AI-Driven Health Monitoring of Distributed Computing Architecture: Insights from XGBoost and SHAP

Pochun Li; Xiaoxuan Sun; Xiaoye Wang; Xuan Li; Yue Yao

REVIEW 4 major objections 4 minor 21 references

AI-Driven Health Monitoring of Distributed Computing Architecture: Insights from XGBoost and SHAP

T0 review · 4 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read The paper claims that an XGBoost classifier, with SHAP explanations, is the most accurate of six models for judging edge-computing node health and that response time, power consumption, and disk I/O drive the judgments.

desk verdict Routine XGBoost+SHAP application whose only table is uninterpretable: 47.5% accuracy is chance-level and no protocol or data is reported. read the letter →

arxiv 2501.14745 v1 pith:BGGUQN2R submitted 2024-12-16 cs.DC cs.LG

Xiaoxuan Sun , Yue Yao , Xiaoye Wang , Pochun Li , Xuan Li This is my paper · ORCID

classification cs.DCcs.LG

keywords edgecomputinghealthmonitoringXGBoostSHAPinterpretabilitydistributedsystemsnodestatus

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 an XGBoost classifier, explained with SHAP values, gives the best health-status judgments for edge-computing nodes among six common models on its collected dataset, while also making the reasoning behind each judgment visible. XGBoost reaches 47.5% accuracy and a 50.39% F1 score, ahead of support vector machines, random forests, k-nearest neighbors, naive Bayes, and a multilayer perceptron. SHAP analysis identifies response time, power consumption, and disk I/O as the features that most influence predictions, offering a concrete list of indicators for operators to monitor. If these claims hold, edge-computing systems could move from opaque anomaly scoring to interpretable, indicator-driven health monitoring.

What carries the argument

The machinery is XGBoost, a regularized gradient-boosted decision-tree ensemble that uses first- and second-order gradients to grow trees, combined with SHAP, which computes each feature's contribution as a weighted average of marginal effects over all feature subsets. XGBoost supplies the classifier; SHAP supplies global and local interpretability. The paper's argument depends on the difference between XGBoost's split-frequency weight importance and SHAP's prediction-contribution importance, using the contrast to claim that response time is the decisive operational indicator even though power consumption and disk I/O dominate split counts.

What would settle it

Re-run the six models on the same dataset with a documented train/test split, then shuffle the health labels and repeat; if shuffled labels yield a similar ranking or if all models sit near the majority-class baseline, the claimed superiority of XGBoost would be refuted.

Watch

Extended reading notes

Core claim

The central claim is that gradient boosting with XGBoost handles the nonlinear, multi-feature patterns of edge-computing node data better than the five baselines, and that SHAP values provide a trustworthy account of which features matter. The paper reports two complementary importance rankings: XGBoost's built-in weight metric, which counts how often a feature is used at tree splits, puts power consumption and disk I/O first, while SHAP's contribution-based ranking puts response time first, followed by power consumption and disk I/O. The paper interprets this contrast as a feature, not a flaw: weight reflects construction, SHAP reflects actual contribution to predictions, and together they guide monitoring priorities and anomaly diagnosis.

Load-bearing premise

The reported comparison rests on the assumption that the privately collected dataset is correctly labeled and representative of edge-node health, and that the train/test split is sound; if labels are noisy or the split leaks, the performance ranking is meaningless.

Editorial extensions

If this is right

If the central claim is correct, monitoring dashboards for edge nodes should lead with response time, power consumption, and disk I/O.
If the central claim is correct, SHAP explanations can turn each health alert into an attributable cause, such as elevated response time, which supports faster remediation.
If the central claim is correct, the same XGBoost-plus-SHAP pipeline can be applied to other distributed-system health monitoring tasks with comparable tabular telemetry.
If the central claim is correct, operators can use the weight-versus-SHAP distinction to avoid over-trusting split-based feature rankings.

