Pith. sign in

REVIEW 4 major objections 6 minor 47 references

A Bayesian free-energy agent controls mobile network slices more adaptably than deep RL.

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

2026-08-02 23:18 UTC pith:CSD3M2RQ

load-bearing objection A credible first application of active inference to O-RAN slicing with a real testbed, but the headline 28.3% robustness gain exists only in the abstract and the evaluation never actually demonstrates it. the 4 major comments →

arxiv 2602.14033 v1 pith:CSD3M2RQ submitted 2026-02-15 cs.IT cs.AImath.IT

BRAIN: Bayesian Reasoning via Active Inference for Agentic and Embodied Intelligence in Mobile Networks

Osman Tugay Basaran , Martin Maier , Falko Dressler This is my paper
classification cs.IT cs.AImath.IT
keywords active inferencefree energy minimizationO-RANnetwork slicingradio resource allocationexplainable AI6Gcontinual adaptation
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 sets out to show that active inference, a neuroscience-inspired framework in which an agent minimizes variational free energy to unify perception and action, can serve as the control "brain" for 6G radio access networks. It argues that deep reinforcement learning (DRL) agents used today for network slicing are brittle: they are black boxes, overfit to training conditions, and suffer catastrophic forgetting when traffic patterns shift. The proposed BRAIN agent instead maintains an explicit generative model of network slices, infers hidden traffic-demand states from real-time O-RAN measurements, and selects resource-allocation actions by minimizing expected free energy, balancing goal satisfaction against information gain. On a GPU-accelerated O-RAN testbed with eMBB, URLLC, and mMTC slices, the agent meets slice-specific QoS targets and recovers from a sudden traffic surge with up to 28.3% higher robustness than DRL baselines, without any retraining, while exposing interpretable belief states and free-energy decompositions.

Core claim

BRAIN is an active-inference agent implemented as an O-RAN xApp. At each control interval it receives O-RAN KPM reports (throughput, buffer occupancy, transport-block count), performs a Bayesian update of its posterior belief over hidden slice-demand states, then selects the physical-resource-block allocation and scheduler whose one-step expected free energy is smallest. The expected free energy is decomposed into an extrinsic term, the KL divergence between predicted outcomes and a hand-specified preference distribution encoding slice QoS targets, and an epistemic term, the expected information gain about hidden state. On a GPU-accelerated O-RAN testbed with eMBB, URLLC, and mMTC slices, BR

What carries the argument

The generative model is the central object: a joint distribution over hidden slice states, observations, and actions, with a transition model B, an observation model A, and a preference distribution over outcomes. The variational free energy F_t and expected free energy G(a) carry the argument: minimizing F_t performs perception (Bayesian belief updating), and minimizing G(a) selects actions by trading extrinsic preference-alignment against epistemic information gain. Named in the paper, the 'expected free energy (EFE) decomposition' is what makes decisions auditable.

Load-bearing premise

The whole inference loop presupposes that the O-RAN KPM reports (throughput, buffer occupancy, transport-block count) are a sufficient statistic of the true network condition, so that no unobserved factor such as channel quality or interference can silently mislead the state estimate.

What would settle it

Build a scenario where two different hidden conditions (e.g., high traffic demand vs. poor channel quality) produce identical KPM reports over several control intervals. If BRAIN's posterior belief and chosen action cannot distinguish them and QoS degrades, the sufficient-statistic premise fails. Concretely: induce an interference event while holding reported metrics constant, and observe whether the agent's allocation diverges from an oracle that sees the true state.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Near-real-time RIC controllers can host agents that keep heterogeneous slice QoS targets under changing traffic without episodic retraining, reducing operational cost.
  • Network AI decisions become auditable: operators can inspect posterior beliefs over slice demand states and the extrinsic/epistemic terms behind each action.
  • Exploration–exploitation emerges from the inference objective itself, so agents do not need crafted reward signals or explicit exploration schedules.
  • If the approach generalizes, the same active-inference loop can be applied to other O-RAN control tasks, such as power control, handover, or energy saving.

Where Pith is reading between the lines

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

  • The hand-coded generative model is domain-specific; scaling to city-wide networks would likely require learning the A and B matrices from data, a hard inference problem the paper does not address.
  • The 28.3% robustness figure compares against one non-stationary event on one testbed; a broader distribution of shifts is needed to know whether the advantage is structural or scenario-specific.
  • Because active inference subsumes reward-based control through preference distributions, a testable extension is to derive the preference model directly from operator service-level agreements and verify the agent follows them under multi-cell interference.
  • The sufficient-statistic assumption about KPM reports could be probed by feeding the agent augmented state (e.g., channel quality indicator) and checking whether beliefs and actions change materially.

