Pith. sign in

REVIEW 4 major objections 6 minor 3 cited by

A compact state-space world model can simulate 'what-if' PRB choices for near-real-time 6G O-RAN control.

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 ↗

T0 review · deepseek-v4-flash

2026-08-04 00:03 UTC pith:VCFJCMOK

load-bearing objection Solid incremental architecture with a genuinely careful leakage-safe pipeline, but the headline accuracy gains are fragile and the causal/what-if framing outruns the evidence. the 4 major comments →

arxiv 2511.02748 v2 pith:VCFJCMOK submitted 2025-11-04 cs.NI cs.LG

Agentic World Modeling for 6G: Near-Real-Time Generative State-Space Reasoning

Farhad Rezazadeh , Amir Ashtari Gargari , Hatim Chergui , Sandra Lagen , Merouane Debbah , Houbing Song , Lingjia Liu This is my paper
classification cs.NI cs.LG
keywords world modelsO-RAN6Gstate-space modelscounterfactual forecastinguncertainty quantificationmodel predictive controlnear-real-time control
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 paper is trying to establish that 6G O-RAN control should be built on an action-conditioned generative world model rather than on LLM-style token prediction. Its concrete proposal, WM-MS3M, treats physical resource blocks (PRBs) as first-class control inputs, learns a strictly causal state-space forecaster with a compact stochastic latent and dual decoders, and uses the same model for factual prediction, counterfactual 'what-if' rollouts, and short-horizon MPC/CEM planning. The empirical payoff claimed on realistic O-RAN traces is a 1.69% MAE improvement over the pure SSM backbone with about 32% fewer parameters, and 35-80% lower RMSE than attention/hybrid baselines with 2.3-4.1x faster inference. A sympathetic reader would care because this points to a leakage-safe, edge-friendly substrate for risk-aware, auditable network control that can be evaluated offline before deployment.

Core claim

The paper's central discovery claim is that a strictly causal, action-conditioned generative state-space model can serve as the world-model substrate for Near-RT O-RAN control. WM-MS3M couples a multi-scale structured state-space mixture (MS3M) front end—depthwise HiPPO-LegS kernels discretized at multiple learned time scales—with a compact diagonal-Gaussian latent and two decoders: a full-frame reconstruction head and a heteroscedastic target head with a bounded autoregressive skip. The model is trained leakage-safe (train-only scalers, chronological splits, KL annealing, scheduled posterior/prior mixing) and, at test time, MC prior sampling yields calibrated means and a variance decomposit

What carries the argument

The load-bearing object is WM-MS3M, a strictly causal conditional VAE whose backbone is a multi-scale structured state-space mixture (MS3M): per-channel HiPPO-LegS kernels discretized bilinearly at multiple learned time scales, depthwise causal convolutions with left padding, squeeze-excitation channel gating, and GLU mixers. A compact diagonal-Gaussian latent z is inferred from history (prior) or history-plus-future (posterior, training only) and conditions two decoders; the heteroscedastic target head outputs mean and log-variance with a bounded AR skip. This machinery does two jobs: it gives calibrated predictive uncertainty via MC prior sampling, and it makes the model action-conditioned

Load-bearing premise

The whole what-if analysis rests on treating the recorded PRB choices as freely swappable control knobs, even though in the data they were chosen by a scheduler that was simultaneously reacting to load and radio conditions; if those hidden reactions also affect the KPIs, imagined PRB changes may not produce the predicted KPI changes.

What would settle it

Run the CEM planner on a live O-RAN testbed for a sustained period and compare its predicted counterfactual KPI trajectories with actual KPIs under the issued PRB actions; if the predicted response to a PRB change is consistently wrong in sign or magnitude, the counterfactual claim fails.

