Pith. sign in

REVIEW 5 major objections 5 minor 44 references

An Intelligent Fault Self-Healing Mechanism for Cloud AI Systems via Integration of Large Language Models and Deep Reinforcement Learning

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 The paper claims a hybrid framework combining a large language model with deep reinforcement learning cuts cloud fault recovery time by 37 percent on unknown fault scenarios while keeping the highest detection accuracy among four baselines.

desk verdict The 37% recovery-time claim fails on the paper's own numbers, and the 'unknown fault scenarios' in the abstract are never operationalized; the architecture is a reasonable hybrid but the evaluation is not trustworthy. read the letter →

arxiv 2506.07411 v1 pith:SRQXEJZJ submitted 2025-06-09 cs.AI

Ze Yang , Yihong Jin , Juntian Liu , Xinhe Xu This is my paper
classification cs.AI
keywords faultself-healingcloudAIsystemslargelanguagemodelsdeepreinforcementlearningdetectionaccuracyrecoverytimecatastrophicforgettingprompttuning
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 proposes IFSHM, a fault self-healing framework for cloud-based AI systems that combines a large language model with deep reinforcement learning. The LLM reads logs, resource metrics, and alarms to build a semantic picture of what is failing, and the reinforcement-learning module picks a recovery action from a hierarchical menu of restart, migrate, or scale operations. A memory-guided meta-controller replays rare fault experiences and fine-tunes prompts so the system adapts to new failure modes without forgetting old ones. On an OpenStack fault-injection benchmark, the authors report that IFSHM shortens system recovery time by 37 percent on unknown fault scenarios and maintains the highest fault detection accuracy, rising from 0.89 to 0.92 as training data grows. If these results hold, automated self-healing could replace rigid rule-based recovery for failures that were never seen in advance.

What carries the argument

The load-bearing mechanism is the two-stage hybrid architecture: an LLM-driven fault semantic interpretation module plus a DRL recovery strategy optimizer, regulated by a memory-guided meta-controller. The LLM module maps multimodal inputs $\mathcal{X}_t = (x^L_t, x^M_t, x^A_t)$ through a modal-aware encoder and deep attention into a semantic state vector $z_t$, which is what lets the policy reason about semantically equivalent faults (a CPU spike, a GC failure, memory overflow, response timeout) as one root cause. The DRL module uses a hierarchical action space with high-level policy categories and low-level parameterized executions, optimized by the clipped PPO objective. The meta-controller performs TD-error-weighted replay of rare trajectories and gradient-based prompt tuning, expressed in the multi-task loss $\mathcal{L}_{\text{total}}$, to keep adapting to new failure modes without catastrophic forgetting. This combination is what the paper claims improves exploration efficiency and generalization enough to cut recovery time by 37 percent.

What would settle it

Re-run the comparison on a held-out slice of the same OpenStack fault-injection data containing fault types excluded from training, with the four baseline methods' exact implementations and hyperparameters; the paper's claim predicts IFSHM's mean recovery time is at least 37 percent shorter than every baseline, so overlapping confidence intervals or a smaller gap would falsify the headline result.

Watch

Extended reading notes

Core claim

The central claim is that semantic understanding is the missing ingredient in DRL-based fault recovery: mapping raw multimodal signals into an LLM-generated state vector makes the reinforcement-learning policy generalize to fault types it was not trained on. To show this, the paper builds a two-stage architecture in which a modal-aware encoder concatenates log fragments, metric time windows, and alarm outputs into a unified context, and an LLM transforms that context into a semantic state vector through attention. A hierarchical policy then selects a high-level recovery category (restart, migration, elastic scaling) and low-level execution parameters, trained with a clipped PPO objective plus generalized advantage estimation. A memory-guided meta-controller stores high-value trajectories, samples them by TD error, and applies gradient-based prompt tuning so the LLM can recognize new semantic patterns. The reported evidence is that IFSHM reaches the highest fault detection accuracy (0.89 to 0.92) and the shortest recovery time (falling from 100 s to 85 s as data grows) against four comparison methods, yielding the 37 percent recovery-time reduction claimed for unknown fault scenarios.