Reading between the lines

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

Editorial inference: the accuracy gap between XGBoost (47.5%) and MLP (46.8%) is small, so the practical benefit on real workloads may be modest even if the ranking is stable.
Editorial inference: the SHAP color-direction claims, such as high response times aligning with healthy nodes, could reflect a dataset-specific pattern and should be validated against operational logs before being used to set alarms.
Editorial inference: the reference markers attached to GBDT, XGBoost, and SHAP in the method section point to unrelated work, so the paper's central claim rests on its own experiment rather than on the cited sources.

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.

Desk Editor's Note

read the letter

Honest take: this is a routine XGBoost+SHAP application to a private edge-node health dataset, and the reported results don't support the title. The one thing a reader should know: the best accuracy in Table 1 is 47.5%, which is at chance for a balanced binary problem, and the paper never reports class balance, sample size, split, or error bars. So the headline claim that XGBoost is superior collapses.

What is actually new: nothing in method. The contribution is the private dataset, which is not described beyond a list of eight indicators. To the paper's credit, the methodology section correctly states the XGBoost objective and SHAP equations, and the writing is clear about what weight-based importance vs SHAP importance mean. The SHAP discussion does distinguish feature usage frequency from actual contribution. That is honest. The conclusion also admits feature selection depends on the existing dataset and dynamic modeling is insufficient; those are real limitations.

Soft spots are large. The experiment is uninterpretable. No majority baseline, no confusion matrix, no precision/recall, no repeated runs. The dataset section is qualitative; no labeling procedure or collection period. The references are mostly irrelevant to edge computing health monitoring; many are about LLMs, GNNs, and medical text. The SHAP narrative is hard to reconcile: the abstract says response time and power consumption are the key features, but Figure 1 ranks power and disk IO highest by weight, and Figure 2 says response time is highest by SHAP. That difference is not necessarily contradictory, but the paper doesn't explain how the two rankings align with the claimed captured features.

Proportionate verdict: this is not a fraudulent paper, it is an under-developed workshop-level report. It might have been acceptable as a short paper if the full experimental protocol, data description, and code were included, and if the accuracy had been put next to a majority baseline. As it stands, the central claim is unsupported. I would not send this to peer review in its current form; a serious editor would desk reject and invite a resubmission with real experiments, data release, and a comparison against trivial baselines. If the authors can show the dataset is balanced and the split is clean, 47.5% might still be at chance, so the method needs a lot more work before it supports the conclusions.

Referee Report

4 major / 4 minor

Summary. The paper proposes an XGBoost-based binary classifier for edge-computing node health status and uses SHAP to interpret feature contributions. It reports an experiment comparing XGBoost against SVM, Random Forest, KNN, Naive Bayes, and MLP on a privately collected dataset, with XGBoost achieving the best accuracy of 47.5% and F1-score of 50.39%. The paper further presents weight-based and SHAP-based feature importance analyses and claims that response time, power consumption, and disk IO are key drivers of node health. The central claim is that XGBoost is superior for this task and that SHAP provides actionable interpretability for system optimization.

Significance. If the empirical claims were substantiated, the paper would offer a moderately useful application of standard machine learning and interpretability tools to edge-node monitoring. However, the manuscript provides no reproducible experimental protocol: the dataset is described only qualitatively, no data or code are made available, no statistical validation is reported, and the headline accuracy is at or near chance for a balanced binary problem. The SHAP analysis is descriptive of a fitted model whose predictive performance is not established, so the practical conclusions are unsupported. The paper's acknowledged limitations in the conclusion confirm, rather than repair, the absence of evidence for dynamic and generalizable monitoring. I find no machine-checked proofs, reproducible code, or parameter-free derivations to credit; the contribution is currently an unvalidated case study.

major comments (4)

