Pith. sign in

REVIEW 3 major objections 5 minor 56 references

CENTILE, a single pretrained generative model over irregular telemetry event streams, claims to improve both HPC backfilling and network capacity provisioning through the calibrated quantiles it serves.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

One pretrained telemetry model, CENTILE, improves both HPC backfilling and ISP capacity provisioning decisions under replay, with zero-shot transfer across months and domains.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection The decision-replay evaluation is a genuine contribution, and the cross-domain transfer is the most interesting result, but the 'any horizon' claim rests on a training procedure the paper never describes. the 3 major comments →

arxiv 2608.01725 v1 pith:Q3YJN5CL submitted 2026-08-03 cs.NI cs.LG

CENTILE: A Telemetry Foundation Model Evaluated by the Decisions It Drives

Zifan Zhang , Zhichao Hou , Tingxiang Ji , Yuchen Liu This is my paper
classification cs.NI cs.LG
keywords telemetry foundation modelconditional quantile forecastingdecision replay evaluationHPC backfillingcapacity provisioningirregularly timed event streamszero-shot transfer
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 claims that the right way to build a foundation model for system and network telemetry is to train one generative model over irregularly timed event streams and evaluate it by the operational decisions its forecasts drive, not by point-forecast error. The model, CENTILE, learns the conditional distribution of each entity's next measurement and serves any requested quantile in a single pass, so a scheduler can consume a runtime quantile as a walltime estimate and a provisioner can consume a traffic quantile as capacity. On production HPC job logs and a national ISP trace, re-running the recorded decisions with CENTILE's forecasts substituted for the deployed ones lowers mean bounded slowdown by up to about 77% over user estimates and roughly halves the deployed rule's violation rate. One small pretrained estimator transfers zero-shot to later months, and its pretrained weights transfer across domains from a day of target data, so the promise is a single artifact replacing a fleet of per-task, per-horizon predictors.

Core claim

CENTILE's central claim is that operational telemetry decisions consume quantiles, not point forecasts, so a generative model that learns the calibrated conditional distribution of an entity's next measurement is exactly the artifact both scheduling and provisioning need. The paper presents CENTILE as the first pretrained telemetry model to improve both HPC backfilling and network capacity provisioning under decision replay: it treats telemetry as per-entity, irregularly timed event streams, uses intensity-preserving attention and a heavy-tailed mixture output, and decodes every forecast horizon directly in one pass without requiring future timestamps. Its runtime estimator transfers zero-sh

What carries the argument

The central object is the served conditional quantile: CENTILE learns the predictive distribution of an entity's next measurement and reduces it to the q-quantile used as a job walltime estimate or the tau-quantile used as provisioned capacity. The carrying mechanisms are intensity-preserving attention (pointwise scores without softmax, so submission bursts and idle periods are not normalized away, plus an update gate that suppresses stale history), a Student-t mixture head that keeps heavy tails and serves any quantile on demand, and direct multi-horizon conditioning in which the requested horizon enters only as an additive Fourier feature added to the state, so every horizon is a single pa

Load-bearing premise

The load-bearing premise is that the model was trained to predict several steps ahead directly, but the paper's training section describes only a one-step next-event objective, so the long-horizon results and the 'any horizon in one pass' claim depend on a training procedure that is not described.

What would settle it

Train CENTILE with only the stated one-step likelihood, then evaluate the h=24 forecasts on the ISP trace both with and without the additive horizon-shift term; if the long-horizon forecast-quality numbers persist without horizon-conditioned training, direct multi-horizon decoding is not the mechanism producing them, and if the numbers vanish without the trained head, the paper's training description is incomplete.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • One pretrained artifact per domain can replace the fleet of per-task, per-horizon, per-entity predictors, collapsing training pipelines, retraining cadence, and monitoring overhead.
  • Decision-replay evaluation can supplement or replace point-forecast error for telemetry models, since error saturates near naive baselines while the replayed decisions continue to diverge.
  • Cross-domain pretraining gives new deployments calibrated estimates from hours to a day of local data, lowering the barrier for cold-start systems.
  • Audited calibration turns the served quantile into an a priori contract: operators can choose the level from cost or risk preferences and expect measured violation rates to respect it.
  • The 2.5 MiB CPU-only estimator serves decisions in milliseconds, so the decision gains are compatible with existing scheduling cycles and control-plane budgets.

