Pith. sign in

REVIEW 3 major objections 5 minor 37 references

Harnessing Generative Pre-Trained Transformer for Datacenter Packet Trace Generation

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read DTG-GPT: a GPT model that learns a datacenter's traffic pattern from source-destination IDs alone and writes new traces that match it.

desk verdict A credible preliminary report on GPT-based datacenter trace generation, but the evaluation needs Markov baselines and code before the central claim can be taken as established. read the letter →

arxiv 2501.12033 v1 pith:LLJWKJJR submitted 2025-01-21 cs.NI cs.AI

Chen Griner This is my paper · ORCID
classification cs.NIcs.AI
keywords datacentertrafficgenerationGPTnetworktracesynthesismatrixcomplexityburstinessn-gramnoveltyprivacy-preservingrelease
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 a transformer model of the kind used for language generation can learn the traffic pattern of a datacenter from nothing more than the ordered list of source and destination node IDs, and can then write new traces that resemble the original. The proposed model, DTG-GPT, treats a trace as a token sequence and adds a metadata embedding that tells the model which header field, which segment of the trace, and which trace each token belongs to. The paper evaluates fidelity with traffic matrices, temporal and non-temporal trace complexity, burst statistics, and n-gram novelty on seven traces from HPC and datacenter domains, and reports that generated traces match the spatial pattern, complexity profile, and burstiness of the originals while producing largely novel n-grams. The practical hope is that operators could one day release trained models rather than raw private traces, giving researchers access to realistic traffic without exposing the traces themselves.

What carries the argument

The load-bearing machinery is DTG-GPT's tokenization plus its meta-data embedding. Each source or destination ID becomes a token, and a second embedding adds three pieces of side information per token: a field vector marking whether the token is a source or destination, a segment vector marking which of $s=24$ equal-length blocks of the trace the token falls in, and a trace-ID vector marking which of the seven training traces it came from; these are combined into a single metadata token via $\mathrm{TOK}^m_i = \mathrm{TOK}^s_i + \mathrm{TOK}^f_i \cdot s + \mathrm{TOK}^t_i \cdot f \cdot s$. The field vector teaches the model the deterministic source-then-destination header order, the segment vector gives it a notion of where in time a token sits, and the trace-ID vector lets the same model serve multiple traces; output masking confines generation to the node IDs of the target network scale.

What would settle it

Retain timestamps from a real trace, train DTG-GPT on the same source-destination ID sequence, generate a trace, and compare timestamp-derived statistics such as the inter-arrival-time distribution or the number of microbursts in fixed 10 µs bins. If the generated trace matches the original's traffic matrix but its bursty short-timescale statistics are indistinguishable from those of a random permutation of the original's ID sequence, the claim that it mimics temporal structure is falsified.

Watch

Extended reading notes

Core claim

The central claim is that DTG-GPT — a GPT-2-sized transformer adapted with a metadata embedding layer — can generate novel packet traces whose spatiotemporal behavior mirrors a real trace even though its input omits timestamps, ports, flow sizes, and protocols. Trained on sequences of source and destination IDs from HPC and datacenter rack traces, the model reproduces the visible diagonal structure of traffic matrices, lands close to the original on a complexity map (temporal and non-temporal), and follows the mean-burst and peak-to-average burst profiles over windows of 20k requests. The generated traces are mostly novel: beyond n-grams of length roughly five to seven, nearly all sampled n-grams in the generated trace do not appear in the original, with the CNS trace as a partial exception. The paper also shows that masking the output distribution lets the same model generate a 512-node trace that mimics the pattern of a 1024-node trace, at least for the tested MOCFE case.

Load-bearing premise

The paper's result depends on the assumption that the spatiotemporal structure worth preserving in a datacenter trace is fully carried by the order of source and destination node IDs, since timestamps, ports, flow sizes, and protocol fields are dropped before training and evaluation.

Editorial extensions

If this is right

  • If DTG-GPT is right, datacenter operators can publish a trained model (the weights) instead of the trace itself, and researchers can draw arbitrarily long synthetic traces from it.
  • Because the model already respects the deterministic field order of source and destination, adding ports, MAC addresses, or protocol fields only requires enlarging the field vector and retraining, not a new architecture.
  • Temperature is an effective fidelity knob: lower temperatures push generated traces toward lower complexity and higher burstiness, higher temperatures do the reverse, so a trace's best-matching temperature can be chosen per source trace.
  • The masking-based scale transfer suggests one model could generate traffic for a network of a different size than any trace it saw, useful for testing demand-aware designs at unobserved scales.