Load-bearing premise

The 37 percent recovery-time advantage rests on the unstated premise that 'unknown fault scenarios' means the same thing for IFSHM and for the four baseline models, and that those models are implemented well enough for the comparison to be fair.

Editorial extensions

If this is right

  • If the 37 percent recovery-time reduction is real, cloud operators can use the same framework for failure modes they have not seen before, instead of writing a new rule or script for each incident.
  • The LLM's semantic state makes recovery decisions invariant to surface symptoms: the same policy would treat a CPU spike, a garbage-collection failure, memory overflow, and response timeout as manifestations of one root cause.
  • The memory-guided meta-controller provides a direct path for continuous adaptation: new fault types can be folded into the replay buffer and prompt template without retraining the whole policy from scratch.
  • On the reported benchmark, IFSHM dominates the four comparison methods on both detection accuracy and recovery time, which suggests the LLM's environment modeling actually improves DRL sample efficiency rather than just adding latency.
  • Hierarchical recovery actions (category plus parameters) would let operators audit and constrain what the agent is allowed to do, helping to make automated self-healing safe enough for production.

Reading between the lines

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

  • The 37 percent figure is computed against baseline implementations that are named but not specified in detail; a fair reproduction would need the exact architectures and hyperparameters of Deformable DETR, GCN-FR, TL-FD/FR, and SSL-AD to confirm the comparison.
  • If the LLM state vector is the real driver of generalization, then a cheaper distilled encoder or a smaller model should preserve most of the recovery-time gain; this is a testable way to isolate the LLM's contribution from the DRL trainer.
  • The paper's own limitations (offline LLM fine-tuning, centralized log access) imply that the framework's next test is in federated or privacy-preserving settings, where the semantic encoder must learn from distributed logs rather than one central stream.
  • The reported accuracy plateau of 0.89 to 0.92 might reflect a benchmark ceiling; injecting rarer or adversarial fault types would reveal whether the generalization claim holds beyond the OpenStack scenarios.
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. This manuscript proposes IFSHM, a two-stage fault self-healing framework for cloud AI systems that combines an LLM-based semantic interpretation module with a PPO-based DRL recovery policy, augmented by a memory-guided meta-controller and prompt tuning. The central claim, stated in the abstract and Section IV-B, is that IFSHM shortens system recovery time by 37% relative to existing DRL and rule methods under unknown fault scenarios, with fault detection accuracy of 0.89-0.92 and recovery time from 100s down to 85s on the DessertLab Failure-Dataset-OpenStack. The paper also presents the architecture equations (Eqs. 1-13) and reports comparisons against four baselines: Deformable DETR, GCN-FR, TL-FD/FR, and SSL-AD.

Significance. If the claimed 37% recovery-time improvement were substantiated with a well-defined held-out-fault protocol, the LLM-plus-DRL architecture would be a potentially useful contribution to cloud reliability. The paper's honest limitation section and the modular design (semantic encoding, hierarchical actions, meta-controller) are positive aspects. However, as presented, the numerical claim is not derivable from the reported data, 'unknown fault scenarios' is not operationalized, and the comparison to 'rule methods' is not actually evaluated. The contribution therefore reduces to an architectural proposal with an unverified headline result.