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 introduces BRAIN, an active-inference agent for O-RAN network slicing that maintains a hand-coded deep generative model of the RAN, performs variational Bayesian belief updating over latent slice-demand states, and selects resource-allocation actions by minimizing expected free energy. The authors claim that BRAIN outperforms tuned heuristic and DRL baselines (DQN, A2C, REINFORCE, PPO, SAC) in a GPU-accelerated testbed, achieving higher QoS satisfaction, faster adaptation to sudden traffic shifts (up to 28.3% higher robustness), and real-time interpretability via posterior beliefs and EFE decomposition. The paper presents the framework, implementation details, and qualitative results in Figures 4–8, but quantitative evidence for the headline claims is missing.

Significance. If fully substantiated, this would be a meaningful contribution: it is one of the first demonstrations of active inference for closed-loop RAN control on a real O-RAN/AI-RAN testbed, with a plausible case for intrinsic explainability and non-stationarity robustness. The deployment as an xApp and the comparison against a wide set of DRL baselines are both valuable. However, the current evidence is insufficient to support the central empirical claims: the abstract's 28.3% robustness figure is not reproduced in Section VI, no confidence intervals or per-seed statistics are reported despite promises in Section V-C, and the 'explainability' is largely a restatement of the agent's internal computation. The paper's strengths are its real-world testbed and the principled formulation; its weaknesses are the missing quantitative evaluation and the unvalidated generative-model assumptions.

major comments (4)
  1. [§VI, Fig. 8, Abstract] The headline claim of 'up to 28.3% higher robustness' appears only in the Abstract and is never derived in the evaluation. Section V-C states that 5 independent runs and 95% confidence intervals are reported, but Figures 4, 7, and 8 show single curves with no error bands, and no table provides means, variances, or per-seed values. Without a quantitative comparison, the central 'superior adaptability' conclusion is unsupported; the observed differences could lie within run-to-run variance. Please report the actual robustness numbers, the QoS-satisfaction values before/after the shift, and appropriate statistical tests.
  2. [§IV-D, §IV-C, Algorithm 1] The explainability claim is introspective and circular in the sense that the quantities used to explain decisions (posterior beliefs Q(s_t) and EFE components) are exactly the quantities used to select actions in Eq. (8) and Algorithm 1. This is not a causal explanation validated independently: it merely restates the agent's decision rule. To substantiate 'human-interpretable' and 'causal' explanations, the paper should validate against ground-truth causes (e.g., injected traffic-demand changes), compare against an independent explainer, or show an ablation where the explanation actually predicts the action's effect. Otherwise, it is just a visualization of the algorithm's internal variables.
  3. [§III-A, §IV-A, §V-C] The correctness of the agent hinges on two assumptions that are not tested: (i) the O-RAN KPM reports are a sufficient statistic of the network's condition (stated in §III-A), and (ii) the hand-coded generative model—transition matrix B, observation matrix A, and preference distribution P_pref—accurately represents the real RAN environment. No learning of A or B from data is described, and the paper mentions sensitivity analyses in §V-C but does not report them in Section VI. If the generative model is misspecified (e.g., hidden states do not capture channel quality or mobility), the EFE-based action selection can be systematically wrong. Please provide a sensitivity study or a description of how A, B, and P_pref were calibrated and validated.
  4. [§V-C, Fig. 8] The non-stationarity experiment is not fully specified. It is unclear whether the DRL baselines were allowed to continue online learning after the shift (which the paper's own 'catastrophic forgetting' argument suggests they would need to do), or whether they were frozen. The comparison is only meaningful if the baselines are treated fairly: if they are kept in training mode, the drop/recovery in Fig. 8 may reflect their ability to adapt; if frozen, it reflects only their initial policy. The paper should state the exact training status of each baseline during the stress test and report the recovery time and post-shift QoS satisfaction quantitatively.
minor comments (6)
  1. [§VI, Fig. 7] The CDF plots show qualitative curves but no legend for line styles; color alone is used, which will not be accessible in grayscale. Consider adding markers or different line styles.
  2. [§IV-B, Eq. (10)] The KL divergence in Eq. (10) has a sum from t=0 to infinity over observations, which is notationally incorrect for a single-step divergence. It should be a sum over observation outcomes o, not over time.
  3. [§III-A] Typo: 'massive machine-yype communications' should be 'massive machine-type communications'.
  4. [§II, References] Reference [44] is listed as 'Deep Reinforcement Learning with Double Q-learning' but is cited for A2C. The A2C baseline is missing an explicit reference (e.g., the asynchronous advantage actor-critic paper).
  5. [§V-C] Typo: 'Tranining' should be 'Training'. Also, the sentence 'We compare both convergence speed in terms of training iterations and actual wall-clock time' is incomplete—the actual wall-clock comparison is not reported in Section VI.
  6. [§I, Fig. 1] The preliminary experiment on catastrophic forgetting is described qualitatively; if it is based on an actual run, the setup (environment, DRL algorithm, hyperparameters) should be specified, or it should be clearly labeled as an illustrative sketch.