Reading between the lines

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

  • Inference: if the fidelity claim is assumed, the natural next step is to use the same architecture on traces that keep timestamps as a separate token field; the paper's own discussion suggests time fields are the hardest to mimic, so inter-arrival distributions would be the first place to stress-test the model.
  • Inference: because the model copies at least a few long n-grams from the original trace, releasing trained weights is not automatically privacy-safe; a motivated party could plausibly extract rare long sequences, so the privacy argument would need a membership-inference check.
  • Inference: the smaller-scale result hints that the model learns a reusable notion of 'diagonal communication pattern' rather than memorizing node identities, which, if confirmed across more traces, would make generation for unseen topologies far more useful.
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

3 major / 5 minor

Summary. The paper introduces DTG-GPT, a GPT-2-style transformer model that generates synthetic datacenter packet traces by modeling the sequence of source and destination node IDs. The model augments the standard token and position embeddings with a meta-data embedding that encodes the packet-header field (source or destination), a segment index, and a trace ID. The authors train on seven publicly available traces from Facebook and HPC workloads and evaluate fidelity using traffic matrices, trace complexity, burst statistics (mean burst size and peak-to-average ratio), and n-gram novelty. They also present a preliminary experiment on generating traces for a smaller network scale via output masking. The central claim is that DTG-GPT can synthesize novel traces that mimic the spatiotemporal patterns of real traces and that this approach could eventually allow operators to release traffic information as trained model weights.

Significance. The paper addresses a real and important problem: the scarcity of realistic datacenter traffic traces for network research. The proposed architecture is simple and the evaluation on seven diverse public traces is a reasonable starting point. The n-gram novelty analysis and the smaller-scale generation experiment are useful contributions. However, the central empirical claim that the GPT architecture and its context window are responsible for the observed fidelity is not yet established, because the fidelity metrics are aggregate statistics that a low-order Markov baseline could plausibly match. In addition, the abstract's 'spatiotemporal patterns' claim is stronger than what the source-destination ID sequence alone can support. If the authors add baseline comparisons and clarify the temperature-selection protocol, the work would be a valuable preliminary contribution to synthetic traffic generation.

major comments (3)
  1. [Section V (Evaluation), subsections V-A to V-D] The four fidelity metrics—traffic matrices (V-A), trace complexity (V-B), burst statistics (V-C), and n-gram novelty (V-D)—are all aggregate or local-statistics functions of the flattened source-destination ID sequence. A first- or second-order Markov chain trained on the same token sequence can, in principle, reproduce the pair-frequency matrix, match the entropy-rate estimates used by trace complexity, and generate the same local repeat/burst statistics, because these metrics depend on short-range transition frequencies. The paper dismisses the trivial empirical-pair sampler in Section I but reports no comparison against a Markov baseline. Without such baselines, the evidence is consistent with the hypothesis that the model has learned little beyond pairwise transition frequencies, and the claim that the GPT architecture, context window, and meta-data embedding are necessary for the reported fidelity is unsupported.
  2. [Section V-B, Figure 7] The temperature is tuned per trace to minimize the Euclidean distance in the complexity map, and Figure 7 shows that the optimal temperature varies by trace (0.95 for CNS, 1.1 for MOCFE, and potentially outside the tested range for DB and Hadoop). The paper does not state which temperature was used for the fidelity results in Figures 3, 5, 8, 9, and 11. If these results were obtained at the default t=1, the complexity-based evaluation understates the model's best-case performance; if they were obtained at per-trace optima, the evaluation is not a fixed-procedure comparison. The authors should report the temperature for every result and provide a protocol for selecting the temperature without access to the target trace, which is essential for a claimed generation method.
  3. [Section II-A and Abstract] The trace model deliberately includes only source and destination node IDs, dropping timestamps, ports, flow sizes, and protocol fields. The abstract and conclusion claim that DTG-GPT generates traces that mimic 'spatiotemporal patterns,' but temporal structure such as microbursts, flow durations, and inter-arrival times is not encoded in the ID sequence. If such temporal information cannot be inferred from the ID sequence alone, the model cannot reproduce it, and the fidelity claim is only valid for the spatiotemporal structure observable in the ID sequence. The paper should state this limitation explicitly in the abstract or introduction and temper the 'spatiotemporal' claim accordingly.
minor comments (5)
  1. [Section V-C, Equation (3)] Equation (3) defines MB(W) as the sum over all requests in W of 1/w, which evaluates to 1 for any window of size w. The accompanying text correctly describes the mean burst size as the number of unique requests in W divided by w, so the formula should sum over unique source-destination pairs, not over all occurrences. This appears to be a typo, but the formal definition should be corrected.
  2. [Section III, paragraph on context length] The sentence 'we, therefore, would this value to be as large as possible' is missing a verb (likely 'want' or 'like'). Please rephrase for clarity.
  3. [Section V-C, last paragraph] The phrase 'beyond this parliamentary work's scope' appears to be a typographical error for 'preliminary work.'
  4. [Figure 4 caption] The caption says 'Original WEB trace matrix,' but the surrounding text (Section V-A) describes the figure as showing the DB trace. Please reconcile the caption with the text.
  5. [Section I, paragraph 1] The sentence 'However, these tracks are usually available only to network operators' should read 'these traces' instead of 'tracks.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the fidelity claims are empirical and the only self-citation is an external metric, not a derivation input.

