arxiv: 2604.23139 · v1 · submitted 2026-04-25 · 💻 cs.DC

Recognition: unknown

GreenDyGNN: Runtime-Adaptive Energy-Efficient Communication for Distributed GNN Training

Arefin Niam, M. S. Q. Zulkar Nine, Tevfik Kosar

Authors on Pith no claims yet

Pith reviewed 2026-05-08 07:18 UTC · model grok-4.3

classification 💻 cs.DC

keywords distributed GNN trainingenergy efficiencycache managementreinforcement learningDouble-DQNnetwork congestionruntime adaptationdistributed training

0 comments

The pith

GreenDyGNN adapts cache rebuild windows and per-owner allocations at runtime with a Double-DQN agent to cut energy from remote feature fetches in distributed GNN training.

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

Distributed GNN training incurs high energy costs from fine-grained remote feature fetches that cross partition boundaries and stall GPUs. Static caching and presampling cannot react when network congestion changes during an epoch, and can even raise energy use by up to 45 percent. GreenDyGNN recasts cache-window decisions as a sequential control problem solved by a Double-DQN agent trained in a simulator whose congestion is deliberately randomized. The agent chooses rebuild intervals and allocation sizes at each boundary; an asynchronous double-buffered pipeline hides the decision cost. Under congestion the policy lowers total energy by up to 43 percent versus default DGL and 4-24 percent versus the best fixed schedule, while staying near optimal when the network remains clean.

Core claim

GreenDyGNN formulates cache window management as a sequential decision problem solved by a Double-DQN agent that adapts rebuild window size and per-owner cache allocation at each boundary. An asynchronous double-buffered pipeline makes adaptation effectively free. Under time-varying congestion the method cuts total energy by up to 43 percent compared with default DGL and 4-24 percent compared with the best static policy, while remaining close to optimal when congestion is absent.

What carries the argument

Double-DQN agent trained in a calibrated simulator with domain-randomized congestion that selects rebuild window sizes and cache allocations at partition boundaries to minimize energy from RPC stalls.

If this is right

Training runs consume less power when network load fluctuates during an epoch.
Static presampling and caching become less necessary because adaptation handles variation.
GPU utilization improves because fewer remote fetches stall computation.
Clusters can run more training jobs per unit of energy when using the adaptive policy.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Similar reinforcement-learning controllers could manage other runtime resources such as batch sizes or communication compression in distributed training.
The approach may generalize to other sampling-based graph workloads beyond GNNs that cross partition boundaries.
Production clusters could monitor energy directly and fine-tune the reward signal without simulator retraining.

Load-bearing premise

The calibrated simulator with randomized congestion patterns produces policies that transfer to real hardware with only small loss in the energy savings.

What would settle it

Measure total energy and training time on a real multi-GPU cluster while injecting controlled network congestion and compare the observed savings against both the simulated prediction and the best static cache schedule.

Figures

Figures reproduced from arXiv: 2604.23139 by Arefin Niam, M. S. Q. Zulkar Nine, Tevfik Kosar.

**Figure 1.** Figure 1: Per-RPC energy decomposed into initiation (red) and view at source ↗

**Figure 2.** Figure 2: GreenDyGNN runtime training pipeline (single-worker view at source ↗

**Figure 3.** Figure 3: RL agent state inputs and action outputs. The agent view at source ↗

**Figure 4.** Figure 4: Total energy (GPU + CPU, all nodes) at B=2000 under congestion. Annotations show GreenDyGNN’s reduction relative to Default DGL. where T˜ recent is the median of the 30 most recent fetch times. If T˜ recent/Tˆ base ≤ 1.1, then ˆδ is clamped to zero. The estimate requires only O(1) arithmetic per decision and is passed directly to the RL state constructor. Algorithm 2 gives the complete per-boundary contro… view at source ↗

**Figure 6.** Figure 6: compares total energy at B=2000 without congestion. Under clean conditions, GreenDyGNN closely matches the strongest static baseline: the gap between GreenDyGNN and RapidGNN stays within 2% on every dataset (178.0 vs. 180.0 kJ on Products, 180.0 vs. 182.0 kJ on Reddit, 258.0 vs. 262.0 kJ on Papers100M). This is the expected behavior. GreenDyGNN is designed to adapt when network conditions change; when the… view at source ↗