major comments (5)
  1. [Abstract and Section IV-B] The headline 37% recovery-time reduction is not supported by the numbers reported in Section IV-B. At the largest training size (5k), IFSHM's 85s is a 19% reduction from the best baseline (Deformable DETR at 105s) and about 25% from the average of all baselines; at 1k, 100s versus 120s is 17%. No pair or average reported in the text yields 37%, so the central quantitative claim in the abstract is not substantiated by the paper's own data.
  2. [Section IV-A and IV-B] The term 'unknown fault scenarios' is never defined. The experiments vary training data size from 1k to 5k, which measures learning-curve behavior, not generalization to held-out fault types; no fault-type split or injection protocol is described. The paper says '20 injected fault scenarios' are used for recovery evaluation, but it does not state whether these include fault types excluded from training. Thus the central claim of superior recovery on unknown faults is not testable from the manuscript.
  3. [Section IV-A] The comparison class is mismatched with the claim. The abstract refers to 'existing DRL and rule methods', but none of the four baselines (Deformable DETR, GCN-FR, TL-FD/FR, SSL-AD) is a rule-based method. Moreover, no implementation details, hyperparameters, or training protocols are provided for any baseline, which makes the comparison neither fair nor reproducible.
  4. [Section III, Eqs. (6)-(13)] The method description is largely generic equations with unspecified coefficients. The multi-task loss in Eq. (13) includes lambda_1 through lambda_4, the information-bottleneck term in Eq. (8) includes beta, the sampling probability in Eq. (10) includes epsilon, and the prompt template w in Eq. (11) is said to be learned but no training procedure or values are given. These free parameters are never reported, so the actual IFSHM algorithm is not fully specified and the equations do not establish a concrete connection to the implemented system.
  5. [Section IV-B] The paper states that each experiment was repeated five times with random seeds to ensure statistical significance, but only point values are reported (e.g., 85s versus 105s) without error bars, standard deviations, or significance tests. The text mentions 'error interval' in Figure 2, but its numerical values are absent. This makes it impossible to assess whether the reported improvements are statistically meaningful.
minor comments (5)
  1. [Section I] The Introduction contains an incomplete sentence: 'Various studies have been done t' immediately before the paragraph break.
  2. [Index Terms] The index term 'Strengthen policy optimization' should be 'Proximal Policy Optimization' or similar.
  3. [Figures 2 and 3] Figures 2 and 3 are referenced but not included in the text; axis labels, units, and the definition of the error bars should be provided.
  4. [Section III-A] Equation formatting is sloppy (e.g., Equation 1 uses mismatched display brackets), and variables nL, nM, and nA are introduced without defining the dimension conventions.
  5. [References] The reference list contains an unusually large number of self-citations and loosely related citations, such as [39]-[44] on multi-agent mechanisms, which are cited only in the future-work sentence.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: the IFSHM equations are standard RL/LLM objectives, and the headline 37% recovery-time improvement is an externally measured experimental outcome, not a quantity fitted by construction or defined in terms of the method's own outputs.

full rationale

The paper's derivation chain is Eq. (1)-(13): multimodal input encoding, LLM semantic state, hierarchical policy, PPO objective, information bottleneck, replay buffer, prompt loss, and a weighted multi-task objective. None of these equations defines the reported recovery time or detection accuracy in terms of itself; the benchmark claims in Section IV-B are measured outcomes on the external DessertLab 'Failure-Dataset-OpenStack' under four named baselines. The only arguable circularity-adjacent feature is a cluster of self-citations ([1], [2], [10], [11]) by the authors, but these are used for general background motivation (e.g., federated learning frameworks and LLM fault tolerance) and are not load-bearing premises from which the 37% result is derived. The abstract's '37%' is not reproducible from the numbers given in Section IV-B (85s versus 105-125s gives roughly 19-25%, not 37%), and 'unknown fault scenarios' is not formally operationalized; these are correctness and verifiability defects, not circularity. Because the central claim is an empirical comparison rather than a self-referential derivation, there is no specific reduction of an equation to itself by construction, and no fitted parameter is relabeled as a prediction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or formal entities; its 'meta-controller' is a renamed combination of replay buffer and prompt tuning. The central claim depends mostly on untested assumptions about the dataset, the MDP formulation, and the sufficiency of the LLM encoder.