[III-B, Table 1] The central claim that XGBoost is superior rests entirely on Table 1, where XGBoost achieves 47.5% accuracy and 50.39% F1. For a binary classification task this is at or near chance if the classes are balanced, and if the classes are imbalanced the value could be below or above a trivial majority-class baseline. The paper reports no dataset size, class distribution, train/test split, cross-validation procedure, repeated runs, error bars, confusion matrix, or majority-class baseline, so the ranking in Table 1 is uninterpretable and the comparison is not a controlled experiment.
[III-A and IV] Section III-A describes the dataset only qualitatively: it lists eight indicators and states that healthy and abnormal samples are present, but gives no sample size, class balance, labeling procedure, collection period, feature preprocessing, normalization, or division into training and test sets. Without this metadata the experiment cannot be reproduced or assessed for leakage or label noise. The conclusion's statement that feature selection depends on the existing dataset does not supply the missing information; it merely acknowledges that the analysis is dataset-specific.
[III-C] The SHAP analysis is computed from the same fitted XGBoost model whose accuracy is not demonstrated to be above chance, so the feature-importance narrative cannot validate the paper's practical recommendations. In addition, the directional claims appear internally inconsistent: Figure 3's text states that high response times are primarily associated with healthy nodes while low response times align with abnormalities, and Figure 4 states that low CPU usage is associated with abnormal nodes, whereas the abstract and introduction identify response time and power consumption as indicators of abnormal status. The label encoding and the direction of effects are not defined, so these interpretation claims are unsupported.
[III-C, Figures 1-4] The paper does not provide any quantitative evaluation of the SHAP or weight-based importance results, such as feature importance values, rankings with confidence, or validation against independent outcomes. The figures appear to be screenshots without axis labels or numeric scales in the manuscript text. Since the model's predictive performance is near chance, the feature importance rankings cannot be taken as evidence about the actual drivers of edge-node health.

minor comments (4)

[II] The equations in Section II contain notation errors and OCR artifacts: the second-order gradient is incorrectly also denoted g_i, and in the SHAP formula the symbol A is used for the feature-subset size without definition, where the standard notation is |S|. The displayed formulas should be rewritten cleanly.
[III-C, Figures 1-4] Figures 1-4 lack axis labels, units, and numeric scales, making it impossible to read the claimed importance values or SHAP magnitudes from the text.
[References] Several citations do not support the claims at the positions where they appear; for example, [19] is cited for the GBDT basis and [20] for XGBoost, but the listed titles are about few-shot learning and graph neural networks, respectively. The reference list should be checked and corrected throughout.
[General] The manuscript contains formatting artifacts in the keyword line and in some equations, and the Introduction and Conclusion discuss 'traditional methods' without identifying or citing any such methods for edge-node health monitoring. These issues should be corrected in a revision.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical XGBoost + SHAP study; SHAP post hoc explains the fitted model, and the performance claim rests on a benchmark table rather than on a self-citation or definitional reduction.

full rationale

The paper makes an empirical performance claim (XGBoost accuracy/F1 on a private dataset, Table 1) and an interpretability claim (SHAP values of that fitted model). No step in the derivation chain equates an output with an input by construction. The XGBoost objective in Section II is the standard gradient-boosting loss, not defined in terms of the experimental outcome. The SHAP values in Section III-C are computed from the already-trained XGBoost model, but that is the intended and standard use of SHAP as a post hoc explanation; it is not a case of fitting a parameter to one quantity and then presenting a closely related quantity as an independent prediction. The reference list contains no load-bearing self-citation, and no uniqueness theorem or ansatz is imported from prior work by these authors. The conclusion's admitted limitation that 'feature selection depends on existing data sets' is a scope caveat rather than a circular reduction. The absence of dataset size, class balance, train/test split, label source, and error bars is a serious rigor and reproducibility concern, but under the specified criteria that is correctness risk, not circularity. Therefore no circular step is identified; score 0.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

No new entities or formal axioms are introduced. The paper's central claim depends entirely on unstated properties of the private dataset and on standard XGBoost and SHAP assumptions.

free parameters (1)

XGBoost hyperparameters = not reported
The paper does not report learning rate, tree depth, number of estimators, regularization, or any hyperparameter settings; model performance depends on these choices.

assumptions (3)

domain assumption The eight listed operating indicators (CPU, memory, disk IO, network latency, power consumption, temperature, process count, response time) are sufficient and correctly measured to determine node health.
Section III-A asserts the features 'comprehensively describe the operating status of the node' without validation or feature selection analysis.
domain assumption The binary health labels (y=1 healthy, y=0 abnormal) are reliably annotated.
Section II defines the labels but provides no labeling procedure, annotation criteria, or validation.
domain assumption The dataset is representative of real edge computing node operating conditions.
Section III-A states the data comes from actual nodes but gives no collection protocol, time span, or node population, so generalizability is assumed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AI-Driven Health Monitoring of Distributed Computing Architecture: Insights from XGBoost and SHAP." pith.science (2026). https://pith.science/paper/BGGUQN2R

