public dataset and reproducible setup. The paper includes concrete measurements (Table I, Figures 5-7) and explicit model variants. However, the headline quantitative claims are contradicted by the paper's own data (41% best-case latency, not <30%; 2.7% Hit Rate loss for the combined pruning+quantization variant), and the system-level components are not evaluated. As a result, the stated contribution is not credible in its current form.

major comments (5)
  1. [Abstract and Sections I, V.A, Table I] The abstract states that the methods 'cut latency to less than 30% of the baseline,' but Table I reports the best V100 latency as 21.5 ms versus 52.4 ms baseline, which is 41.0%, and no model in the table is below 30%; even the Pruned+Quantized variant is 56.9% of baseline. The introduction and conclusion claim 'under 40%' and '60% latency reduction' respectively, which are inconsistent with the abstract. This internal contradiction means the paper's primary advertised result is not supported by its own evidence.
  2. [Section V.B, Figure 6] The abstract claims 'maintaining the original recommendation accuracy' and the introduction claims 'keeping accuracy loss below 1%.' However, Section V.B reports that pruning alone reduces Hit Rate by approximately 2.0%, and the combined Pruned+Quantized model reaches only 97.3% of baseline Hit Rate (2.7% loss). Only the Distilled+RT variant has losses below 0.6%. The accuracy-preservation claim must be restricted to the distilled variant, and the trade-offs for pruning and quantization should be stated honestly.
  3. [Sections IV and V] The system-level contributions (heterogeneous platform integration, elastic inference scheduling, load balancing, Kubernetes autoscaling) are described architecturally but never evaluated. All reported experiments measure per-model inference latency and throughput on a truncated offline dataset (latest 100 user entries, candidate set 50) with no end-to-end pipeline, no dynamic load test, and no serving deployment. The conclusion that the framework is a 'practical solution for deploying large-scale online recommendation services' is therefore unsupported.
  4. [Section V.A, Table I] The evaluation lacks comparisons against existing acceleration baselines (e.g., running the baseline through TensorRT FP16, ONNX Runtime INT8, or OpenVINO) and lacks repeated trials or error bars. Without these, it is unclear whether the observed gains come from the proposed 'closed-loop' pipeline or from standard inference engine optimizations, and the reliability of differences between variants (e.g., 44.1 vs 36.7 ms) cannot be assessed.
  5. [Section III.A] The mathematical notation is badly corrupted: complexity terms appear as 'O(d)!' and 'O("!#)', formula (4) is incomplete, and the low-rank decomposition description is garbled. These errors prevent verification of the claimed computational savings and must be corrected.
minor comments (4)
  1. [Throughout] There are numerous typos and formatting errors, e.g., 'To To' at the start of Section V.A, and inconsistent references to 'Table 5-3' and 'Table 5-4' that do not exist in the text.
  2. [References] Many references are unrelated to recommendation systems or acceleration (e.g., COVID-19 detection, baseball teaching methods, stock market prediction), which weakens the related-work context and suggests the bibliography is not carefully curated.
  3. [Section V.B] The text describes 'online recommendation quality metrics' but the evaluation is offline; the term 'online' is misleading in this context.
  4. [Figures and Tables] Figures 5-7 are referenced but their axes and units are not described in the text, and some tables are listed but not included; the paper should ensure all cited figures/tables are present and annotated.

Circularity Check

0 steps flagged · score 1.0 of 10

No meaningful circularity: the paper's quantitative claims are empirical measurements of standard acceleration techniques, not derived predictions; the headline discrepancy with Table I is a correctness issue, not a circularity.

full rationale

The paper's central results (latency 21.5 ms, throughput 460 req/s, Hit Rate ratios) are direct measurements on a truncated Taobao dataset for five configurations. There is no equation in which an output quantity is defined in terms of the quantity it is used to predict. The pruning ratio (p=0.4) and quantization bit-width (8-bit) are design inputs; reporting reduced parameter counts consistent with those inputs is definitional but not load-bearing for the latency/throughput/accuracy claims, which are independently measured. No uniqueness theorem is invoked, and no weight-bearing argument rests on a self-citation by the present authors. The abstract's 'less than 30%' latency claim is contradicted by Table I's best measured value of 21.5/52.4 = 41%, and the text also contains internal inconsistencies (e.g., ~45% quantization latency cut vs. 15.8% in Table I; <1% vs. 2.0% pruning accuracy loss), but these are correctness or reporting flaws, not circular reductions. Because the evaluation is self-contained and no prediction reduces to a fitted constant by construction, the circularity burden is low; the small nonzero score reflects only the trivial sense in which the reported parameter compression follows from the chosen compression targets.