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

If this is right

  • If correct, O-RAN Near-RT control can use one model for prediction, what-if simulation, and planning, so offline policy screening can replace online trial-and-error.
  • The compact SSM forecaster is deployable at the edge: about 32% fewer parameters than MS3M, 2.3-4.1x faster inference than attention/hybrid baselines, and roughly 0.65 ms per-sample latency.
  • Operators can ask concrete risk questions, such as how BLER exceedance changes if PRBs are cut 20%, and receive calibrated intervals rather than point estimates.
  • The leakage-safe training protocol (train-only scalers, chronological splits) means offline accuracy scores are more likely to transfer to real-time deployment.
  • PRB-bound constrained CEM planning yields auditable, reproducible decisions, so reward weights and constraints can be tuned by the operator.

Where Pith is reading between the lines

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

  • Editorial extension: because the training traces are observational—the scheduler chooses PRBs in response to load and radio state—the counterfactual interpretation of PRB changes depends on an unstated no-confounding assumption; a randomized or instrumental-variable perturbation of PRBs would test whether the model's what-if rollouts track true causal responses.
  • Editorial extension: the same compact-latent-plus-SSM recipe could transfer to other continuous control channels (power, MCS, beam) and to multi-cell coordination, where the paper's single-sector PRB planner is a special case.
  • Editorial extension: the paper validates one-step and H=8 rollouts, but not closed-loop stability over many receding-horizon steps; a testable extension is to measure long-horizon rollout error and whether compounding errors erode the planner's advantage.
  • Editorial extension: if the 35-80% RMSE gap over attention baselines persists under distribution shift (different RAN vendor, traffic mix, or load regime), the SSM inductive bias would be a robust default for Near-RT telemetry; otherwise the gap may be data-specific.

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

4 major / 6 minor

Summary. The paper proposes WM-MS3M, an action-conditioned generative state-space forecaster for near-real-time O-RAN control. The architecture extends the authors' previous MS3M backbone with a diagonal-Gaussian stochastic latent and dual decoders (full-frame reconstruction plus a heteroscedastic target head with an AR skip), trained under a leakage-safe protocol (chronological splits, train-only scalers, KL annealing, posterior/prior mixing). At test time, prior-sampling Monte Carlo averaging produces point predictions and uncertainty decompositions. The same model is used for counterfactual what-if forecasting under hypothetical PRB sequences and for short-horizon MPC/CEM planning. The paper reports that WM-MS3M improves MAE by 1.69% over MS3M with 32% fewer parameters, achieves 35–80% lower RMSE than attention/hybrid baselines, and produces a what-if/planning analysis on a single O-RAN test window in which a 20% PRB reduction is scored as optimal.

Significance. If the central claims held, the paper would make a useful contribution: it demonstrates a compact, strictly causal SSM-based forecaster with calibrated uncertainty and an explicit action channel, plus a reproducible leakage-safe pipeline and released code. The separation of forecasting from decision making and the use of data-driven PRB bounds for planning are sensible engineering choices for Near-RT RIC applications. The manuscript is also transparent about its data provenance and provides algorithms for training, inference, and planning. However, the significance is currently limited by two load-bearing gaps: the causal/counterfactual interpretation of the action-conditioned model is not supported by the observational data or the identification assumptions stated, and the headline empirical gains over MS3M rest on single-run differences far smaller than typical seed noise. The comparison with external baselines is additionally weakened by the authors' own statement that those baselines are non-exact reimplementations. These issues must be addressed before the central novelty—counterfactual what-if reasoning and planning—can be accepted as established.