Where Pith is reading between the lines

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

  • The same quantile-serving design would likely apply to other decisions that consume upper bounds or percentiles, such as autoscaling and workload placement, but the paper only demonstrates scheduling and provisioning.
  • The scheduling gains do not require the model to be more accurate than true runtimes; they come from EASY backfilling's non-monotone response to estimate inflation, so similar headroom may exist in other backfill schedulers whose user estimates are inflated.
  • Direct multi-horizon decoding removes rollout latency from the inference path, suggesting a natural next experiment is closing a live control loop rather than replaying history.
  • If the shared event-stream structure across network, HPC, and cloud telemetry is real, adding storage or power telemetry to the pretraining pool could strengthen the cross-domain transfer further; the paper does not test that.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces CENTILE, a generative foundation model for irregularly timed telemetry event streams. The model embeds per-entity events, processes causal histories with intensity-preserving attention layers, and emits a Student-t mixture head from which arbitrary quantiles are sampled. The authors evaluate CENTILE not by point-forecast error but by replaying two operational decisions: EASY backfilling on the F-DATA Fugaku job log and next-hour capacity provisioning on the CESNET ISP traffic trace. They report that CENTILE lowers mean bounded slowdown by up to about 77% over deployed user estimates, outperforms history- and learning-based walltime baselines across months, transfers cross-domain from network/cloud telemetry to HPC scheduling, and roughly halves the operator rule's provisioning violation rate. The paper also claims an 'any horizon in a single pass' capability via direct multi-horizon decoding and reports forecast-quality improvements at h=1, 4, and 24 on CESNET.

Significance. The decision-replay evaluation is a genuine methodological contribution: it evaluates telemetry forecasts by the scheduling and provisioning decisions they drive, with external baselines (user estimates, history q90, Tsafrir, GBM, Chronos-Bolt, EWMA, GRU/TGR) rather than by fitted values. The empirical tables are extensive, the calibration audits are a useful check, and the cross-month zero-shot and cross-domain transfer results are plausibly significant if reproducible. A 1.3M-parameter, 2.5 MiB estimator serving decisions in milliseconds is also a practical strength. The main concern is that the direct multi-horizon decoding mechanism, which underpins the 'any horizon' claim and the long-horizon forecast results, is not supported by the training objective as described; this needs to be resolved before the paper's central claims can be accepted.