@misc{pith2026250114745,
  author       = {Pith},
  title        = {Pith review of: AI-Driven Health Monitoring of Distributed Computing Architecture: Insights from XGBoost and SHAP},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BGGUQN2R}},
  note         = {Machine review of arXiv:2501.14745}
}

read the original abstract

With the rapid development of artificial intelligence technology, its application in the optimization of complex computer systems is becoming more and more extensive. Edge computing is an efficient distributed computing architecture, and the health status of its nodes directly affects the performance and reliability of the entire system. In view of the lack of accuracy and interpretability of traditional methods in node health status judgment, this paper proposes a health status judgment method based on XGBoost and combines the SHAP method to analyze the interpretability of the model. Through experiments, it is verified that XGBoost has superior performance in processing complex features and nonlinear data of edge computing nodes, especially in capturing the impact of key features (such as response time and power consumption) on node status. SHAP value analysis further reveals the global and local importance of features, so that the model not only has high precision discrimination ability but also can provide intuitive explanations, providing data support for system optimization. Research shows that the combination of AI technology and computer system optimization can not only realize the intelligent monitoring of the health status of edge computing nodes but also provide a scientific basis for dynamic optimization scheduling, resource management and anomaly detection. In the future, with the in-depth development of AI technology, model dynamics, cross-node collaborative optimization and multimodal data fusion will become the focus of research, providing important support for the intelligent evolution of edge computing systems.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 4 canonical work pages

[1]

Analyze the Impact of the Epidemic on New York Taxis by Machine Learning Algorithms and Recommendations for Optimal Prediction Algorithms,

Z. Liu, X. Xia, H. Zhang and Z. Xie, "Analyze the Impact of the Epidemic on New York Taxis by Machine Learning Algorithms and Recommendations for Optimal Prediction Algorithms," Proceedings of the 2021 3rd International Conference on Robotics Systems and Automation Engineering, pp. 46-52, May 2021

2021
[2]

Adaptive Optimization for Enhanced Efficiency in Large-Scale Language Model Training,

J. Chen, B. Liu, X. Liao, J. Gao, H. Zheng, and Y. Li, "Adaptive Optimization for Enhanced Efficiency in Large-Scale Language Model Training," arXiv preprint, 2024

work page 2024
[3]

Harnessing LLMs for API Interactions: A Framework for Classification and Synthetic Data Generation,

C. Tao, X. Fan, and Y. Yang, "Harnessing LLMs for API Interactions: A Framework for Classification and Synthetic Data Generation," arXiv preprint arXiv:2409.11703, 2024

arXiv 2024
[4]

Deep Learning for Medical Text Processing: BERT Model Fine-Tuning and Comparative Study

J. Hu, Y. Cang, G. Liu, M. Wang, W. He and R. Bao, "Deep Learning for Medical Text Processing: BERT Model Fine-Tuning and Comparative Study", arXiv preprint arXiv:2410.20792, 2024

arXiv 2024
[5]

Improving the RAG- based Personalized Discharge Care System by Introducing the Memory Mechanism

Y. Yang, C. Xu, J. Guo, T. Feng, and C. Ruan, "Improving the RAG- based Personalized Discharge Care System by Introducing the Memory Mechanism", Preprints, doi: 10.20944/preprints202410.1696.v1, 2024

arXiv 2024
[6]

Enhancing Recommendation Systems with GNNs and Addressing Over- Smoothing

W. Liu, Z. Zhang, X. Li, J. Hu, Y. Luo and J. Du, "Enhancing Recommendation Systems with GNNs and Addressing Over- Smoothing", arXiv preprint arXiv:2412.03097, 2024

arXiv 2024
[7]

Metric Learning for Tag Recommendation: Tackling Data Sparsity and Cold Start Issues