**Figure 7.** Figure 7: RL agent behavior on OGBN-Papers100M. Top: re view at source ↗

**Figure 9.** Figure 9: Cumulative energy under congestion across all three view at source ↗

**Figure 10.** Figure 10: Accuracy vs. wall time under congestion across view at source ↗

**Figure 11.** Figure 11: Ablation study under congestion. Removing RL view at source ↗

read the original abstract

Distributed GNN training is dominated by remote feature fetching, which can be very costly. Multi-hop neighborhood sampling crosses partition boundaries and triggers fine-grained RPCs whose fixed initiation cost and GPU-stall latency waste energy. Prior systems try to reduce this overhead with presampling and static caching, but cache policies cannot react to runtime network variation. We show that under time-varying congestion, static caching can increase energy by up to 45% because a fixed rebuild schedule is insufficient. We present GreenDyGNN, which formulates cache window management as a sequential decision problem. GreenDyGNN performs intra-epoch cache rebuilds and uses a Double-DQN agent, trained in a calibrated simulator with domain-randomized congestion, to adapt rebuild window size and per-owner cache allocation at each boundary. An asynchronous double-buffered pipeline makes adaptation effectively free. Under congestion, GreenDyGNN cuts total energy by up to 43% over Default DGL and 4-24% over the best static policy, while closely matching the optimum under clean conditions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper applies Double-DQN to runtime cache decisions in distributed GNN training and reports energy savings from simulation, but lacks hardware validation for those gains.

read the letter

The paper's core contribution is using a Double-DQN agent to handle dynamic cache window sizing and per-owner allocation decisions during distributed GNN training. This lets the system react to changing network congestion instead of sticking with a fixed static policy. They train the agent inside a simulator that includes domain-randomized congestion, then deploy it with an asynchronous double-buffered pipeline that keeps the overhead low. The abstract reports up to 43% energy reduction versus default DGL under congestion and 4-24% better than the best static approach, while staying close to optimal when the network is clean. This approach makes sense for the problem. Static caching can indeed waste energy if congestion patterns shift, and turning cache management into an online decision process is a logical move. The use of Double-DQN to avoid overestimation in Q-values is a standard choice here. The main limitation is the lack of real hardware validation. All the quantitative results come from the simulator. If the simulator's models for RPC latency, GPU stalls, and congestion do not match actual Ethernet or InfiniBand behavior, the learned policy may not deliver the same savings on a real cluster. The abstract gives no cross-validation against measured traces or end-to-end energy numbers from physical runs. This work is aimed at researchers and engineers working on systems for large-scale graph neural networks, particularly those focused on energy efficiency and runtime adaptation. It offers a practical technique that could be built upon. I would recommend sending it to peer review. The idea is clearly presented and addresses a genuine pain point, even though the evaluation would benefit from additional real-system experiments.

Referee Report

2 major / 2 minor

Summary. The paper presents GreenDyGNN, a system for distributed GNN training that formulates cache window management as a sequential decision problem solved by a Double-DQN agent. The agent is trained in a calibrated simulator with domain-randomized congestion to dynamically adapt intra-epoch cache rebuild windows and per-owner allocations; an asynchronous double-buffered pipeline is used to make adaptation overhead negligible. The central claims are that static policies can increase energy by up to 45% under time-varying congestion, while GreenDyGNN reduces total energy by up to 43% versus Default DGL and 4-24% versus the best static policy under congestion, while matching the optimum under clean conditions.