major comments (4)
  1. [§II (Eq. 3), §V-D, Algorithm 3] The manuscript uses causal and counterfactual language throughout: PRBs are 'first-class causes', Eq. (3) is called the counterfactual predictive distribution, and Algorithm 3 rolls out 'hypothetical' PRB sequences. But the model is trained on observational O-RAN traces in which PRB allocations are chosen by an existing scheduler that reacts to load, radio quality, and interference. No ignorability, sequential randomization, or other causal identification condition is stated. Unobserved confounders such as active UE count, inter-cell interference, and scheduler policy state are not in X_t. Under confounding, p(y_{t+1}|X_t,u_t) is an associational quantity, not p(y_{t+1}|do(u_t)); a model can be accurate under the training policy yet systematically wrong under hypothetical actions. The conclusion in Table IV that Step −20% is optimal may reflect that the scheduler allocates more PRBs in h
  2. [Table II, §V-A] The reported comparison between WM-MS3M and MS3M relies on a single run with no seed variability or confidence intervals. The RMSE difference is 0.000119 (~0.04%), far smaller than typical seed noise for stochastic latent models; the MAE improvement is 1.69%, but there is no evidence it is not within run-to-run variation, especially given stochastic latent sampling, dropout, and mixed-precision training. The paper should provide multiple seeds, standard deviations, paired tests or confidence intervals, and state explicitly whether the reported numbers are one run. If the difference is not stable across seeds, the 'cuts MAE' claim must be tempered.
  3. [§IV-D, §V-B] The 35–80% RMSE gains over RWKV, Performers, RetNet, and Chronos are central to the headline, but Section IV-D admits that these are 'not exact reproductions of the original codebases' and that Chronos models are 'trained from scratch on our dataset' without reporting hyperparameter search budgets or final configurations. Under-tuned baselines can easily explain large gaps. The accuracy-efficiency ranking in Section V-B is therefore not yet robust. The authors should use official checkpoints/standardized implementations where possible, or report detailed hyperparameter tuning for each baseline, including the number of configurations tried and the selected hyperparameters.
  4. [§V-D, Table IV, Fig. 3] The what-if and planning evaluation is a single test window with five hand-defined PRB scenarios; no multiple contexts, no uncertainty intervals on the reward, and no comparison against a simple random-policy or baseline scheduler. The conclusion that in this window Step −20% is optimal is presented as if it were a general property ('the model's learned diminishing SINR/RSRP returns'), but the evidence is anecdotal. The abstract's claim of 'rare-event simulation and offline policy screening' is not demonstrated by these experiments. The authors should repeat the evaluation over several test windows, report the distribution of rewards, and either add a non-trivial baseline policy or explicitly scope the claim to the illustrated example.
minor comments (6)
  1. [§II, Eq. (1)] Notation is confusing: x_t is first defined as the observed KPI vector in R^F, then Eq. (1) uses x_t for both the latent state and the observation. Please use distinct symbols (e.g., s_t for latent) throughout.
  2. [§III-E, Eq. (24)] The variance decomposition uses μ_t for both the per-sample decoder mean and the MC average. Define \bar{μ}_t and clarify the estimator; with S=8 the (S−1) denominator also deserves a comment.
  3. [Table I] The notation table is very dense and contains symbols that are only used once or are redundant (e.g., several scaler variants). Consider pruning to reduce reader burden.
  4. [Fig. 3] The y-axis is a scaled reward with no units; the values ±1.5 are not interpretable. Add a note that the reward is in standardized space and, if possible, show a physical-unit secondary axis.
  5. [Alg. 3] CEM hyperparameters (population 256, elite fraction 0.1, 4 iterations, horizon 8, smoothness 0.05) are free and no sensitivity analysis is reported. At least a small grid over horizon and smoothness would help the Near-RT feasibility claim.
  6. [Throughout] There are several minor typos and spacing inconsistencies, e.g., 'MS 3M' vs 'MS3M' and 'withz (s)'. A careful proofread is recommended.

Circularity Check

0 steps flagged

No circular reduction found; out-of-sample comparisons carry the accuracy claims, and the causal what-if wording is a validity concern rather than a circular step.

full rationale

The paper's quantitative claims are genuinely out-of-sample: Algorithm 1 fits scalers on the training split only, splits are chronological, and test-time inference uses prior sampling (Algorithm 2). The headline comparison against MS3M is an empirical ablation: WM-MS3M's 1.69% MAE improvement and ~31.6% parameter reduction are read from Table II, not derived by construction. The attention/hybrid baselines are in-house reimplementations, which is a reproducibility caveat but not a circular step. Equation (3) defines 'counterfactual (what-if) prediction' as the model's conditional output under hypothetical PRB paths; this is a definition of the reported quantity rather than an equivalence that turns inputs into outputs. The lack of causal identification (observational traces, no ignorability argument) undermines the causal/counterfactual interpretation, but that is a correctness and validity limitation, not a circularity of the derivation. Self-citations to [17], [45], [46] provide the MS3M backbone and data provenance; they are not used as a uniqueness theorem or to forbid alternatives. The MPC/CEM conclusion that Step -20% PRB is best is partly driven by the reward's explicit PRB-cost penalty (Eq. 34), which the paper itself acknowledges ('due to PRB cost'); this is transparent objective design, not a hidden circular reduction. Overall, no step reduces a prediction to a fitted constant or to an unverified self-citation chain.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 1 invented entities