major comments (3)
  1. [III-B and III-C, Eq. (10) and Eq. (11)] The training objective in Eq. (11) is strictly one-step: the model minimizes -log p_theta(x_{i+1} | z_i). The direct multi-horizon decoding in Eq. (10) instead conditions the head on a horizon-shifted state z_i + W_psi psi(log(1+h)) and decodes a distribution for the value h steps ahead. For h=4 and h=24 to be valid, the head and W_psi must have been trained on horizon-conditioned targets (e.g., p_theta(x_{i+h} | z_i + W_psi psi(...))). No such training procedure is described anywhere in Section III. As written, the additive horizon term is an untrained perturbation at inference, and the 'any horizon in a single pass' contribution and Table VII's h=4/h=24 rows do not follow from the stated method. Moreover, even h=1 forecaster inference uses a nonzero shift psi(log 2), so the capacity-provisioning results in Table VI and Figure 4 also depend on this mechanism. The authors must either des
  2. [III-D and IV-E] The claimed superiority of direct decoding over autoregressive rollout rests on the direct head being a separately trained estimator of each horizon's conditional law. Since the training procedure for that head is absent (see previous comment), the ablation in Section IV-E comparing 'direct decoding' with 'rollout' using 'identical weights' compares an untrained or unspecified direct head against rollout, so the reported MASE gap (4.35 vs. 8.34 at 64 steps) and the cost comparison do not support the stated conclusion. Additionally, the text states that 'unrolling this recursion bounds the expected rollout error' by a geometric sum, but no proof or formal statement of the Lipschitz/contraction condition is provided. Please supply the proof or a precise reference, and revisit the ablation once the direct head's training is specified.
  3. [IV-A and reproducibility] The code availability statement points to a GitHub URL but gives no commit hash, no environment specification, and no reproduction script. Given that the multi-horizon training inconsistency can only be resolved by inspecting the actual training loop, a precise artifact (with a commit hash and the exact training configuration for the forecaster) is essential for verification. I am not treating this as grounds for rejection, but it should be part of the revision.
minor comments (5)
  1. [Table III] Only CENTILE rows report mean ± std over seeds. The baselines have no variance or significance tests, making it hard to assess whether the 24.3 vs. 22.2 difference is meaningful. Please add seed variance or at least a paired comparison for the main claims.
  2. [IV-A] MASE for h=4 and h=24 is reported but the naive reference is not explicitly defined for multi-step horizons. State whether it is the last observed value at forecast origin or the seasonal naive benchmark.
  3. [Eq. (2)] The bounded-slowdown floor epsilon is left unspecified. Since very short jobs can dominate the ratio, please give the value used in the replay and test sensitivity to it.
  4. [IV-D] The comparison with Chronos-Bolt is useful, but Chronos-Bolt is zero-shot on CESNET while CENTILE is pretrained on the first 80% of the same dataset. This asymmetry should be stated more explicitly in the text so readers do not interpret Table VI as a matched pretraining comparison.
  5. [III-D, Remark] The remark that guarantees are conditional on measured calibration is honest and appropriate. In the revision, please make the same caveat visible at the first mention of Proposition 1, since the abstract and introduction currently state the decision improvements without this condition.

Circularity Check

0 steps flagged

No significant circularity: decision-level results are measured against external baselines and explicit calibration assumptions; the multi-horizon decoding gap is an omitted training description, not a circular reduction.

full rationale

CENTILE's central claims are evaluated by decision replay against external baselines (user estimates, History q90, Tsafrir last-2, GBM quantile, Chronos-Bolt, GRU/TGR) on held-out traces, and the model is pretrained with a strictly proper one-step likelihood (Eq. 11). The theoretical guarantees in Section III-D are explicitly conditional on Assumption 1 (quantile calibration), and calibration is subsequently audited empirically (Fig. 5 and Table VI), so the guarantees are not assumed by construction. Self-citations, including [23] and [50]–[54], appear as related work or as a non-load-bearing remark that replacing softmax attention is an established route; none is used to forbid alternatives or to import a uniqueness theorem. The one notable derivation gap is that Eq. (11) trains only the one-step next-event likelihood, while Eq. (10) is used to serve h=4 and h=24 forecasts without a described horizon-conditioned training objective; this is an omitted training description and an internal-consistency risk, not a circular equation-to-equation reduction, so it does not raise the circularity score.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The ledger is moderate. The architecture reuses published components, so no new physical or formal entities are invented. The main decision-relevant choices are the served quantile and the backlogged evaluation band, both operator/policy selections. The heavier unstated burden is the multi-horizon training mechanism and the fidelity of the replay simulations.