free parameters (5)
  • lambda1, lambda2, lambda3, lambda4 (loss weights) = not reported
    In Eq. 13, Ltotal = LPPO + lambda1*LV + lambda2*LIB + lambda3*LPrompt + lambda4*LCluster; weights are not specified, so the multi-task balance is a free choice affecting behavior.
  • beta (KL penalty) = not reported
    Eq. 8 defines LIB = beta * DKL; beta controls information bottleneck strength; value not given.
  • epsilon in sampling probability = not reported
    Eq. 10 Pi proportional to |delta_i| + epsilon; epsilon is an arbitrary small constant.
  • Prompt soft template w = learned
    Eq. 11 uses a gradient-learned prompt; no initialization or regularization described, making the semantic interpretation dependent on this learned vector.
  • LoRA rank/adaptation hyperparameters = not reported
    The 7B LLM is fine-tuned with low-rank adaptation, but rank and target modules are not stated.
assumptions (4)
  • standard math PPO objective and GAE are valid for this MDP
    Eqs. 6-7 assume the standard clipped surrogate objective; the paper does not establish that the cloud recovery problem is a well-posed MDP with stationary dynamics.
  • domain assumption LLM semantic encoder maps logs/metrics to a latent state sufficient for recovery
    Eqs. 2-3 assume the concatenated encodings capture fault semantics; no evidence is given that the 7B model with LoRA can generalize to unseen fault types.
  • domain assumption Offline fine-tuning and centralized log access
    Section V states these as limitations, but the whole experimental protocol depends on them.
  • domain assumption The fault injection dataset is representative of real cloud AI failures
    Section IV-A introduces the dataset but gives no workload description or distribution of fault types.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Intelligent Fault Self-Healing Mechanism for Cloud AI Systems via Integration of Large Language Models and Deep Reinforcement Learning." pith.science (2026). https://pith.science/paper/SRQXEJZJ

@misc{pith2026250607411,
  author       = {Pith},
  title        = {Pith review of: An Intelligent Fault Self-Healing Mechanism for Cloud AI Systems via Integration of Large Language Models and Deep Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SRQXEJZJ}},
  note         = {Machine review of arXiv:2506.07411}
}
read the original abstract

As the scale and complexity of cloud-based AI systems continue to increase, the detection and adaptive recovery of system faults have become the core challenges to ensure service reliability and continuity. In this paper, we propose an Intelligent Fault Self-Healing Mechanism (IFSHM) that integrates Large Language Model (LLM) and Deep Reinforcement Learning (DRL), aiming to realize a fault recovery framework with semantic understanding and policy optimization capabilities in cloud AI systems. On the basis of the traditional DRL-based control model, the proposed method constructs a two-stage hybrid architecture: (1) an LLM-driven fault semantic interpretation module, which can dynamically extract deep contextual semantics from multi-source logs and system indicators to accurately identify potential fault modes; (2) DRL recovery strategy optimizer, based on reinforcement learning, learns the dynamic matching of fault types and response behaviors in the cloud environment. The innovation of this method lies in the introduction of LLM for environment modeling and action space abstraction, which greatly improves the exploration efficiency and generalization ability of reinforcement learning. At the same time, a memory-guided meta-controller is introduced, combined with reinforcement learning playback and LLM prompt fine-tuning strategy, to achieve continuous adaptation to new failure modes and avoid catastrophic forgetting. Experimental results on the cloud fault injection platform show that compared with the existing DRL and rule methods, the IFSHM framework shortens the system recovery time by 37% with unknown fault scenarios.

Figures

Figures reproduced from arXiv: 2506.07411 by the authors.

Figure 1
Figure 1. Intelligent Fault Self-Healing Mechanism (IFSHM) [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Fault Detection Accuracy Comparison [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Recovery Time Comparison recovery time of Deformable DETR-FD, GCN-FR, and TL￾FD/FR decreased from 120 s, 130 s, and 125 s to 105 s, 115 s, and 110 s, respectively, followed by performance. SSL-AD was the weakest, with recovery times dropping from 140 s to 125 s, but still higher than other methods. V. CONCLUSION In conclusion, we propose IFSHM, a unified self-healing framework for cloud AI systems that integrates LL… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 13 canonical work pages