The central claims rest on one public O-RAN dataset, a set of hand-tuned hyperparameters (latent size, KL schedule, PRB bounds, CEM settings), and an unexamined causal assumption that conditioning on observed history is enough to make PRB counterfactuals valid. No new physical entities are introduced; the stochastic latent is a standard modeling construct.

free parameters (8)
  • Latent dimension dz = 48
    Set in Section IV-C; sizes the stochastic latent that drives uncertainty estimates and what-if rollouts.
  • KL annealing schedule = beta 0.01 to 1.0 over 20 epochs
    Controls how much information the latent stores; affects calibration and training stability.
  • Posterior mixing schedule = pi 1.0 to 0.5 over 20 epochs
    Blends posterior (teacher-forced) and prior phases during training; a leakage-safety/performance trade-off knob.
  • MC prior samples S = 8
    Averaging draws at test time; reduces epistemic variance at latency cost.
  • PRB admissible set = [5th, 95th] percentile of train PRBs
    Defines U in Section II; all planning and what-if actions are clamped to this range.
  • CEM planner hyperparameters (H, P, rho, I) = H=8, P=256, rho=0.1, I=4
    Chosen in Section IV-C; planning quality and wall-clock time depend on them.
  • Reward weights w and smoothness lambda_sm = not reported
    Algorithm 3 and Section IV-C; the conclusion that Step -20% is optimal is conditional on these unreported values.
  • Backbone hyperparameters (d, L_l, N, M, dropout) = d=192, L_l=4, N=64, M=4, dropout 0.1
    Architecture configuration in Section IV-C; parameter-efficiency claims depend on it.
axioms (5)
  • domain assumption O-RAN KPI traces used for training are representative of the Near-RT operating regime.
    Section IV-A points to earlier data collection; all numbers in Table II depend on this dataset.
  • domain assumption Conditioning on observed KPI history and PRB history suffices for valid PRB counterfactuals.
    Eq. (3) and Algorithm 3 treat PRB paths as interventions but no ignorability/unconfoundedness condition is stated; PRBs were assigned by a scheduler correlated with unobserved load.
  • standard math HiPPO-LegS operator with bilinear discretization is stable and kernels can be truncated at length L_k.
    Used in Eqs. (6)-(8); the Hurwitz-to-Schur-stable result is cited to [40]-[42].
  • ad hoc to paper CEM with 4 iterations, population 256, and elite fraction 0.1 is a sufficient optimizer for the planning objective.
    Algorithm 3; no convergence analysis or comparison to other planning optimizers.
  • domain assumption Actions outside the data-driven 5th-95th percentile PRB range are excluded because the model is unreliable there.
    Section II defines U; the paper only claims what-if validity within U.
invented entities (1)
  • Diagonal-Gaussian latent state z no independent evidence
    purpose: Captures residual stochasticity not explained by the deterministic SSM summary d; enables MC uncertainty and what-if rollouts.
    Introduced in Section III-D/E as a standard VAE latent; identifiable only up to affine transformations, with no falsifiable prediction outside the model.

pith-pipeline@v1.3.0-alltime-deepseek · 22933 in / 16607 out tokens · 190355 ms · 2026-08-04T00:03:51.385246+00:00 · methodology