free parameters (5)
  • Served quantile q for walltime and tau for provisioning = q=0.5 primary, q=0.6 sweeps; tau=0.9 for headline violation comparison
    Risk levels chosen by the decision policy; headline gains are reported at these levels. The paper claims robustness across q=0.5-0.8 but does not provide a formal sensitivity analysis for tau.
  • Backlogged capacity band = offered-to-capacity ratios 0.95 to 1.46
    Evaluation subset where queues form; this selection could favor the reported improvements, although multiple capacities are shown.
  • Context length C = 96 events
    Hyperparameter for per-entity history; no sensitivity analysis reported.
  • Mixture components M and network width D = M=6, D=256 for estimator, D=512 for forecaster
    Architecture capacity choices; no ablations varying M or D.
  • Monte Carlo draws N_s = 2048
    Number of draws used to estimate served quantiles from the mixture; no reported error analysis for quantile Monte Carlo error.
axioms (4)
  • domain assumption Assumption 1: at the served level and horizon, the served distribution is calibrated (Eq. 4) for the events the decision consumes.
    Introduced in Section III-D; all theoretical decision guarantees (Proposition 1 and the newsvendor argument) are conditional on it. The paper audits calibration empirically (Figure 5) but cannot guarantee it at deployment.
  • standard math Minimizing the strictly proper scoring rule of Eq. (11) over a sufficiently rich family yields the true conditional distribution.
    Invoked in Sections III-C and III-D; the family is finite in practice and optimization is approximate.
  • domain assumption The EASY backfill simulation with estimate-extension correction [18], capacity set from offered load, and kill limits faithfully represents the deployed scheduler.
    The decision-replay conclusions depend on this replay faithfully modeling the real system; no live deployment is tested.
  • domain assumption Rollout-feedback regularity: bounded expected one-step error and a Lipschitz feedback map.
    Used in Section III-D to argue direct decoding avoids compounding; bursty telemetry may not satisfy the contraction needed for rollout bounds.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of CENTILE: A Telemetry Foundation Model Evaluated by the Decisions It Drives." pith.science (2026). https://pith.science/paper/Q3YJN5CL

@misc{pith2026260801725,
  author       = {Pith},
  title        = {Pith review of: CENTILE: A Telemetry Foundation Model Evaluated by the Decisions It Drives},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q3YJN5CL}},
  note         = {Machine review of arXiv:2608.01725}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Modern computing and networking infrastructure emits telemetry continuously, yet operators convert it into decisions with a separate predictor per task, entity, and horizon. One generative model, pretrained once over an operator's own event streams, could replace this fleet, an approach that already scales to high-cardinality streams in recommendation systems. However, point-forecast error on operational telemetry saturates near simple last-value baselines, so lower error alone need not improve the decisions it feeds. To close this gap, we present \sys, a generative foundation model for network and systems telemetry, evaluated by replaying the decisions its calibrated conditional quantiles drive. \sys treats heterogeneous telemetry as event-driven, irregularly timed entity streams and serves flexible forecast horizons in a single pass, requiring no future timestamps. To our knowledge, \sys is the first pretrained telemetry model to improve both HPC scheduling and network provisioning decisions under replay, its runtime estimator transferring zero-shot across months and its pretrained weights across domains from hours of target data. Extensive experiments on HPC job logs and network traffic confirm that \sys lowers the mean bounded slowdown of backfilling by up to approximately $77\%$ over deployed user estimates and roughly halves the deployed rule's violation rate. Our code is available at https://github.com/ZzZTripleZzZ/all-in-one.

Figures

Figures reproduced from arXiv: 2608.01725 by Tingxiang Ji, Yuchen Liu, Zhichao Hou, Zifan Zhang.

Figure 1
Figure 1. Figure 1: CENTILE overview. Events from telemetry entities flow through the adapter, the pretrained attention layers, and the Student-t mixture head into calibrated quantiles that drive the replayed decisions. The inset expands one of the L attention layers. assume a uniform sampling grid that discards event timing, and serve fixed quantile grids whose most conservative level falls short of what provisioning require… view at source ↗
Figure 2
Figure 2. Figure 2: Mean bounded slowdown across the backlogged capacity band on the [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: CENTILE attains the lowest mean bounded slowdown on every month. Shaded months are served zero-shot by the April-pretrained model. 7 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Provisioning frontier on CESNET. Crosses denote operator rules, and [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Nominal quantile vs. empirical coverage of the runtime estimator [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