full rationale

The paper's central claim is empirical: DTG-GPT, trained with a standard next-token cross-entropy objective on source/destination ID sequences, generates traces that are then compared to the original traces using four measures. None of these measures is used as a training loss, so the reported fidelity is not an input to the generation process. The trace-complexity tool is taken from [8], a prior paper co-authored by the present author, but it is used as an externally defined measurement framework with stated assumptions, not as a premise that encodes the target result that DTG-GPT is faithful. The traffic-matrix check is explicitly acknowledged as reproducible by simple pair-frequency sampling and is offered only as a basic validation, not as the core evidence. The temperature sweeps in Section V-B and V-C are presented openly as fine-tuning exploration, with per-temperature distances reported; the paper does not hide a fitted parameter as a prediction. No uniqueness theorem or ansatz is imported from the authors' prior work, and no equation defines a prediction in terms of the same quantity used as its input. The absence of Markov-baseline comparisons weakens the strength of the GPT-specific claim, but that is a correctness and robustness concern, not a circularity concern. The discarded timestamp and flow-size information is a modeling limitation, not a circular step. Overall, the derivation chain does not reduce to its inputs by construction.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central empirical claim rests on a handful of modeling choices: representing traffic as a flat source-destination ID sequence, using the GPT-2 architecture, using trace complexity from the author's earlier paper as the main fidelity metric, and tuning temperature per trace. No new physical entities are introduced. The main free parameters are the temperature and the segment count; the temperature is fitted to the target trace during evaluation.

free parameters (2)
  • temperature t = 0.9, 0.95, 1.0, 1.1, 1.2 (best per trace: CNS 0.95, MOCFE 1.1, DB/Hadoop >1.2)
    Chosen per trace to minimize the Euclidean distance between generated and original traces in the complexity map (Section V-B, Figure 7).
  • number of segments s = 24
    Hand-chosen to divide each trace into segments for the meta-data embedding (Section IV-D). The segment length is floor(n/s).
assumptions (5)
  • domain assumption The transformer/GPT-2 architecture from [18] and [25] works as specified for next-token prediction on token sequences.
    The model is built directly on the GPT-2 implementation (Section III).
  • domain assumption Trace complexity, as defined in [8], is a valid and sufficient measure of temporal and non-temporal structure in packet traces.
    The fidelity evaluation treats similar temporal and non-temporal complexity values as evidence that the generated trace resembles the original (Section II-B, Section V-B).
  • domain assumption The sequence of source and destination IDs fully captures the spatiotemporal traffic patterns of interest.
    The paper drops all other packet header fields and evaluates fidelity only on the ID sequence (Section II-A).
  • domain assumption Traces used for training are stationary enough that a model trained on one trace can generate novel sequences from the same distribution.
    The model is trained on random segments from the trace and then generates a new trace of similar length (Section IV-D, IV-E).
  • standard math Softmax with temperature as defined in Equation 2 controls output randomness as intended.
    Used to adjust the creativity of generation (Section IV-B).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Harnessing Generative Pre-Trained Transformer for Datacenter Packet Trace Generation." pith.science (2026). https://pith.science/paper/LLJWKJJR