0 comments
read the original abstract

We argue that sixth-generation (6G) intelligence is not fluent token prediction but the capacity to imagine and choose -- to simulate future scenarios, weigh trade-offs, and act with calibrated uncertainty. We reframe open radio access network (O-RAN) near-real-time (Near-RT) control via counterfactual dynamics and a world modeling (WM) paradigm that learns an action-conditioned generative state space. This enables quantitative "what-if" forecasting beyond large language models (LLMs) as the primary modeling primitive. Actions such as physical resource blocks (PRBs) are treated as first-class control inputs in a causal world model, and both aleatoric and epistemic uncertainty are modeled for prediction and what-if analysis. An agentic, model predictive control (MPC)-based cross-entropy method (CEM) planner operates over short horizons, using prior-mean rollouts within data-driven PRB bounds to maximize a deterministic reward. The model couples multi-scale structured state-space mixtures (MS3M) with a compact stochastic latent to form WM-MS3M, summarizing key performance indicators (KPIs) histories and predicting next-step KPIs under hypothetical PRB sequences. On realistic O-RAN traces, WM-MS3M cuts mean absolute error (MAE) by 1.69% versus MS3M with 32% fewer parameters and similar latency, and achieves 35-80% lower root mean squared error (RMSE) than attention/hybrid baselines with 2.3-4.1x faster inference, enabling rare-event simulation and offline policy screening.

Figures

Figures reproduced from arXiv: 2511.02748 by Amir Ashtari Gargari, Farhad Rezazadeh, Hatim Chergui, Houbing Song, Lingjia Liu, Merouane Debbah, Sandra Lagen.

Figure 1
Figure 1. Figure 1: Agentic world-modeling pipeline in the O-RAN Near-RT RIC. Aggregated KPIs and PRB actions from the E2 node feed [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Reward comparison across PRB control scenarios over [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. LLM-Based Agentic Negotiation for 6G: Addressing Uncertainty Neglect and Tail-Event Risk

    cs.NI 2025-11 conditional novelty 6.0

    A CVaR-aware agentic framework for 6G network slicing eliminates URLLC SLA violations by shifting LLM decisions from mean latency to tail-risk distributions predicted by digital twins.

  2. LiQSS: Post-Transformer Linear Quantum-Inspired State-Space Tensor Networks for Real-Time 6G

    cs.NI 2026-01 conditional novelty 5.0

    LiQSS, a linear-time state-space tensor-network forecaster, matches or beats Transformer-based KPI predictors on an O-RAN dataset with up to 155x fewer parameters and faster inference.

  3. Mitigating Anchoring Bias in LLM-Based Agents for Energy-Efficient 6G Autonomous Networks

    cs.NI 2026-06 unverdicted novelty 4.0

    LLM agents for 6G slicing exhibit anchoring bias mitigated by Truncated Weibull randomization plus CVaR digital twins, yielding up to 25% energy savings and sub-second inference on a 1B model.

Reference graph

Works this paper leans on

47 extracted references · 12 linked inside Pith · cited by 3 Pith papers

  1. [1]

    World Models,

    D. Ha and J. Schmidhuber, “World Models,”arXiv preprint arXiv:1803.10122, 2018

    Pith/arXiv arXiv 2018

  2. [2]

    World Models for Cognitive Agents: Transforming Edge Intelligence in Future Networks,

    C. Zhaoet al., “World Models for Cognitive Agents: Transforming Edge Intelligence in Future Networks,” 2025. [Online]. Available: https: //arxiv.org/abs/2506.00417

    Pith/arXiv arXiv 2025

  3. [3]

    Edge General Intelligence Through World Models and Agentic AI: Fundamentals, Solutions, and Challenges,

    C. Zhao, G. Liuet al., “Edge General Intelligence Through World Models and Agentic AI: Fundamentals, Solutions, and Challenges,” 2025. [Online]. Available: https://arxiv.org/abs/2508.09561

    Pith/arXiv arXiv 2025

  4. [4]

    Autonomous networks: Exploring the evolution from level 0 to level 5,

    TM Forum, “Autonomous networks: Exploring the evolution from level 0 to level 5,” TM Forum, Tech. Rep., Dec. 2021, technical Report

    2021

  5. [5]

    LLM-xApp: A Large Language Model Empowered Radio Resource Management xApp for 5G O-RAN,