56 extracted references · 48 canonical work pages · 1 internal anchor

  1. [1]

    Resource central: Understanding and predicting work- loads for improved resource management in large cloud platforms,

    E. Cortez, A. Bonde, A. Muzio, M. Russinovich, M. Fontoura, and R. Bianchini, “Resource central: Understanding and predicting work- loads for improved resource management in large cloud platforms,” in ACM SOSP, 2017

    work page 2017

  2. [2]

    M100 exadata: a data collection campaign on the cineca’s marconi100 tier-0 supercomputer,

    A. Borghesiet al., “M100 exadata: a data collection campaign on the cineca’s marconi100 tier-0 supercomputer,”Scientific Data, 2023

    work page 2023

  3. [3]

    Google cluster-usage traces: Format + schema,

    C. Reiss, J. Wilkes, and J. L. Hellerstein, “Google cluster-usage traces: Format + schema,” Google Inc., Tech. Rep., 2011

    work page 2011

  4. [4]

    Borg: The next generation,

    M. Tirmaziet al., “Borg: The next generation,” inEuroSys, 2020

    work page 2020

  5. [5]

    Challenges in deploying machine learning: A survey of case studies,

    A. Paleyes, R.-G. Urma, and N. D. Lawrence, “Challenges in deploying machine learning: A survey of case studies,”ACM Comput. Surv., 2022

    work page 2022

  6. [6]

    Decision-focused learning: Foundations, state of the art, benchmark and future opportunities,

    J. Mandi, J. Kotary, S. Berden, M. Mulamba, V . Bucarey, T. Guns, and F. Fioretto, “Decision-focused learning: Foundations, state of the art, benchmark and future opportunities,”J. Artif. Intell. Res., 2024

    work page 2024

  7. [7]

    Et-bert: A contextualized datagram representation with pre-training transformers for encrypted traffic classification,

    X. Lin, G. Xiong, G. Gou, Z. Li, J. Shi, and J. Yu, “Et-bert: A contextualized datagram representation with pre-training transformers for encrypted traffic classification,” inWWW, 2022

    work page 2022

  8. [8]

    net- found: Foundation model for network security and performance,

    S. Guthula, N. Battula, R. Beltiukov, W. Guo, and A. Gupta, “net- found: Foundation model for network security and performance,” arXiv:2310.17025, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Mobigpt: A foundation model for mobile wireless networks,

    X. Qi, H. Chai, and Y . Li, “Mobigpt: A foundation model for mobile wireless networks,”arXiv:2509.18166, 2025

    work page arXiv 2025

  10. [10]

    Large wireless model (lwm): A foundation model for wireless channels,

    S. Alikhani, G. Charan, and A. Alkhateeb, “Large wireless model (lwm): A foundation model for wireless channels,”arXiv:2411.08872, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    Netllm: Adapting large language models for networking,

    D. Wu, X. Wang, Y . Qiao, Z. Wang, J. Jiang, S. Cui, and F. Wang, “Netllm: Adapting large language models for networking,” inACM SIGCOMM, 2024

    work page 2024

  12. [12]

    Chronos: Learning the language of time series,

    A. F. Ansariet al., “Chronos: Learning the language of time series,” TMLR, 2024

    work page 2024

  13. [13]

    A decoder-only foundation model for time-series forecasting,

    A. Das, W. Kong, R. Sen, and Y . Zhou, “A decoder-only foundation model for time-series forecasting,” inICML, 2024

    work page 2024

  14. [14]

    Uomo: A universal model of mobile traffic forecasting for wireless network optimization,

    H. Chai, S. Zhang, X. Qi, B. Qiu, and Y . Li, “Uomo: A universal model of mobile traffic forecasting for wireless network optimization,” inACM KDD, 2025

    work page 2025

  15. [15]

    F- data: A fugaku workload dataset for job-centric predictive modelling in hpc systems,