@misc{pith2026250112033,
  author       = {Pith},
  title        = {Pith review of: Harnessing Generative Pre-Trained Transformer for Datacenter Packet Trace Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LLJWKJJR}},
  note         = {Machine review of arXiv:2501.12033}
}
read the original abstract

Today, the rapid growth of applications reliant on datacenters calls for new advancements to meet the increasing traffic and computational demands. Traffic traces from datacenters are essential for further development and optimization of future datacenters. However, traces are rarely released to the public. Researchers often use simplified mathematical models that lack the depth needed to recreate intricate traffic patterns and, thus, miss optimization opportunities found in realistic traffic. In this preliminary work, we introduce DTG-GPT, a packet-level Datacenter Traffic Generator (DTG), based on the generative pre-trained transformer (GPT) architecture used by many state-of-the-art large language models. We train our model on a small set of available traffic traces from different domains and offer a simple methodology to evaluate the fidelity of the generated traces to their original counterparts. We show that DTG-GPT can synthesize novel traces that mimic the spatiotemporal patterns found in real traffic traces. We further demonstrate that DTG-GPT can generate traces for networks of different scales while maintaining fidelity. Our findings indicate the potential that, in the future, similar models to DTG-GPT will allow datacenter operators to release traffic information to the research community via trained GPT models.

Figures

Figures reproduced from arXiv: 2501.12033 by the authors.

Figure 1
Figure 1. Two traffic matrices of 8M requests. Where (a) is a real trace from an HPC cluster (NeckBone) [21] and (b) is a counterpart trace generated with DTG-GPT. The maximal element was clipped to 100 requests to enhance contrast. at traffic matrices. These matrices describe the accumulated traffic between each source and destination in the network, as seen in the trace. This allows us to observe the spatial structure of th… view at source ↗
Figure 2
Figure 2. Simplified scheme for our DTG-GPT model . [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Traffic matrices for several of the communication traces. Colors are scaled individually, and the scale is provided at [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Original WEB trace traffic matrix, along with generated [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: The complexity map of seven real traces and their generated counterparts. An opaque circle designates original ( [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: The Euclidean distance (in the complexity map) of [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: The mean burst size for every overlapping window of size [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: The burst PAR for every overlapping window of size [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 11
Figure 11. Figure 11: Proportion of n-grams that are novel against n-gram [PITH_FULL_IMAGE:figures/full_fig_p010_11.png]
Figure 12
Figure 12. Figure 12: Proportion of n-grams that are unique against n-gram [PITH_FULL_IMAGE:figures/full_fig_p011_12.png]
Figure 13
Figure 13. Figure 13: shows the results of our test on two matrices tests in this paper, traffic matrix similarity and trace complexity. 13 (b) shows the traffic matrix for the smaller-scale generated trace, and [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 24 canonical work pages

  1. [1]

    Unleashing ai data center growth through optics,

    E. Timurdogan, “Unleashing ai data center growth through optics,” https://www.laserfocusworld.com/optics/article/14300952/ unleashing-ai-data-center-growth-through-optics

    work page arXiv

  2. [2]

    Infrastructure for large scale ai:

    A. Bjorlin, “Infrastructure for large scale ai: ”empowering open”,” in Open Compute Project , San Jose, CA, October 2022

    work page 2022

  3. [3]

    A survey of reconfigurable optical networks,

    M. N. Hall, K.-T. Foerster, S. Schmid, and R. Durairajan, “A survey of reconfigurable optical networks,” Optical Switching and Networking , vol. 41, p. 100621, 2021

    work page 2021

  4. [4]

    Toward demand-aware networking: A theory for self-adjusting networks,

    C. Avin and S. Schmid, “Toward demand-aware networking: A theory for self-adjusting networks,” ACM SIGCOMM Computer Communica- tion Review, vol. 48, no. 5, pp. 31–40, 2019

    work page 2019

  5. [5]

    Network traffic generation: A survey and methodology,

    O. A. Adeleke, N. Bastin, and D. Gurkan, “Network traffic generation: A survey and methodology,” ACM Computing Surveys (CSUR), vol. 55, no. 2, pp. 1–23, 2022

    2022

  6. [6]

    High-fidelity cellu- lar network control-plane traffic generation without domain knowledge,