Y. Luo, R. Wang, Y. Liang, A. Liang and W. Liu, "Metric Learning for Tag Recommendation: Tackling Data Sparsity and Cold Start Issues", arXiv preprint arXiv:2411.06374, 2024

arXiv 2024
[8]

Advanced Risk Prediction and Stability Assessment of Banks Using Time Series Transformer Models

W. Sun, Z. Xu, W. Zhang, K. Ma, Y. Wu and M. Sun, "Advanced Risk Prediction and Stability Assessment of Banks Using Time Series Transformer Models", arXiv preprint arXiv:2412.03606, 2024

arXiv 2024

Show all 21 references

[9]

Multi-Source Data-Driven LSTM Framework for Enhanced Stock Price Prediction and Volatility Analysis

Z. Xu, W. Zhang, Y. Sun and Z. Lin, "Multi-Source Data-Driven LSTM Framework for Enhanced Stock Price Prediction and Volatility Analysis", Journal of Computer Technology and Software, vol. 3, no. 8, 2024

2024
[10]

Survival prediction across diverse cancer types using neural networks

X. Yan, W. Wang, M. Xiao, Y. Li, and M. Gao, "Survival prediction across diverse cancer types using neural networks", Proceedings of the 2024 7th International Conference on Machine Vision and Applications, pp. 134-138, 2024

2024
[11]

A Hybrid Model for Predicting Missing Records in Data Using XGBoost,

P. Ndayishimiyepas, C. Wilson and M. Kimwele, "A Hybrid Model for Predicting Missing Records in Data Using XGBoost," 2022 IEEE International Symposium on Product Compliance Engineering - Asia (ISPCE-ASIA), Guangzhou, Guangdong Province, China, 2022, pp. 1-5

2022
[12]

Optimizing Gesture Recognition for Seamless UI Interaction Using Convolutional Neural Networks

Q. Sun, T. Zhang, S. Gao, L. Yang and F. Shao, "Optimizing Gesture Recognition for Seamless UI Interaction Using Convolutional Neural Networks", arXiv preprint arXiv:2411.15598, 2024

2024 arXiv
[13]

Emotion-Aware Interaction Design in Intelligent User Interface Using Multi-Modal Deep Learning

S. Duan, Z. Wang, S. Wang, M. Chen and R. Zhang, "Emotion-Aware Interaction Design in Intelligent User Interface Using Multi-Modal Deep Learning", arXiv preprint arXiv:2411.06326, 2024

2024 arXiv
[14]

Deep Learning with HM- VGG: AI Strategies for Multi-modal Image Analysis

J. Du, Y. Cang, T. Zhou, J. Hu and W. He, "Deep Learning with HM- VGG: AI Strategies for Multi-modal Image Analysis", arXiv preprint arXiv:2410.24046, 2024

2024 arXiv
[15]

Scaling-up Medical Vision-and- Language Representation Learning with Federated Learning,

S. Lu, Z. Liu, T. Liu and W. Zhou, "Scaling-up Medical Vision-and- Language Representation Learning with Federated Learning," Engineering Applications of Artificial Intelligence, vol. 126, Article ID 107037, 2023

2023
[16]

A Combined Encoder and Transformer Approach for Coherent and High- Quality Text Generation

J. Chen, S. Wang, Z. Qi, Z. Zhang, C. Wang and H. Zheng, "A Combined Encoder and Transformer Approach for Coherent and High- Quality Text Generation", arXiv preprint arXiv:2411.12157, 2024

2024 arXiv
[17]

A Self-training Framework for Automated Medical Report Generation,

S. Wang, Z. Liu and B. Peng, "A Self-training Framework for Automated Medical Report Generation," Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp. 16443-16449, December 2023

2023
[18]

Graph Neural Network-Based Entity Extraction and Relationship Reasoning in Complex Knowledge Graphs

J. Du, G. Liu, J. Gao, X. Liao, J. Hu and L. Wu, "Graph Neural Network-Based Entity Extraction and Relationship Reasoning in Complex Knowledge Graphs", arXiv preprint arXiv:2411.15195, 2024

2024 arXiv
[19]