    F. Antici, A. Bartolini, J. Domke, Z. Kiziltan, and K. Yamamoto, “F- data: A fugaku workload dataset for job-centric predictive modelling in hpc systems,”Scientific Data, 2025

    work page 2025

  16. [16]

    Who limits the resource efficiency of my datacenter: An analysis of alibaba datacenter traces,

    J. Guo, Z. Chang, S. Wang, H. Ding, Y . Feng, L. Mao, and Y . Bao, “Who limits the resource efficiency of my datacenter: An analysis of alibaba datacenter traces,” inIEEE/ACM IWQoS, 2019

    work page 2019

  17. [17]

    The ANL/IBM SP scheduling system,

    D. A. Lifka, “The ANL/IBM SP scheduling system,” inJSSPP, 1995

    work page 1995

  18. [18]

    Backfilling using system- generated predictions rather than user runtime estimates,

    D. Tsafrir, Y . Etsion, and D. G. Feitelson, “Backfilling using system- generated predictions rather than user runtime estimates,”IEEE TPDS, 2007

    work page 2007

  19. [19]

    Obtaining dynamic schedul- ing policies with simulation and machine learning,

    D. Carastan-Santos and R. Y . de Camargo, “Obtaining dynamic schedul- ing policies with simulation and machine learning,” inSC, 2017

    work page 2017

  20. [20]

    Strictly proper scoring rules, prediction, and estimation,

    T. Gneiting and A. E. Raftery, “Strictly proper scoring rules, prediction, and estimation,”J. Amer. Statist. Assoc., 2007

    work page 2007

  21. [21]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” inNeurIPS, 2017

    2017

  22. [22]

    Actions speak louder than words: Trillion-parameter sequential transducers for generative recommendations,

    J. Zhaiet al., “Actions speak louder than words: Trillion-parameter sequential transducers for generative recommendations,” inICML, 2024

    work page 2024

  23. [23]

    Protransformer: Robustify transformers via plug-and-play paradigm,

    Z. Hou, W. Gao, Y . Shen, F. Wang, and X. Liu, “Protransformer: Robustify transformers via plug-and-play paradigm,” inNeurIPS, 2024

    work page 2024

  24. [24]

    Going deeper with image transformers,

    H. Touvron, M. Cord, A. Sablayrolles, G. Synnaeve, and H. Jégou, “Going deeper with image transformers,” inICCV, 2021

    work page 2021

  25. [25]

    Mixture density networks,

    C. M. Bishop, “Mixture density networks,” Aston University, Tech. Rep., 1994

    work page 1994

  26. [26]

    A multi- horizon quantile recurrent forecaster,

    R. Wen, K. Torkkola, B. Narayanaswamy, and D. Madeka, “A multi- horizon quantile recurrent forecaster,”arXiv:1711.11053, 2017

    Pith/arXiv arXiv 2017

  27. [27]

    Improving multi-step prediction of learned time series models,

    A. Venkatraman, M. Hebert, and J. A. Bagnell, “Improving multi-step prediction of learned time series models,” inAAAI, 2015

    work page 2015

  28. [28]

    Recurrent neural net- works for time series forecasting: Current status and future directions,

    H. Hewamalage, C. Bergmeir, and K. Bandara, “Recurrent neural net- works for time series forecasting: Current status and future directions,” Int. J. Forecasting, 2021

    work page 2021

  29. [29]

    The big data newsvendor: Practical insights from machine learning,

    G.-Y . Ban and C. Rudin, “The big data newsvendor: Practical insights from machine learning,”Oper. Res., 2019

    work page 2019

  30. [30]

    Cesnet-timeseries24: Time series dataset for network traffic anomaly detection and forecasting,

    J. Koumar, K. Hynek, T. ˇCejka, and P. Šiška, “Cesnet-timeseries24: Time series dataset for network traffic anomaly detection and forecasting,” Scientific Data, 2025

    work page 2025

  31. [31]