Circularity Check

0 steps flagged

No significant circularity: the BRAIN derivation is self-contained; only minor motivational self-citations appear, and the headline robustness figure is a missing-support concern rather than a circular step.

full rationale

The core derivation (Algorithm 1, Eqs. 3-11) is not circular: BRAIN selects actions by minimizing G(a_t) = KL_pref - I_gain computed from an explicitly specified generative model, and the performance comparison against DRL baselines is externally measured on a GPU-accelerated testbed, so the empirical claims are not fitted inputs renamed as predictions. The only self-citations ([16] Maier; [23] Basaran & Dressler) are motivational or related-work and are not load-bearing for the derivation. The claimed 'introspective explainability' uses the same quantities (Q(s_t), G components) that determine the action; this is algorithmic transparency rather than a circular derivation, because it does not claim an independent prediction from those quantities. The 'up to 28.3% higher robustness' figure appears only in the abstract and is not substantiated by quantitative tables, confidence intervals, or per-seed results in Section VI; that is a missing-support/correctness concern, not a circularity, and does not raise the circularity score.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central performance comparison is measured, but the agent's behavior is driven entirely by hand-specified beliefs and preferences. The paper's explainability is generated from the same model that selects actions, so it is not an independent check on the decision process. The absence of learned A and B matrices makes the approach fragile if the model is misspecified.

free parameters (5)
  • Reward weights α, β, γ = not reported
    Eq. 1 defines reward as α·eMBB throughput − β·URLLC buffer + γ·mMTC TB count; these weights encode slice priorities and are hand-chosen.
  • Precision γ for action posterior = not reported
    Eq. 15 uses softmax(−γ G(a|s_t)); γ is an inverse-temperature parameter controlling decisiveness, set by hand.
  • Preference distribution P_pref(o) = not reported
    Section IV-C/D defines P_pref over observations; it encodes desired slice QoS targets and directly determines the extrinsic cost of actions.
  • Generative model matrices A and B = not reported
    Section III-B defines observation model A and transition model B; these are hand-designed, with no learning procedure described in the paper.
  • Planning horizon H = one-step horizon (H=1)
    Section IV-C says 'for simplicity, with a one-step horizon π=a_t'; despite the later claim of sensitivity analysis, no multi-step horizons are evaluated.
axioms (4)
  • domain assumption KPM reports are a sufficient statistic of the network's condition
    Section III-A: 'we assume these reported metrics are a sufficient statistic of the network's condition'; if false, latent-state inference is ungrounded.
  • ad hoc to paper The hand-coded generative model P(s,o,a) adequately captures RAN dynamics
    Eq. 6 and Algorithm 1 use a fixed factorization with hand-chosen A and B; correctness of EFE-based action selection depends on this assumption.
  • domain assumption Variational free-energy minimization is an appropriate objective for network control
    Borrowed from computational neuroscience; the paper does not prove that minimizing expected free energy yields QoS-optimal or even safe control policies.
  • domain assumption The single-cell testbed with two UEs is representative of 6G non-stationary environments
    Conclusions generalize to multi-cell, multi-agent 6G networks, but only a single 100 MHz cell with two UEs is demonstrated.
invented entities (2)
  • Latent traffic-demand states (Low/Medium/High) per slice no independent evidence
    purpose: Explainable state representation used for belief updates and EFE explainability
    These latent states are a modeling choice visualized in Fig. 5; the paper provides no external ground-truth validation that they correspond to real traffic-demand levels.
  • 'Check' (information-gathering) actions no independent evidence
    purpose: Explicit exploratory actions that sharpen belief states
    Mentioned only in the Fig. 5 caption; not defined in the action-space description in Section V-B, and their effects are not evaluated independently.

pith-pipeline@v1.3.0-alltime-deepseek · 20347 in / 9111 out tokens · 88633 ms · 2026-08-02T23:18:18.921135+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of BRAIN: Bayesian Reasoning via Active Inference for Agentic and Embodied Intelligence in Mobile Networks." pith.science (2026). https://pith.science/paper/CSD3M2RQ