Significance. If the simulator fidelity and policy transfer claims hold, the work would offer a practical runtime-adaptive approach to energy-efficient communication in distributed GNN training, addressing a gap where static caching fails under network variation. The use of RL with domain randomization and the double-buffered pipeline are technically interesting contributions that could influence future systems for large-scale graph workloads.

major comments (2)

[Abstract and Evaluation (simulator-trained policy results)] The headline quantitative claims (up to 43% energy reduction vs. Default DGL and 4-24% vs. best static policy) rest entirely on simulator results; the manuscript provides no hardware traces, no cross-validation of simulator RPC latency/GPU-stall/congestion outputs against measured production traces, and no end-to-end energy numbers from physical clusters, which directly undermines the transferability assertion in the abstract.
[Simulator calibration and policy transfer discussion] The weakest assumption—that the domain-randomized congestion model produces policies that remain near-optimal on real Ethernet/InfiniBand networks—is load-bearing for all reported savings; without at least one controlled hardware experiment or sensitivity analysis showing policy degradation under model mismatch, the 43% figure cannot be treated as a verified system result.

minor comments (2)

[Abstract] The abstract states energy savings but supplies no run counts, error bars, or precise baseline definitions; these details should be added to the evaluation section for reproducibility.
[System design and pipeline description] Clarify the exact definition of 'total energy' (e.g., whether it includes only communication or also GPU compute) and how the asynchronous pipeline overhead is measured and shown to be negligible.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive feedback and for recognizing the technical interest in the RL-based adaptive approach. We address the two major comments below regarding the simulator-based evaluation and policy transfer assumptions. We will incorporate revisions to strengthen the discussion of limitations and add further analysis.

read point-by-point responses

Referee: [Abstract and Evaluation (simulator-trained policy results)] The headline quantitative claims (up to 43% energy reduction vs. Default DGL and 4-24% vs. best static policy) rest entirely on simulator results; the manuscript provides no hardware traces, no cross-validation of simulator RPC latency/GPU-stall/congestion outputs against measured production traces, and no end-to-end energy numbers from physical clusters, which directly undermines the transferability assertion in the abstract.

Authors: We acknowledge that all reported quantitative results, including the energy savings, are obtained from the calibrated simulator rather than physical hardware deployments. The simulator was calibrated using domain randomization over a range of congestion parameters to capture time-varying network behavior, and its latency and stall models were validated against synthetic microbenchmarks. We agree that the absence of hardware traces limits direct claims of transferability to production Ethernet/InfiniBand clusters. In the revised manuscript we will (1) expand the simulator calibration subsection with additional validation metrics, (2) add an explicit limitations paragraph on simulation-to-hardware gaps, and (3) temper the abstract and conclusion to frame the 43% figure as simulator-demonstrated potential rather than a verified hardware result. revision: partial
Referee: [Simulator calibration and policy transfer discussion] The weakest assumption—that the domain-randomized congestion model produces policies that remain near-optimal on real Ethernet/InfiniBand networks—is load-bearing for all reported savings; without at least one controlled hardware experiment or sensitivity analysis showing policy degradation under model mismatch, the 43% figure cannot be treated as a verified system result.

Authors: Domain randomization was chosen precisely to encourage robustness across congestion regimes that we expect to appear on real networks. To directly respond to the concern, the revision will include a new sensitivity-analysis subsection that (a) perturbs the congestion model parameters outside the training distribution and (b) quantifies the resulting degradation in policy performance within simulation. We do not have the resources or cluster access to run controlled hardware experiments for this revision cycle; therefore we will also revise the discussion to state that the reported savings are simulator-validated and that hardware transfer remains an open question for future work. revision: partial

standing simulated objections not resolved

We do not possess production hardware traces or end-to-end physical-cluster energy measurements and therefore cannot add them to the manuscript.

Circularity Check

0 steps flagged

No significant circularity in GreenDyGNN derivation chain

full rationale

The paper formulates cache-window management as an MDP and trains a Double-DQN policy inside a calibrated simulator with domain-randomized congestion; the reported energy reductions (43 % vs. Default DGL, 4-24 % vs. best static policy) are then obtained by executing that policy and the baselines inside the same simulator. These are straightforward empirical comparisons, not derivations that reduce to the inputs by construction. No self-definitional equations, fitted parameters renamed as predictions, load-bearing self-citations, or uniqueness theorems imported from prior author work appear in the abstract or described method. The central claim therefore remains an independent optimization result whose validity hinges on simulator fidelity rather than logical circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no explicit free parameters, axioms, or invented entities are stated beyond the standard Double-DQN algorithm and the assumption of a calibrated domain-randomized simulator.