    X. Wuet al., “LLM-xApp: A Large Language Model Empowered Radio Resource Management xApp for 5G O-RAN,” inWorkshop on Security and Privacy of Next-Generation Networks (FutureG) 2025, co-located with NDSS 2025. Internet Society, 2025

    2025

  6. [6]

    LLM-hRIC: LLM- empowered Hierarchical RAN Intelligent Control for O-RAN,

    L. Bao, S. Yun, J. Lee, and T. Q. S. Quek, “LLM-hRIC: LLM- empowered Hierarchical RAN Intelligent Control for O-RAN,”arXiv preprint arXiv:2504.18062, 2025. [Online]. Available: https://arxiv.org/ abs/2504.18062

    arXiv 2025

  7. [7]

    Intent-Based Network for RAN Management with Large Language Models,

    F. A. Bimoet al., “Intent-Based Network for RAN Management with Large Language Models,”arXiv preprint arXiv:2507.14230, 2025. [Online]. Available: https://arxiv.org/abs/2507.14230

    Pith/arXiv arXiv 2025

  8. [8]

    Language Models Need Inductive Biases to Count Inductively,

    Y . Chang and Y . Bisk, “Language Models Need Inductive Biases to Count Inductively,” inProceedings of the International Conference on Learning Representations (ICLR), 2025

    2025

  9. [9]

    Scheduled Sampling for Sequence Prediction with Re- current Neural Networks,

    S. Bengioet al., “Scheduled Sampling for Sequence Prediction with Re- current Neural Networks,” inAdvances in Neural Information Processing Systems (NeurIPS), 2015

    2015

  10. [10]

    Professor Forcing: A New Algorithm for Training Recurrent Networks,

    A. Lambet al., “Professor Forcing: A New Algorithm for Training Recurrent Networks,” inNeurIPS, 2016

    2016

  11. [11]

    Off-Policy Deep Reinforcement Learning without Exploration,

    S. Fujimoto, D. Meger, and D. Precup, “Off-Policy Deep Reinforcement Learning without Exploration,” inProceedings of the 36th International Conference on Machine Learning (ICML), 2019

    2019

  12. [12]

    Time Series Foundation Models: Benchmarking Challenges and Requirements,

    M. Meyeret al., “Time Series Foundation Models: Benchmarking Challenges and Requirements,”arXiv preprint arXiv:2510.13654, 2025

    arXiv 2025

  13. [13]

    Fidel-TS: A High-Fidelity Benchmark for Multimodal Time Series Forecasting,

    Z. Xuet al., “Fidel-TS: A High-Fidelity Benchmark for Multimodal Time Series Forecasting,”arXiv preprint arXiv:2509.24789, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Counterfactual Dynamics Forecasting — A New Setting of Spatio-Temporal Forecasting,

    Y . Liuet al., “Counterfactual Dynamics Forecasting — A New Setting of Spatio-Temporal Forecasting,” inProceedings of the AAAI Conference on Artificial Intelligence, 2023

    2023

  15. [15]

    Understanding Physical Dynamics with Counter- factual World Modeling,

    R. Venkateshet al., “Understanding Physical Dynamics with Counter- factual World Modeling,” inEuropean Conference on Computer Vision (ECCV), 2024. [Online]. Available: https://arxiv.org/abs/2312.06721

    Pith/arXiv arXiv 2024

  16. [16]

    Simplified State Space Layers for Sequence Modeling,

    J. T. H. Smithet al., “Simplified State Space Layers for Sequence Modeling,” inICLR, 2023

    2023

  17. [17]

    Rivaling Transformers: Multi-Scale Structured State-Space Mixtures for Agentic 6G O-RAN,

    F. Rezazadeh, H. Chergui, M. Debbah, H. Song, D. Niyato, and L. Liu, “Rivaling Transformers: Multi-Scale Structured State-Space Mixtures for Agentic 6G O-RAN,” 2025. [Online]. Available: https://arxiv.org/abs/ 2510.05255

    Pith/arXiv arXiv 2025

  18. [18]

    Stochastic Backpropaga- tion and Approximate Inference in Deep Generative Models,

    D. J. Rezende, S. Mohamed, and D. Wierstra, “Stochastic Backpropaga- tion and Approximate Inference in Deep Generative Models,” inICML, 2014

    2014

  19. [19]

    What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?