  1. [1]

    Advances in appfl: A comprehensive and extensible federated learning framework,

    Z. Li, S. He, Z. Yang, M. Ryu, K. Kim, and R. Madduri, “Advances in appfl: A comprehensive and extensible federated learning framework,” arXiv preprint arXiv:2409.11585, 2024

    arXiv 2024

  2. [2]

    Research on large language model cross-cloud privacy protection and collaborative training based on federated learning,

    Z. Yang, Y . Jin, Y . Zhang, J. Liu, and X. Xu, “Research on large language model cross-cloud privacy protection and collaborative training based on federated learning,”arXiv preprint arXiv:2503.12226, 2025

    arXiv 2025

  3. [3]

    Ai-driven fault detection and predictive maintenance in elec- trical power systems: A systematic review of data-driven approaches, digital twins, and self-healing grids,

    S. Rana, “Ai-driven fault detection and predictive maintenance in elec- trical power systems: A systematic review of data-driven approaches, digital twins, and self-healing grids,”American Journal of Advanced Technology and Engineering Solutions, vol. 1, no. 01, pp. 258–289, 2025

    work page 2025

  4. [4]

    T-GAE: Transferable Graph Autoencoder for Network Alignment,

    J. He, C. I. Kanatsoulis, and A. Ribeiro, “T-GAE: Transferable Graph Autoencoder for Network Alignment,”arXiv e-prints, p. arXiv:2310.03272, Oct. 2023

    arXiv 2023

  5. [5]

    Graphsnapshot: Graph machine learning acceleration with fast storage and retrieval,

    D. Liu, R. Waleffe, M. Jiang, and S. Venkataraman, “Graphsnapshot: Graph machine learning acceleration with fast storage and retrieval,” arXiv preprint arXiv:2406.17918, 2024

    arXiv 2024

  6. [6]

    Confidence trigger detection: Accelerating real-time tracking-by-detection systems,

    Z. Ding, Z. Lai, S. Li, P. Li, Q. Yang, and E. Wong, “Confidence trigger detection: Accelerating real-time tracking-by-detection systems,” in2024 5th International Conference on Electronic Communication and Artificial Intelligence (ICECAI). IEEE, 2024, pp. 587–592

    work page 2024

  7. [7]

    Deception detection from linguistic and physiological data streams using bimodal convolutional neural networks,

    P. Li, M. Abouelenien, R. Mihalcea, Z. Ding, Q. Yang, and Y . Zhou, “Deception detection from linguistic and physiological data streams using bimodal convolutional neural networks,” in2024 5th International Conference on Information Science, Parallel and Distributed Systems (ISPDS). IEEE, 2024, pp. 263–267

    2024

  8. [8]

    Assertion detection in clinical natural language processing using large language models,

    Y . Ji, Z. Yu, and Y . Wang, “Assertion detection in clinical natural language processing using large language models,” in2024 IEEE 12th International Conference on Healthcare Informatics (ICHI), 2024, pp. 242–247

    2024

Show all 44 references
  1. [9]

    Optimization and prediction tech- niques for self-healing and self-learning applications in a trustworthy cloud continuum,

    J. Alonso, L. Orue-Echevarria, E. Osaba, J. L ´opez Lobo, I. Martinez, J. Diaz de Arcaya, and I. Etxaniz, “Optimization and prediction tech- niques for self-healing and self-learning applications in a trustworthy cloud continuum,”Information, vol. 12, no. 8, p. 308, 2021

    2021

  2. [10]

    Hades: Hardware accelerated decoding for efficient speculation in large language models,

    Z. Yang, Y . Jin, and X. Xu, “Hades: Hardware accelerated decoding for efficient speculation in large language models,”arXiv preprint arXiv:2412.19925, 2024

    2024 arXiv

  3. [11]

    Adaptive fault tolerance mechanisms of large language models in cloud computing environ- ments,