Enhancing Few- Shot Learning with Integrated Data and GAN Model Approaches

Y. Feng, A. Shen, J. Hu, Y. Liang, S. Wang and J. Du, "Enhancing Few- Shot Learning with Integrated Data and GAN Model Approaches", arXiv preprint arXiv:2411.16567, 2024

2024 arXiv
[20]

Self- Supervised Graph Neural Networks for Enhanced Feature Extraction in Heterogeneous Information Networks

J. Wei, Y. Liu, X. Huang, X. Zhang, W. Liu and X. Yan, "Self- Supervised Graph Neural Networks for Enhanced Feature Extraction in Heterogeneous Information Networks", 2024 5th International Conference on Machine Learning and Computer Application (ICMLCA), pp. 272-276, 2024

2024
[21]

Leveraging Semi-Supervised Learning to Enhance Data Mining for Image Classification under Limited Labeled Data

A. Shen, M. Dai, J. Hu, Y. Liang, S. Wang and J. Du, "Leveraging Semi-Supervised Learning to Enhance Data Mining for Image Classification under Limited Labeled Data", arXiv preprint arXiv:2411.18622, 2024

2024 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

Analyze the Impact of the Epidemic on New York Taxis by Machine Learning Algorithms and Recommendations for Optimal Prediction Algorithms,

Z. Liu, X. Xia, H. Zhang and Z. Xie, "Analyze the Impact of the Epidemic on New York Taxis by Machine Learning Algorithms and Recommendations for Optimal Prediction Algorithms," Proceedings of the 2021 3rd International Conference on Robotics Systems and Automation Engineering, pp. 46-52, May 2021

2021

[2] [2]

Adaptive Optimization for Enhanced Efficiency in Large-Scale Language Model Training,

J. Chen, B. Liu, X. Liao, J. Gao, H. Zheng, and Y. Li, "Adaptive Optimization for Enhanced Efficiency in Large-Scale Language Model Training," arXiv preprint, 2024

work page 2024

[3] [3]

Harnessing LLMs for API Interactions: A Framework for Classification and Synthetic Data Generation,

C. Tao, X. Fan, and Y. Yang, "Harnessing LLMs for API Interactions: A Framework for Classification and Synthetic Data Generation," arXiv preprint arXiv:2409.11703, 2024

arXiv 2024

[4] [4]

Deep Learning for Medical Text Processing: BERT Model Fine-Tuning and Comparative Study

J. Hu, Y. Cang, G. Liu, M. Wang, W. He and R. Bao, "Deep Learning for Medical Text Processing: BERT Model Fine-Tuning and Comparative Study", arXiv preprint arXiv:2410.20792, 2024

arXiv 2024

[5] [5]

Improving the RAG- based Personalized Discharge Care System by Introducing the Memory Mechanism

Y. Yang, C. Xu, J. Guo, T. Feng, and C. Ruan, "Improving the RAG- based Personalized Discharge Care System by Introducing the Memory Mechanism", Preprints, doi: 10.20944/preprints202410.1696.v1, 2024

arXiv 2024

[6] [6]

Enhancing Recommendation Systems with GNNs and Addressing Over- Smoothing

W. Liu, Z. Zhang, X. Li, J. Hu, Y. Luo and J. Du, "Enhancing Recommendation Systems with GNNs and Addressing Over- Smoothing", arXiv preprint arXiv:2412.03097, 2024

arXiv 2024

[7] [7]

Metric Learning for Tag Recommendation: Tackling Data Sparsity and Cold Start Issues

Y. Luo, R. Wang, Y. Liang, A. Liang and W. Liu, "Metric Learning for Tag Recommendation: Tackling Data Sparsity and Cold Start Issues", arXiv preprint arXiv:2411.06374, 2024

arXiv 2024

[8] [8]

Advanced Risk Prediction and Stability Assessment of Banks Using Time Series Transformer Models

W. Sun, Z. Xu, W. Zhang, K. Ma, Y. Wu and M. Sun, "Advanced Risk Prediction and Stability Assessment of Banks Using Time Series Transformer Models", arXiv preprint arXiv:2412.03606, 2024

arXiv 2024