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

No new entities are introduced; the paper combines existing compression and serving techniques. The ledger lists the free design parameters that control the experimental outcome and the unproven assumptions about complexity models and dataset representativeness.

free parameters (6)
  • Target pruning ratio p = 0.4
    Used in Eq. 5 to set the threshold θ; the paper states 'Given a target pruning ratio p' and later reports a 40% parameter reduction, so p=0.4 is a hand-chosen parameter that directly controls the pruned model size.
  • Pruning fine-tuning rounds K = 3
    Section III.B: 'Empirically, after K=3 rounds, we reduce total parameters by ≈40%'; K is a chosen hyperparameter.
  • Quantization bit-width b = 8
    Section III.A/B: weights and activations quantized from 32-bit to 8-bit; standard choice but a design parameter.
  • Sparse attention window size w = not specified
    Section III.A introduces a local window of size w and sparse sampling but never reports the actual value used in experiments.
  • Low-rank decomposition rank r = not specified
    Section III.A: projection matrices are rank-reduced from d to r≪d, but the chosen r is not reported.
  • Distillation loss weight = not specified
    Knowledge distillation is applied with KL divergence (Eq. 3), but the loss weighting or temperature is not stated.
assumptions (4)
  • standard math Complexity model P ≈ O(Lh²) and FLOPs ≈ O(mLh²) for the fully connected recommender network.
    Used in Section II.B as the foundation for the need for pruning/quantization; standard complexity accounting for MLPs.
  • standard math Self-attention has O(L²d) time and O(L²) space complexity.
    Section III.A, used to motivate sparse attention; standard known result.
  • domain assumption The Taobao User Behavior Dataset, truncated to 100 latest entries and 50 candidates, is representative of real-time e-commerce recommendation workloads.
    Section V.A, this assumption underlies the external validity of all reported latency/throughput/accuracy numbers.
  • domain assumption Latency can be approximated by τ ≈ α·mLh² + β with α a hardware-dependent constant.
    Section II.B, this linear scaling model is assumed to justify the acceleration approach, but the constants are not measured.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Learning Model Acceleration and Optimization Strategies for Real-Time Recommendation Systems." pith.science (2026). https://pith.science/paper/CIVTCQ65

@misc{pith2026250611421,
  author       = {Pith},
  title        = {Pith review of: Deep Learning Model Acceleration and Optimization Strategies for Real-Time Recommendation Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CIVTCQ65}},
  note         = {Machine review of arXiv:2506.11421}
}
read the original abstract

With the rapid growth of Internet services, recommendation systems play a central role in delivering personalized content. Faced with massive user requests and complex model architectures, the key challenge for real-time recommendation systems is how to reduce inference latency and increase system throughput without sacrificing recommendation quality. This paper addresses the high computational cost and resource bottlenecks of deep learning models in real-time settings by proposing a combined set of modeling- and system-level acceleration and optimization strategies. At the model level, we dramatically reduce parameter counts and compute requirements through lightweight network design, structured pruning, and weight quantization. At the system level, we integrate multiple heterogeneous compute platforms and high-performance inference libraries, and we design elastic inference scheduling and load-balancing mechanisms based on real-time load characteristics. Experiments show that, while maintaining the original recommendation accuracy, our methods cut latency to less than 30% of the baseline and more than double system throughput, offering a practical solution for deploying large-scale online recommendation services.

Figures

Figures reproduced from arXiv: 2506.11421 by the authors.

Figure 1
Figure 1. Illustration of the end-to-end reasoning process of a deep learning driven real-time recommender system In Offline recommendation allows higher latency, but real-time systems must complete the full pipeline—from user action to result delivery—within tens to hundreds of milliseconds ( [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of feature weighting of user behavior sequence based on self-attention Deep recommendation models use self-attention to capture temporal and contextual dependencies. As in [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Schematic diagram of the stepwise threshold driven neuron pruning with weight quantization process Specifically, let the weight matrix of a layer be W ∈ R+×4 , and the elements be denoted w56 . The process is divided into two main stages. Stage 1: Dynamic-Threshold Pruning Threshold Computation: Given a target pruning ratio p, sort {∣ w56 ∣} in ascending order and choose the initial threshold θ(7) as shown in Formul… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Deep Recommendation Model Training and Deployment Architecture Based on Weighted Knowledge Distillation For mobile and edge deployment, models run on TensorFlow Lite or SNPE, targeting NPU/DSP for efficiency. In the cloud, models use asynchronous microservices with Kub…
Figure 5
Figure 5. Figure 5: shows the average latency and maximum throughput on GPU (V100) and CPU platforms for each model [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Accuracy Comparison As shown in [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]
Figure 7
Figure 7. Figure 7: Resource Consumption Comparison [PITH_FULL_IMAGE:figures/full_fig_p004_7.png]

Discussion (0). Sign in 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. Instructional Prompt Optimization for Few-Shot LLM-Based Recommendations on Cold-Start Users

    cs.AI 2025-09 reject novelty 3.0 of 10

    A manuscript claims instructional prompt engineering improves LLM-based cold-start recommendation, but provides no reproducible evidence.