    Y . Jin, Z. Yang, X. Xu, Y . Zhang, and S. Ji, “Adaptive fault tolerance mechanisms of large language models in cloud computing environ- ments,”arXiv preprint arXiv:2503.12228, 2025

    2025 arXiv

  4. [12]

    Exploring diverse methods in visual question answering,

    P. Li, Q. Yang, X. Geng, W. Zhou, Z. Ding, and Y . Nian, “Exploring diverse methods in visual question answering,” in2024 5th International Conference on Electronic Communication and Artificial Intelligence (ICECAI). IEEE, 2024, pp. 681–685

    2024

  5. [13]

    Contextual hourglass network for semantic segmentation of high resolution aerial imagery,

    P. Li, Y . Lin, and E. Schultz-Fellenz, “Contextual hourglass network for semantic segmentation of high resolution aerial imagery,” in2024 5th International Conference on Electronic Communication and Artificial Intelligence (ICECAI). IEEE, 2024, pp. 15–18

    2024

  6. [14]

    A comparative study on enhancing prediction in social network advertisement through data augmentation,

    Q. Yang, P. Li, X. Xu, Z. Ding, W. Zhou, and Y . Nian, “A comparative study on enhancing prediction in social network advertisement through data augmentation,” in2024 4th International Conference on Machine Learning and Intelligent Systems Engineering (MLISE). IEEE, 2024, pp. 214–218

    2024

  7. [15]

    A decision-language model (dlm) for dynamic restless multi-armed bandit tasks in public health,

    N. Behari, E. Zhang, Y . ZHAO, A. Taneja, D. M. Nagaraj, and M. Tambe, “A decision-language model (dlm) for dynamic restless multi-armed bandit tasks in public health,” inThe Thirty-eighth Annual Conference on Neural Information Processing Systems

  8. [16]

    Scalable neural network kernels,

    A. Sehanobish, K. M. Choromanski, Y . ZHAO, K. A. Dubey, and V . Likhosherstov, “Scalable neural network kernels,” inThe Twelfth International Conference on Learning Representations

  9. [17]

    Automated assessment in math education: A comparative analysis of llms for open-ended responses,

    S. Baral, E. Worden, W.-C. Lim, Z. Luo, C. Santorelli, and A. Gurung, “Automated assessment in math education: A comparative analysis of llms for open-ended responses,” 2024

    2024

  10. [18]

    Cloud-based ai systems: Leveraging large language models for intelligent fault detection and autonomous self-healing,

    C. Ji and H. Luo, “Cloud-based ai systems: Leveraging large language models for intelligent fault detection and autonomous self-healing,” arXiv preprint arXiv:2505.11743, 2025

    2025 arXiv

  11. [19]

    Give: Structured reasoning of large language models with knowledge graph inspired veracity extrapolation,

    J. He, M. D. Ma, J. Fan, D. Roth, W. Wang, and A. Ribeiro, “Give: Structured reasoning of large language models with knowledge graph inspired veracity extrapolation,” 2025. [Online]. Available: https://arxiv.org/abs/2410.08475

    2025 arXiv

  12. [20]

    A large-scale empirical study on large language models for election prediction,

    C. Yu, Z. Weng, Y . Li, Z. Li, X. Hu, and Y . Zhao, “A large-scale empirical study on large language models for election prediction,”arXiv preprint arXiv:2412.15291

    arXiv

  13. [21]

    Ai-powered self-healing systems for fault tolerant platform engineering: Case studies and challenges,

    M. J. Karamthulla, J. N. A. Malaiyappan, and S. Prakash, “Ai-powered self-healing systems for fault tolerant platform engineering: Case studies and challenges,”Journal of Knowledge Learning and Science Technol- ogy ISSN: 2959-6386 (online), vol. 2, no. 2, pp. 327–338, 2023

    2023

  14. [22]

    Ai-enhanced self-healing cloud architectures for data integrity, privacy, and sustainable learning,