    Z. J. Kong, N. Hu, Y . C. Hu, J. Meng, and Y . Koral, “High-fidelity cellu- lar network control-plane traffic generation without domain knowledge,” in Proceedings of the 2024 ACM on Internet Measurement Conference , 2024, pp. 530–544

    work page 2024

  7. [7]

    Survey of reconfigurable data center net- works: Enablers, algorithms, complexity,

    K.-T. Foerster and S. Schmid, “Survey of reconfigurable data center net- works: Enablers, algorithms, complexity,” ACM SIGACT News, vol. 50, no. 2, pp. 62–79, 2019

    work page 2019

  8. [8]

    On the complexity of traffic traces and implications,

    C. Avin, M. Ghobadi, C. Griner, and S. Schmid, “On the complexity of traffic traces and implications,” Proceedings of the ACM on Mea- surement and Analysis of Computing Systems , vol. 4, no. 1, pp. 1–29, 2020

    work page 2020

Show all 37 references
  1. [9]

    Projector: Agile reconfigurable data center interconnect,

    M. Ghobadi, R. Mahajan, A. Phanishayee, N. Devanur, J. Kulkarni, G. Ranade, P.-A. Blanche, H. Rastegarfar, M. Glick, and D. Kilper, “Projector: Agile reconfigurable data center interconnect,” in Proc. ACM SIGCOMM 2016 Conference . ACM, 2016, pp. 216–229

    2016

  2. [10]

    Demand-aware network designs of bounded degree,

    C. Avin, K. Mondal, and S. Schmid, “Demand-aware network designs of bounded degree,” in Proc. International Symposium on Distributed Computing (DISC), 2017

    2017

  3. [11]

    Legal issues surrounding monitoring during network research,

    D. C. Sicker, P. Ohm, and D. Grunwald, “Legal issues surrounding monitoring during network research,” in Proceedings of the 7th ACM SIGCOMM conference on Internet measurement , 2007, pp. 141–148

    2007

  4. [12]

    Pac-gan: Packet generation of network traffic using gen- erative adversarial networks,

    A. Cheng, “Pac-gan: Packet generation of network traffic using gen- erative adversarial networks,” in 2019 IEEE 10th Annual Information Technology, Electronics and Mobile Communication Conference (IEM- CON). IEEE, 2019, pp. 0728–0734

    2019

  5. [13]

    Practical gan-based synthetic ip header trace generation using netshare,

    Y . Yin, Z. Lin, M. Jin, G. Fanti, and V . Sekar, “Practical gan-based synthetic ip header trace generation using netshare,” in Proceedings of the ACM SIGCOMM 2022 Conference , 2022, pp. 458–472. 12

    2022

  6. [14]

    Generative spatiotemporal image exploitation for datacenter traffic prediction,

    G. Byun, H. Yang, S. M. Raza, M. Kim, M. Y . Chung, and H. Choo, “Generative spatiotemporal image exploitation for datacenter traffic prediction,” Computer Networks, vol. 254, p. 110755, 2024

    2024

  7. [15]

    Flow-based network traffic generation using generative adversarial networks,

    M. Ring, D. Schl ¨or, D. Landes, and A. Hotho, “Flow-based network traffic generation using generative adversarial networks,” Computers & Security, vol. 82, pp. 156–172, 2019

    2019

  8. [16]

    Netgpt: Generative pretrained transformer for network traffic,

    X. Meng, C. Lin, Y . Wang, and Y . Zhang, “Netgpt: Generative pretrained transformer for network traffic,” arXiv preprint arXiv:2304.09513, 2023

    2023 arXiv

  9. [17]

    Trafficgpt: Breaking the token barrier for efficient long traffic analysis and generation,

    J. Qu, X. Ma, and J. Li, “Trafficgpt: Breaking the token barrier for efficient long traffic analysis and generation,” arXiv preprint arXiv:2403.05822, 2024

    2024 arXiv

  10. [18]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” Advances in Neural Information Processing Systems, 2017

    2017

  11. [19]

    Beyond matchings: Dynamic multi- hop topology for demand-aware datacenters,

    C. Griner, C. Avin, and G. Einziger, “Beyond matchings: Dynamic multi- hop topology for demand-aware datacenters,” Computer Networks, vol. 240, p. 110143, 2024