    A. Kendall and Y . Gal, “What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?” inNeurIPS, 2017

    2017

  20. [20]

    J. B. Rawlings, D. Q. Mayne, and M. M. Diehl,Model Predictive Control: Theory, Computation, and Design, 2nd ed. Nob Hill Publishing,

  21. [21]

    The Cross-Entropy Method for Combinatorial and Continuous Optimization,

    R. Y . Rubinstein, “The Cross-Entropy Method for Combinatorial and Continuous Optimization,”Methodology and Computing in Applied Probability, vol. 1, no. 2, pp. 127–190, 1999

    1999

  22. [22]

    Sample-Efficient and Smooth Cross-Entropy Method Model Predictive Control Using Deterministic Samples,

    M. Walker, D. Frisch, and U. D. Hanebeck, “Sample-Efficient and Smooth Cross-Entropy Method Model Predictive Control Using Deterministic Samples,”arXiv preprint arXiv:2510.05706, 2025. [Online]. Available: https://arxiv.org/pdf/2510.05706

    Pith/arXiv arXiv 2025

  23. [23]

    RWKV: Reinventing RNNs for the Transformer Era,

    B. Penget al., “RWKV: Reinventing RNNs for the Transformer Era,” arXiv preprint arXiv:2305.13048, 2023

    Pith/arXiv arXiv 2023

  24. [24]

    Rethinking Attention with Performers,

    K. Choromanskiet al., “Rethinking Attention with Performers,” inICLR, 2021

    2021

  25. [25]

    Retentive Network: A Successor to Transformer for Large Language Models,

    Y . Sunet al., “Retentive Network: A Successor to Transformer for Large Language Models,”arXiv preprint arXiv:2307.08621, 2023

    Pith/arXiv arXiv 2023

  26. [26]

    Chronos: Learning the Language of Time Series,

    A. F. Ansariet al., “Chronos: Learning the Language of Time Series,” Transactions on Machine Learning Research, 2024

    2024

  27. [27]

    HiPPO: Recurrent Memory with Optimal Polynomial Projections,

    A. Guet al., “HiPPO: Recurrent Memory with Optimal Polynomial Projections,” inNeurIPS, 2020

    2020

  28. [28]

    Dreamer: Reinforcement Learning with World Models,

    D. Hafneret al., “Dreamer: Reinforcement Learning with World Models,” inICLR, 2020

    2020

  29. [29]

    Mastering Atari with Discrete World Models,

    D. Hafner, T. Lillicrapet al., “Mastering Atari with Discrete World Models,” inICLR, 2021

    2021

  30. [30]

    Dyna, an Integrated Architecture for Learning, Planning, and Reacting,

    R. S. Sutton, “Dyna, an Integrated Architecture for Learning, Planning, and Reacting,”ACM SIGART Bulletin, vol. 2, no. 4, pp. 160–163, 1991

    1991

  31. [31]

    Generalized Kullback-Leibler Divergence Loss,

    J. Cuiet al., “Generalized Kullback-Leibler Divergence Loss,”arXiv preprint arXiv:2503.08038, 2025. [Online]. Available: https://arxiv.org/ abs/2503.08038

    Pith/arXiv arXiv 2025

  32. [32]

    (2023) O-RAN Architecture Overview

    O-RAN Software Community. (2023) O-RAN Architecture Overview. [Online]. Available: https://docs.o-ran-sc.org/en/latest/architecture/ architecture.html

    2023

  33. [33]

    ETSI TS 104 040 V4.0.0: Publicly Available Specification (PAS); E2 interface: RAN Function-specific Service Models (O-RAN E2SMs),