pith-pipeline@v0.9.0 · 5494 in / 1168 out tokens · 68154 ms · 2026-05-08T07:18:34.318405+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

32 extracted references · 17 canonical work pages

[1]

Semi-supervised classification with graph convolutional networks,

T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” inProceedings of the 5th International Con- ference on Learning Representations (ICLR), 2017

2017
[2]

Inductive representation learning on large graphs,

W. L. Hamilton, R. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” inAdvances in Neural Information Processing Systems 30 (NeurIPS), 2017

2017
[3]

DistDGL: Distributed graph neural network training for billion-scale graphs,

D. Zheng, C. Ma, M. Wang, J. Zhou, Q. Su, X. Song, Q. Gan, Z. Zhang, and G. Karypis, “DistDGL: Distributed graph neural network training for billion-scale graphs,” in10th IEEE/ACM Workshop on Irregular Applications: Architectures and Algorithms (IA3), 2020, pp. 36–44, https://doi.org/10.1109/IA351965.2020.00011

work page doi:10.1109/ia351965.2020.00011 2020
[4]

P3: Distributed deep graph learning at scale,

S. Gandhi and A. P. Iyer, “P3: Distributed deep graph learning at scale,” in15th USENIX Symposium on Operating Systems Design and Implementation (OSDI), 2021, pp. 551–568

2021
[5]

Improving the accuracy, scalability, and performance of graph neural networks with Roc,

Z. Jia, S. Lin, M. Gao, M. Zaharia, and A. Aiken, “Improving the accuracy, scalability, and performance of graph neural networks with Roc,”Proceedings of Machine Learning and Systems, vol. 2, pp. 187– 198, 2020

2020
[6]

ByteGNN: Efficient graph neural network training at large scale,

C. Zheng, H. Chen, Y . Cheng, Z. Song, Y . Wu, C. Li, J. Cheng, H. Yang, and S. Zhang, “ByteGNN: Efficient graph neural network training at large scale,”Proceedings of the VLDB Endowment, vol. 15, no. 6, pp. 1228–1242, 2022, https://doi.org/10.14778/3514061.3514069

work page doi:10.14778/3514061.3514069 2022
[7]

Rapidgnn: Energy and communication-efficient distributed training on large-scale graph neural networks,

A. Niam, T. Kosar, and M. Nine, “Rapidgnn: Energy and communication-efficient distributed training on large-scale graph neural networks,”arXiv preprint arXiv:2509.05207, 2025

work page arXiv 2025
[8]

MassiveGNN: Efficient training via prefetching for massively connected distributed graphs,

A. Sarkar, S. Ghosh, N. R. Tallent, and A. Jannesari, “MassiveGNN: Efficient training via prefetching for massively connected distributed graphs,” in2024 IEEE International Conference on Cluster Comput- ing (CLUSTER), 2024, https://doi.org/10.1109/CLUSTER59578.2024. 00013

work page doi:10.1109/cluster59578.2024 2024
[9]

PaGraph: Scaling GNN training on large graphs via computation-aware caching,

Z. Lin, C. Li, Y . Miao, Y . Liu, and Y . Xu, “PaGraph: Scaling GNN training on large graphs via computation-aware caching,” inProceedings of the 11th ACM Symposium on Cloud Computing (SoCC), 2020, pp. 401–415, https://doi.org/10.1145/3419111.3421281

work page doi:10.1145/3419111.3421281 2020
[10]

Communication-efficient graph neural networks with probabilistic neighborhood expansion analysis and caching,

T. Kaler, A.-S. Iliopoulos, P. Murzynowski, T. B. Schardl, C. E. Leiserson, and J. Chen, “Communication-efficient graph neural networks with probabilistic neighborhood expansion analysis and caching,” in Proceedings of Machine Learning and Systems, vol. 5, 2023

2023
[11]