    2024

  12. [20]

    On- line dynamic b-matching: With applications to reconfigurable datacenter networks,

    M. Bienkowski, D. Fuchssteiner, J. Marcinkowski, and S. Schmid, “On- line dynamic b-matching: With applications to reconfigurable datacenter networks,” ACM SIGMETRICS Performance Evaluation Review, vol. 48, no. 3, pp. 99–108, 2021

    2021

  13. [21]

    Characterization of the DOE mini-apps,

    U. DOE, “Characterization of the DOE mini-apps,” https://portal.nersc. gov/project/CAL/doe-miniapps.htm, 2016

    2016

  14. [22]

    A mathematical theory of communication,

    C. E. Shannon, “A mathematical theory of communication,” Bell system technical journal, vol. 27, no. 3, pp. 379–423, 1948

    1948

  15. [23]

    On the complexity of finite sequences,

    A. Lempel and J. Ziv, “On the complexity of finite sequences,” IEEE Transactions on information theory , vol. 22, no. 1, pp. 75–81, 1976

    1976

  16. [24]

    Universal prediction of indi- vidual sequences,

    M. Feder, N. Merhav, and M. Gutman, “Universal prediction of indi- vidual sequences,” IEEE transactions on Information Theory , vol. 38, no. 4, pp. 1258–1270, 1992

    1992

  17. [25]

    Language models are unsupervised multitask learners,

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever et al., “Language models are unsupervised multitask learners,” OpenAI blog , vol. 1, no. 8, p. 9, 2019

    2019

  18. [26]

    Improving language understanding by generative pre- training,

    A. Radford, “Improving language understanding by generative pre- training,” 2018

    2018

  19. [27]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805 , 2018

    2018 arXiv

  20. [28]

    Language mod- els are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell et al., “Language mod- els are few-shot learners,” Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020

    1901

  21. [29]

    Inside the social network’s (datacenter) network,

    A. Roy, H. Zeng, J. Bagga, G. Porter, and A. C. Snoeren, “Inside the social network’s (datacenter) network,” in Proceedings of the 2015 ACM Conference on Special Interest Group on Data Communication . ACM, 2015, pp. 123–137

    2015

  22. [30]

    COLLECTION, https://trace-collection.net/

    T. COLLECTION, https://trace-collection.net/

  23. [31]

    Automated assistance for creative writing with an rnn language model,

    M. Roemmele and A. S. Gordon, “Automated assistance for creative writing with an rnn language model,” in Companion Proceedings of the 23rd International Conference on Intelligent User Interfaces , 2018, pp. 1–2

    2018

  24. [32]

    Is temperature the creativity parameter of large language models?

    M. Peeperkorn, T. Kouwenhoven, D. Brown, and A. Jordanous, “Is temperature the creativity parameter of large language models?” arXiv preprint arXiv:2405.00492, 2024

    2024 arXiv

  25. [33]

    A learning algorithm for boltzmann machines,

    D. H. Ackley, G. E. Hinton, and T. J. Sejnowski, “A learning algorithm for boltzmann machines,” Cognitive science, vol. 9, no. 1, pp. 147–169, 1985

    1985

  26. [34]

    High-resolution measurement of data center microbursts,

    Q. Zhang, V . Liu, H. Zeng, and A. Krishnamurthy, “High-resolution measurement of data center microbursts,” in Proceedings of the 2017 Internet Measurement Conference , 2017, pp. 78–85

    2017

  27. [35]

    oxfordreference, https://www.oxfordreference.com/display/10.1093/oi/ authority.20110803100312821/

    doi

  28. [36]

    How much do language models copy from their training data? eval- uating linguistic novelty in text generation using raven,

    R. T. McCoy, P. Smolensky, T. Linzen, J. Gao, and A. Celikyilmaz, “How much do language models copy from their training data? eval- uating linguistic novelty in text generation using raven,” Transactions of the Association for Computational Linguistics , vol. 11, pp. 652–670, 2023

    2023

  29. [37]

    Quantifying memorization across neural language models,

    N. Carlini, D. Ippolito, M. Jagielski, K. Lee, F. Tramer, and C. Zhang, “Quantifying memorization across neural language models,” arXiv preprint arXiv:2202.07646, 2022. 13 APPENDIX An interesting use case for our approach is that it also lends itself to generating traces with ...

    2022 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.