    The M4 com- petition: 100,000 time series and 61 forecasting methods,

    S. Makridakis, E. Spiliotis, and V . Assimakopoulos, “The M4 com- petition: 100,000 time series and 61 forecasting methods,”Int. J. Forecasting, 2020

    work page 2020

  32. [32]

    Improving backfilling by using machine learning to predict running times,

    E. Gaussier, D. Glesser, V . Reis, and D. Trystram, “Improving backfilling by using machine learning to predict running times,” inSC, 2015

    work page 2015

  33. [33]

    Uarp: Uncertainty-aware runtime prediction for preventing scheduler termination under wallclock constraints in hpc,

    J. Choi and S. Oh, “Uarp: Uncertainty-aware runtime prediction for preventing scheduler termination under wallclock constraints in hpc,”J. Supercomput., 2026

    work page 2026

  34. [34]

    Netgpt: Generative pretrained transformer for network traffic,

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

    Pith/arXiv arXiv 2023

  35. [35]

    NetMamba: Efficient network traffic classification via pre-training unidirectional mamba,

    T. Wang, X. Xie, W. Wang, C. Wang, Y . Zhao, and Y . Cui, “NetMamba: Efficient network traffic classification via pre-training unidirectional mamba,” inIEEE ICNP, 2024

    work page 2024

  36. [36]

    Unified training of universal time series forecasting transformers,

    G. Woo, C. Liu, A. Kumar, C. Xiong, S. Savarese, and D. Sahoo, “Unified training of universal time series forecasting transformers,” in ICML, 2024

    work page 2024

  37. [37]

    Lag-llama: Towards foundation models for probabilistic time series forecasting,

    K. Rasulet al., “Lag-llama: Towards foundation models for probabilistic time series forecasting,”arXiv:2310.08278, 2023

    Pith/arXiv arXiv 2023

  38. [38]

    MOMENT: A family of open time-series foundation models,

    M. Goswami, K. Szafer, A. Choudhry, Y . Cai, S. Li, and A. Dubrawski, “MOMENT: A family of open time-series foundation models,” inICML, 2024

    work page 2024

  39. [39]

    Time-LLM: Time series forecasting by reprogramming large language models,

    M. Jin, S. Wang, L. Ma, Z. Chu, J. Y . Zhang, X. Shi, P.-Y . Chen, Y . Liang, Y .-F. Li, S. Pan, and Q. Wen, “Time-LLM: Time series forecasting by reprogramming large language models,” inICLR, 2024

    work page 2024

  40. [40]

    One fits all: Power general time series analysis by pretrained LM,

    T. Zhou, P. Niu, X. Wang, L. Sun, and R. Jin, “One fits all: Power general time series analysis by pretrained LM,” inNeurIPS, 2023

    work page 2023

  41. [41]

    Tiny time mixers (ttms): Fast pre-trained models for enhanced zero/few-shot forecasting of multivariate time series,

    V . Ekambaram, A. Jati, P. Dayama, S. Mukherjee, N. H. Nguyen, W. M. Gifford, C. Reddy, and J. Kalagnanam, “Tiny time mixers (ttms): Fast pre-trained models for enhanced zero/few-shot forecasting of multivariate time series,”arXiv:2401.03955, 2024

    Pith/arXiv arXiv 2024

  42. [42]

    Time-series foundation models for isp traffic forecasting,

    F. Liu, B. Farkiani, and P. Crowley, “Time-series foundation models for isp traffic forecasting,”arXiv:2511.17529, 2025

    work page arXiv 2025

  43. [43]

    CLOUDADV: Decision-Aligned Instance Sizing with Zero-Shot Foundation Models under Drift

    J. Bell, G. Carfi, G. Gramaglia, A. Simioni, D. Fontani, and V . Lomonaco, “Cloudadv: Decision-aligned instance sizing with zero- shot foundation models under drift,”arXiv:2606.31470, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  44. [44]

    Decision- focused fine-tuning of time series foundation models for dispatchable feeder optimization,