BGL: GPU-efficient GNN training by optimizing graph data I/O and preprocessing,

T. Liu, Y . Chen, D. Li, C. Wu, Y . Zhu, J. He, Y . Peng, H. Chen, H. Chen, and C. Guo, “BGL: GPU-efficient GNN training by optimizing graph data I/O and preprocessing,” in20th USENIX Symposium on Networked Systems Design and Implementation (NSDI), 2023, pp. 103–118

2023
[12]

Chameleon: a scalable production testbed for computer sci- ence research,

K. Keahey, P. Riteau, D. Stanzione, T. Cockerill, J. Mambretti, P. Rad, and P. Ruth, “Chameleon: a scalable production testbed for computer sci- ence research,” inContemporary High Performance Computing. CRC Press, 2019, pp. 123–148

2019
[13]

PyTorch: An imperative style, high- performance deep learning library,

A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Köpf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, “PyTorch: An imperative style, high- performance deep learning library,” inAdvances in Neural Information Processing S...

2019
[14]

BNS-GCN: Efficient full- graph training of graph convolutional networks with partition-parallelism and random boundary node sampling,

C. Wan, Y . Li, A. Li, N. S. Kim, and Y . Lin, “BNS-GCN: Efficient full- graph training of graph convolutional networks with partition-parallelism and random boundary node sampling,”Proceedings of Machine Learn- ing and Systems, vol. 4, pp. 673–693, 2022

2022
[15]

GNNLab: A factored system for sample-based GNN training over GPUs,

J. Yang, D. Tang, X. Song, L. Wang, Q. Yin, R. Chen, W. Yu, and J. Zhou, “GNNLab: A factored system for sample-based GNN training over GPUs,” inProceedings of the 17th European Conference on Computer Systems (EuroSys), 2022, pp. 417–434, https://doi.org/10. 1145/3492321.3519557

work page arXiv 2022
[16]

DUCATI: A dual-cache training system for graph neural networks on giant graphs with the GPU,

X. Zhang, Y . Shen, Y . Shao, and L. Chen, “DUCATI: A dual-cache training system for graph neural networks on giant graphs with the GPU,”Proceedings of the ACM on Management of Data, vol. 1, no. 2, pp. 166:1–166:24, 2023, https://doi.org/10.1145/3589311

work page doi:10.1145/3589311 2023
[17]

DGCL: An efficient communication library for distributed GNN training,

Z. Cai, X. Yan, Y . Wu, K. Ma, J. Cheng, and F. Yu, “DGCL: An efficient communication library for distributed GNN training,” inProceedings of the 16th European Conference on Computer Systems (EuroSys), 2021, pp. 130–144, https://doi.org/10.1145/3447786.3456233

work page doi:10.1145/3447786.3456233 2021
[18]

Strubell, A

E. Strubell, A. Ganesh, and A. McCallum, “Energy and policy consid- erations for deep learning in NLP,” inProceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), 2019, pp. 3645–3650, https://doi.org/10.18653/v1/P19-1355

work page doi:10.18653/v1/p19-1355 2019
[19]

Schwartz, J

R. Schwartz, J. Dodge, N. A. Smith, and O. Etzioni, “Green AI,” Communications of the ACM, vol. 63, no. 12, pp. 54–63, 2020, https: //doi.org/10.1145/3381831

work page doi:10.1145/3381831 2020
[20]

So, Maud Texier, and Jeff Dean

D. Patterson, J. Gonzalez, U. Hölzle, Q. Le, C. Liang, L.-M. Munguia, D. Rothchild, D. So, and M. Texier, “The carbon footprint of machine learning training will plateau, then shrink,”IEEE Computer, vol. 55, no. 7, 2022, https://doi.org/10.1109/MC.2022.3148714

work page doi:10.1109/mc.2022.3148714 2022
[21]

Zeus: Understanding and optimizing GPU energy consumption of DNN training,

J. You, J.-W. Chung, and M. Chowdhury, “Zeus: Understanding and optimizing GPU energy consumption of DNN training,” in20th USENIX Symposium on Networked Systems Design and Implementation (NSDI), 2023, pp. 119–139

2023
[22]