    V . R. Vemula, “Ai-enhanced self-healing cloud architectures for data integrity, privacy, and sustainable learning,” inSmart Education and Sustainable Learning Environments in Smart Cities. IGI Global Scientific Publishing, 2025, pp. 93–106

    2025

  15. [23]

    Artificial intelligence for self- healing automation testing frameworks: Real-time fault prediction and recovery,

    P. Nama, P. Reddy, and S. K. Pattanayak, “Artificial intelligence for self- healing automation testing frameworks: Real-time fault prediction and recovery,”Artificial Intelligence, vol. 64, no. 3S, 2024

    2024

  16. [24]

    Integration of multi-agent systems and artificial intelligence in self-healing subway power supply systems: Advancements in fault diagnosis, isolation, and recovery,

    J. Feng, T. Yu, K. Zhang, and L. Cheng, “Integration of multi-agent systems and artificial intelligence in self-healing subway power supply systems: Advancements in fault diagnosis, isolation, and recovery,” Processes, vol. 13, no. 4, p. 1144, 2025

    2025

  17. [25]

    Artificial intelligence for fault detection in cloud- optimized data engineering systems,

    D. K. Pentyala, “Artificial intelligence for fault detection in cloud- optimized data engineering systems,”International Journal of Social Trends, vol. 2, no. 4, pp. 8–44, 2024

    2024

  18. [26]

    Automated intelligent healing in cloud- scale data centers,

    R. Li, Z. Cheng, P. P. Lee, P. Wang, Y . Qiang, L. Lan, C. He, J. Lu, M. Wang, and X. Ding, “Automated intelligent healing in cloud- scale data centers,” in2021 40th International Symposium on Reliable Distributed Systems (SRDS). IEEE, 2021, pp. 244–253

    2021

  19. [27]

    Time series modeling for heart rate prediction: From arima to transformers,

    H. Ni, S. Meng, X. Geng, P. Li, Z. Li, X. Chen, X. Wang, and S. Zhang, “Time series modeling for heart rate prediction: From arima to transformers,” in2024 6th International Conference on Electronic Engineering and Informatics (EEI). IEEE, 2024, pp. 584–589

    2024

  20. [28]

    RAG-RLRC-LaySum at BioLaySumm: Integrating retrieval-augmented generation and readability control for layman summarization of biomedical texts,

    Y . Ji, Z. Li, R. Meng, S. Sivarajkumar, Y . Wang, Z. Yu, H. Ji, Y . Han, H. Zeng, and D. He, “RAG-RLRC-LaySum at BioLaySumm: Integrating retrieval-augmented generation and readability control for layman summarization of biomedical texts,” inProceedings of the 23rd Workshop on...

    2024

  21. [29]

    Self-give: Associative thinking from limited structured knowledge for enhanced large language model reasoning,

    J. He, J. Fan, B. Jiang, I. Houine, D. Roth, and A. Ribeiro, “Self-give: Associative thinking from limited structured knowledge for enhanced large language model reasoning,” 2025. [Online]. Available: https://arxiv.org/abs/2505.15062

    2025

  22. [30]

    Curriculum guided reinforcement learning for efficient multi hop retrieval augmented generation,

    Y . Ji, R. Meng, Z. Li, and D. He, “Curriculum guided reinforcement learning for efficient multi hop retrieval augmented generation,”arXiv preprint arXiv:2505.17391, 2025

    2025 arXiv

  23. [31]

    Mt2st: Adaptive multi-task to single-task learning,

    D. Liu and Y . Yu, “Mt2st: Adaptive multi-task to single-task learning,” arXiv preprint arXiv:2406.18038, 2024

    2024 arXiv

  24. [32]

    The bandit whisperer: Communication learning for restless bandits,

    Y . Zhao, T. Wang, D. M. Nagaraj, A. Taneja, and M. Tambe, “The bandit whisperer: Communication learning for restless bandits,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 22, 2025, pp. 23 404–23 413

    2025

  25. [33]

    Bias evaluation and mitigation in retrieval-augmented medical question-answering systems,