    M. Beichter, N. Friederich, J. Pinter, D. Werling, K. Phipps, S. Beichter, O. Neumann, R. Mikut, V . Hagenmeyer, and B. Heidrich, “Decision- focused fine-tuning of time series foundation models for dispatchable feeder optimization,”Energy and AI, 2025

    work page 2025

  45. [45]

    Walltime prediction and its impact on job scheduling performance and predictability,

    D. Klusá ˇcek and M. Soysal, “Walltime prediction and its impact on job scheduling performance and predictability,” inJSSPP, 2020

    work page 2020

  46. [46]

    A reinforcement learning based backfilling strategy for hpc batch jobs,

    E. Kolker-Hicks, D. Zhang, and D. Dai, “A reinforcement learning based backfilling strategy for hpc batch jobs,” inSC Workshops (SC-W), 2023

    work page 2023

  47. [47]

    Autopilot: Workload autoscaling at Google,

    K. Rzadca, P. Findeisen, J. Swiderski, P. Zych, P. Broniek, J. Kusmierek, P. Nowak, B. Strack, P. Witusowski, S. Hand, and J. Wilkes, “Autopilot: Workload autoscaling at Google,” inACM EuroSys, 2020

    work page 2020

  48. [48]

    Zoom2Net: Constrained network telemetry imputation,

    F. Gong, D. Raghunathan, A. Gupta, and M. Apostolaki, “Zoom2Net: Constrained network telemetry imputation,” inACM SIGCOMM, 2024

    work page 2024

  49. [49]

    DOTE: Rethinking (predictive) W AN traffic engineer- ing,

    Y . Perry, F. V . Frujeri, C. Hoch, S. Kandula, I. Menache, M. Schapira, and A. Tamar, “DOTE: Rethinking (predictive) W AN traffic engineer- ing,” inUSENIX NSDI, 2023. 10

    work page 2023

  50. [50]

    Hlogformer: A hierarchical transformer for representing log data,

    Z. Hou, M. Ghashami, M. Kuznetsov, and M. Torkamani, “Hlogformer: A hierarchical transformer for representing log data,”arXiv:2408.16803, 2024

    Pith/arXiv arXiv 2024

  51. [51]

    OpenTwin: Digital twin driven closed-loop KPM inference and control for open RAN,

    M. Hossen, Z. Zhang, D. Ron, Y . Liu, and V . K. Shah, “OpenTwin: Digital twin driven closed-loop KPM inference and control for open RAN,”arXiv, 2026

    work page 2026

  52. [52]

    Digital twin- assisted data-driven optimization for reliable edge caching in wireless networks,

    Z. Zhang, Y . Liu, Z. Peng, M. Chen, D. Xu, and S. Cui, “Digital twin- assisted data-driven optimization for reliable edge caching in wireless networks,”IEEE JSAC, 2024

    work page 2024

  53. [53]

    Network digital untwinning: Towards backward optimization of digital twins,

    Z. Zhang, D. Chen, A. Gao, M. Wang, M. Chen, M. Fang, X. Yang, and Y . Liu, “Network digital untwinning: Towards backward optimization of digital twins,” inIEEE ICDCS, 2026

    work page 2026

  54. [54]

    Digital network twins for next-generation wireless: Creation, optimization, and challenges,

    Z. Zhang, Z. Peng, H. Yu, M. Chen, and Y . Liu, “Digital network twins for next-generation wireless: Creation, optimization, and challenges,” IEEE Network, 2025

    work page 2025

  55. [55]

    Learning in situ: A randomized experiment in video streaming,

    F. Y . Yan, H. Ayers, C. Zhu, S. Fouladi, J. Hong, K. Zhang, P. Levis, and K. Winstein, “Learning in situ: A randomized experiment in video streaming,” inUSENIX NSDI, 2020

    work page 2020

  56. [56]

    Anthropic, “Claude,” https://www.anthropic.com/claude, 2026. 11

    work page 2026

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.