Reference graph

Works this paper leans on

34 extracted references · 7 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 multiprocessor system

    Su, Pei-Chiang, et al. "A Mixed-Heuristic Quantum-Inspired Simplified Swarm Optimization Algorithm for scheduling of real-time tasks in the multiprocessor system." Applied Soft Computing 131 (2022): 109807

    work page 2022

  2. [2]

    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[J]. arXiv preprint arXiv:2411.15721,

    work page Pith review arXiv

  3. [3]

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

    Duan, Chenming, et al. "Real-Time Prediction for Athletes' Psychological States Using BERT-XGBoost: Enhancing Human-Computer Interaction." arXiv preprint arXiv:2412.05816 (2024)

    arXiv 2024

  4. [4]

    In the cloud, models use asynchronous microservices with Kubernetes/Kubeflow, supporting dynamic replica scaling

    Deep Recommendation Model Training and Deployment Architecture Based on Weighted Knowledge Distillation For mobile and edge deployment, models run on TensorFlow Lite or SNPE, targeting NPU/DSP for efficiency. In the cloud, models use asynchronous microservices with Kubernetes/Kubeflow, supporting dynamic replica scaling. Mixed-precision training and auto-...

    work page 2021

  5. [5]

    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

  6. [6]

    Deep Learning-based Anomaly Detection and Log Analysis for Computer Networks

    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,

    work page Pith review arXiv

  7. [7]

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

    2022

  8. [8]

    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]. arXiv preprint arXiv:2501.04996,

    arXiv

Show all 34 references
  1. [9]

    Analysis of collective response reveals that covid-19-related activities start from the end of 2019 in mainland china[J]

    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

  2. [10]

    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,

    arXiv

  3. [11]

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

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

    arXiv

  4. [12]

    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,

    arXiv

  5. [13]

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

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

    arXiv

  6. [14]

    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

    2024

  7. [15]

    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,

    arXiv

  8. [16]

    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,

    arXiv

  9. [17]

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

    2024 arXiv

  10. [18]

    (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 Conference on Sensors, Electronics and Computer Engineering (ICSECE) (pp. 1-4). IEEE

    2024

  11. [19]

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

    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[J]. arXiv preprint arXiv:2504.08751,

    arXiv

  12. [20]

    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,

    arXiv

  13. [21]

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

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

    arXiv

  14. [22]

    Data Augmentation Through Random Style Replacement[J]

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

  15. [23]

    Mao, Y., Tao, D., Zhang, S., Qi, T., & Li, K. (2025). Research and Design on Intelligent Recognition of Unordered Targets for Robots Based on Reinforcement Learning. arXiv preprint arXiv:2503.07340

    2025 arXiv

  16. [24]

    & Shi, T

    Yi, Q., He, Y., Wang, J., Song, X., Qian, S., Zhang, M., ... & Shi, T. (2025). SCORE: Story Coherence and Retrieval Enhancement for AI Narratives. arXiv preprint arXiv:2503.23512

    2025

  17. [25]

    Qiu, S., Wang, Y., Ke, Z., Shen, Q., Li, Z., Zhang, R., & Ouyang, K. (2025). A Generative Adversarial Network-Based Investor Sentiment Indicator: Superior Predictability for the Stock Market. Mathematics, 13(9),

    2025

  18. [26]

    Ouyang, K., Fu, S., & Ke, Z. (2024). Graph Neural Networks Are Evolutionary Algorithms. arXiv preprint arXiv:2412.17629

    2024

  19. [27]

    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,

    arXiv

  20. [28]

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

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

    arXiv

  21. [29]

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

    2024

  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,

    arXiv

  23. [31]

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

    Zhao H, 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,

    arXiv

  24. [34]

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

    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[J]. arXiv preprint arXiv:2504.18631,

    arXiv

  25. [35]

    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,

    arXiv

  26. [36]

    Contextual bandits for unbounded context distributions[J]

    Zhao P, Wu J, Liu Z, et al. Contextual bandits for unbounded context distributions[J]. arXiv preprint arXiv:2408.09655,

    arXiv

Pith tools

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