Reducing energy bloat in large model training,

J.-W. Chung, Y . Gu, I. Jang, L. Meng, N. Bansal, and M. Chowdhury, “Reducing energy bloat in large model training,” inACM SIGOPS 30th Symposium on Operating Systems Principles (SOSP), 2024, https://doi. org/10.1145/3694715.3695970

work page doi:10.1145/3694715.3695970 2024
[23]

EnvPipe: Performance- preserving DNN training framework for saving energy,

S. Choi, I. Koo, J. Ahn, M. Jeon, and Y . Kwon, “EnvPipe: Performance- preserving DNN training framework for saving energy,” in2023 USENIX Annual Technical Conference (USENIX ATC), 2023, pp. 851–864

2023
[24]

In: ACM SIGCOMM

H. Mao, R. Netravali, and M. Alizadeh, “Neural adaptive video stream- ing with Pensieve,” inProceedings of the ACM Special Interest Group on Data Communication (SIGCOMM), 2017, https://doi.org/10.1145/ 3098822.3098843

work page arXiv 2017
[25]

HPCC: high precision congestion control,

H. Mao, M. Schwarzkopf, S. B. Venkatakrishnan, Z. Meng, and M. Al- izadeh, “Learning scheduling algorithms for data processing clusters,” in Proceedings of the ACM Special Interest Group on Data Communication (SIGCOMM), 2019, https://doi.org/10.1145/3341302.3342080

work page doi:10.1145/3341302.3342080 2019
[26]

Device placement optimization with reinforcement learning,

A. Mirhoseini, H. Pham, Q. V . Le, B. Steiner, R. Larsen, Y . Zhou, N. Kumar, M. Norouzi, S. Bengio, and J. Dean, “Device placement optimization with reinforcement learning,” inProceedings of the 34th International Conference on Machine Learning (ICML), 2017, pp. 2430– 2439

2017
[27]

Domain randomization for transferring deep neural networks from simulation to the real world,

J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel, “Domain randomization for transferring deep neural networks from simulation to the real world,” in2017 IEEE/RSJ International Con- ference on Intelligent Robots and Systems (IROS), 2017, pp. 23–30, https://doi.org/10.1109/IROS.2017.8202133

work page doi:10.1109/iros.2017.8202133 2017
[28]

Learning dexterous in-hand manipulation,

OpenAI, M. Andrychowicz, B. Baker, M. Chociej, R. Jozefowicz, B. McGrew, J. Pachocki, A. Petron, M. Plappert, G. Powell, A. Ray, J. Schneider, S. Sidor, J. Tobin, P. Welinder, L. Weng, and W. Zaremba, “Learning dexterous in-hand manipulation,”The International Journal of Robotics Research, vol. 39, no. 1, pp. 3–20, 2020, https://doi.org/10. 1177/0278364919887447

2020
[29]

A fast and high quality multilevel scheme for partitioning irregular graphs,

G. Karypis and V . Kumar, “A fast and high quality multilevel scheme for partitioning irregular graphs,”SIAM Journal on Scientific Com- puting, vol. 20, no. 1, pp. 359–392, 1998, https://doi.org/10.1137/ S1064827595287997

1998
[30]

doi:10.1609/aaai.v30i1.10295

H. van Hasselt, A. Guez, and D. Silver, “Deep reinforcement learning with double Q-learning,” inProceedings of the 30th AAAI Conference on Artificial Intelligence (AAAI), 2016, pp. 2094–2100, https://doi.org/ 10.1609/aaai.v30i1.10295

work page doi:10.1609/aaai.v30i1.10295 2016
[31]

Deep graph library: A graph-centric, highly-performant package for graph neural networks,

M. Wang, D. Zheng, Z. Ye, Q. Gan, M. Li, X. Song, J. Zhou, C. Ma, L. Yu, Y . Gai, T. Xiao, T. He, G. Karypis, J. Li, and Z. Zhang, “Deep graph library: A graph-centric, highly-performant package for graph neural networks,”arXiv preprint arXiv:1909.01315, 2019

work page arXiv 1909
[32]

Anthropic, “Claude,” https://www.anthropic.com/claude, 2024, aI assis- tant used for grammar and readability improvements

2024