    Y . Ji, H. Zhang, and Y . Wang, “Bias evaluation and mitigation in retrieval-augmented medical question-answering systems,”arXiv preprint arXiv:2503.15454, 2025

    2025 arXiv

  26. [34]

    Contemporary model compression on large language models inference,

    D. Liu, “Contemporary model compression on large language models inference,”arXiv preprint arXiv:2409.01990, 2024

    2024 arXiv

  27. [35]

    Llmeasyquant: Scalable quantization for parallel and distributed llm inference,

    D. Liu and Y . Yu, “Llmeasyquant: Scalable quantization for parallel and distributed llm inference,” inAI4X 2025 International Conference

    2025

  28. [36]

    Cross-cloud data privacy protection: Optimizing collaborative mechanisms of ai systems by integrating federated learning and llms,

    H. Luo and C. Ji, “Cross-cloud data privacy protection: Optimizing collaborative mechanisms of ai systems by integrating federated learning and llms,”arXiv preprint arXiv:2505.13292, 2025

    2025 arXiv

  29. [37]

    A stochastic gda method with backtracking for solving nonconvex (strongly) concave minimax problems,

    Q. Xu, X. Zhang, N. S. Aybat, and M. G ¨urb¨uzbalaban, “A stochastic gda method with backtracking for solving nonconvex (strongly) concave minimax problems,”arXiv preprint arXiv:2403.07806, 2024

    2024 arXiv

  30. [38]

    Agda+: Proximal alternating gradient descent ascent method with a nonmonotone adaptive step-size search for nonconvex minimax problems,

    X. Zhang, Q. Xu, and N. S. Aybat, “Agda+: Proximal alternating gradient descent ascent method with a nonmonotone adaptive step-size search for nonconvex minimax problems,”arXiv preprint arXiv:2406.14371, 2024

    2024 arXiv

  31. [39]

    Chateval: Towards better llm-based evaluators through multi- agent debate,

    C.-M. Chan, W. Chen, Y . Su, J. Yu, W. Xue, S. Zhang, J. Fu, and Z. Liu, “Chateval: Towards better llm-based evaluators through multi- agent debate,”arXiv preprint arXiv:2308.07201, 2023

    2023 arXiv

  32. [40]

    Multi-agent collaboration: Harnessing the power of intelligent llm agents,

    Y . Talebirad and A. Nadiri, “Multi-agent collaboration: Harnessing the power of intelligent llm agents,”arXiv preprint arXiv:2306.03314, 2023

    2023 arXiv

  33. [41]

    Composerx: Multi-agent symbolic music composition with llms,

    Q. Deng, Q. Yang, R. Yuan, Y . Huang, Y . Wang, X. Liu, Z. Tian, J. Pan, G. Zhang, H. Linet al., “Composerx: Multi-agent symbolic music composition with llms,”arXiv preprint arXiv:2404.18081, 2024

    2024 arXiv

  34. [42]

    From critique to clarity: A pathway to faithful and personalized code explanations with large language models,

    Z. Xu, Z. Luo, Y . Li, K. Lee, and S. R. Etesami, “From critique to clarity: A pathway to faithful and personalized code explanations with large language models,”arXiv preprint arXiv:2501.14731, 2024

    2024 arXiv

  35. [43]

    Towards foundation- model-based multiagent system to accelerate ai for social impact,

    Y . Zhao, N. Boehmer, A. Taneja, and M. Tambe, “Towards foundation- model-based multiagent system to accelerate ai for social impact,”arXiv preprint arXiv:2412.07880, 2024

    2024 arXiv

  36. [44]

    Faithfulpersona: Balancing faithfulness and personalization in code explanations through self-critique,

    Z. Luo, Y . Li, Z. Xu, K. Lee, and S. R. Etesami, “Faithfulpersona: Balancing faithfulness and personalization in code explanations through self-critique,” inFindings of the Association for Computational Linguis- tics: NAACL 2025, 2025, pp. 930–944

    2025

Pith tools

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