    ETSI, “ETSI TS 104 040 V4.0.0: Publicly Available Specification (PAS); E2 interface: RAN Function-specific Service Models (O-RAN E2SMs),” European Telecommunications Standards Institute (ETSI), Tech. Rep. TS 104 040 V4.0.0, 2024. [Online]. Available: https://www.etsi.org/deliver/ etsi ts/104000 104099/104040/04.00.00 60/ts 104040v040000p.pdf

    2024

  34. [34]

    Squeeze-and-Excitation Networks,

    J. Huet al., “Squeeze-and-Excitation Networks,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 7132–7141

    2018

  35. [35]

    Language Modeling with Gated Convolutional Networks,

    Y . N. Dauphinet al., “Language Modeling with Gated Convolutional Networks,” inICML, 2017, pp. 933–941

    2017

  36. [36]

    Autoregressive Models: What Are They Good For?

    M. Dalalet al., “Autoregressive Models: What Are They Good For?” in NeurIPS 2019 Workshop on Information Theory and Machine Learning (ITML), 2019

    2019

  37. [37]

    C. P. Robert and G. Casella,Monte Carlo Statistical Methods, 2nd ed. Springer, 2004

    2004

  38. [38]

    Deep Learning for Time Series Forecasting: Tutorial and Survey,

    K. Benidiset al., “Deep Learning for Time Series Forecasting: Tutorial and Survey,”ACM Computing Surveys, vol. 55, no. 6, pp. 1–36, 2022

    2022

  39. [39]

    DeepAR: Probabilistic forecasting with autoregressive recurrent networks,

    D. Salinaset al., “DeepAR: Probabilistic forecasting with autoregressive recurrent networks,”International Journal of Forecasting, vol. 36, no. 3, pp. 1181–1191, 2020

    2020

  40. [40]

    J. P. Hespanha,Linear Systems Theory, 2nd ed. Princeton, NJ: Princeton University Press, 2018

    2018

  41. [41]

    Preservation of Common Quadratic Lyapunov Functions and Pad ´e Approximations,

    S. Sajjaet al., “Preservation of Common Quadratic Lyapunov Functions and Pad ´e Approximations,” inProceedings of the 49th IEEE Conference on Decision and Control (CDC), Atlanta, GA, USA, 2010, pp. 7334– 7338

    2010

  42. [42]

    On Pad ´e Approximations, Quadratic Stability and Discretization of Switched Linear Systems,

    R. N. Shortenet al., “On Pad ´e Approximations, Quadratic Stability and Discretization of Switched Linear Systems,”Systems & Control Letters, vol. 60, no. 5, pp. 351–359, 2011

    2011

  43. [43]

    Auto-Encoding Variational Bayes,

    D. P. Kingma and M. Welling, “Auto-Encoding Variational Bayes,”ICLR, 2014

    2014

  44. [44]

    R. J. Hyndman and G. Athanasopoulos,Forecasting: Principles and Practice, 2nd ed. OTexts, 2018, https://otexts.com/fpp2/

    2018

  45. [45]

    An Experimental Reservoir-Augmented Foundation Model: 6G O-RAN Case Study,

    F. Rezazadeh, R. Zhao, J. Dai, A. Ashtari Gargari, H. Chergui, and L. Liu, “An Experimental Reservoir-Augmented Foundation Model: 6G O-RAN Case Study,” 2025, accepted to publish at GACLM 2025; arXiv:2508.07778

    Pith/arXiv arXiv 2025

  46. [46]

    Learning Low-Dimensional Representation for O-RAN Testing via Transformer- ESN,

    J. Dai, R. Zhao, F. Rezazadeh, L. Zheng, H. Wang, and L. Liu, “Learning Low-Dimensional Representation for O-RAN Testing via Transformer- ESN,” 2025, accepted for publication at IEEE MASS, 2025

    2025

  47. [2017]

    Available: https://sites.engineering.ucsb.edu/ ∼jbraw/mpc/ MPC-book-2nd-edition-1st-printing.pdf

    [Online]. Available: https://sites.engineering.ucsb.edu/ ∼jbraw/mpc/ MPC-book-2nd-edition-1st-printing.pdf