@misc{pith2026260214033,
  author       = {Pith},
  title        = {Pith review of: BRAIN: Bayesian Reasoning via Active Inference for Agentic and Embodied Intelligence in Mobile Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CSD3M2RQ}},
  note         = {Machine review of arXiv:2602.14033}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Future sixth-generation (6G) mobile networks will demand artificial intelligence (AI) agents that are not only autonomous and efficient, but also capable of real-time adaptation in dynamic environments and transparent in their decisionmaking. However, prevailing agentic AI approaches in networking, exhibit significant shortcomings in this regard. Conventional deep reinforcement learning (DRL)-based agents lack explainability and often suffer from brittle adaptation, including catastrophic forgetting of past knowledge under non-stationary conditions. In this paper, we propose an alternative solution for these challenges: Bayesian reasoning via Active Inference (BRAIN) agent. BRAIN harnesses a deep generative model of the network environment and minimizes variational free energy to unify perception and action in a single closed-loop paradigm. We implement BRAIN as O-RAN eXtended application (xApp) on GPU-accelerated testbed and demonstrate its advantages over standard DRL baselines. In our experiments, BRAIN exhibits (i) robust causal reasoning for dynamic radio resource allocation, maintaining slice-specific quality of service (QoS) targets (throughput, latency, reliability) under varying traffic loads, (ii) superior adaptability with up to 28.3% higher robustness to sudden traffic shifts versus benchmarks (achieved without any retraining), and (iii) real-time interpretability of its decisions through human-interpretable belief state diagnostics.

Figures

Figures reproduced from arXiv: 2602.14033 by Falko Dressler, Martin Maier, Osman Tugay Basaran.

Figure 2
Figure 2. Figure 2: Architecture of conventional DRL and proposed explainable deep [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 1
Figure 1. Figure 1: Preliminary experiment on catastrophic forgetting issue for DRL [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of GPU-Accelerated Testbed. the Kullback–Leibler (KL) divergence between the predicted outcome distribution and the preferred distribution. Formally, for action at: DKL Q(ot+1 | at) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Training dynamics of agentic (RL/DRL) and embodied (active-inference) Agents on AI-RAN testbed: [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Agent’s posterior belief trajectory over hidden traffic demand levels ( [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Interpretation of the EFE (G(π)) decomposition. At each time step, the chosen action’s epistemic (soft green) and extrinsic (soft red) terms presented. Epistemic value dominates early on (favoring exploration) and then gives way to extrinsic value (favoring QoS/exploitation). TABLE I HYPERPARAMETERS OF BASELINE RL MODELS. Model Learning R. Discount F. Replay Buffer Deep Q-Network (DQN) 1 × 10−3 0.99 105 ex… view at source ↗
Figure 7
Figure 7. Figure 7: Comparative network slicing performance of the proposed [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Resilience to a controlled non-stationarity event (traffic distribution [PITH_FULL_IMAGE:figures/full_fig_p009_8.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

47 extracted references · 3 linked inside Pith

  1. [1]

    Embodied Intelligence: A Synergy of Morphology, Action, Perception and Learning,

    H. Liu, D. Guo, and A. Cangelosi, “Embodied Intelligence: A Synergy of Morphology, Action, Perception and Learning,”ACM Computing Surveys, vol. 57, no. 7, pp. 1–36, Mar. 2025

    2025

  2. [2]

    Artificial General Intelligence (AGI)-Native Wireless Systems: A Journey Beyond 6G,

    W. Saad, O. Hashash, C. K. Thomas, C. Chaccour, M. Debbah, N. Mandayam, and Z. Han, “Artificial General Intelligence (AGI)-Native Wireless Systems: A Journey Beyond 6G,”Proceedings of the IEEE, vol. 113, no. 9, pp. 849–887, Sep. 2025

    2025

  3. [3]

    Empowering the 6G Cellular Architecture with Open RAN,

    M. Polese, M. Dohler, F. Dressler, M. Erol-Kantarci, R. Jana, R. Knopp, and T. Melodia, “Empowering the 6G Cellular Architecture with Open RAN,”IEEE Journal on Selected Areas in Communications, vol. 42, no. 2, pp. 245–262, Feb. 2024

    2024

  4. [4]

    Channel Nonstationarity and Consistency for Beyond 5G and 6G: A Survey,

    X. Cheng, Z. Huang, and L. Bai, “Channel Nonstationarity and Consistency for Beyond 5G and 6G: A Survey,”IEEE Communications Surveys & Tutorials, vol. 24, no. 3, pp. 1634–1669, 2022

    2022

  5. [5]

    A Vision of 6G Wireless Systems: Applications, Trends, Technologies, and Open Research Problems,

    W. Saad, M. Bennis, and M. Chen, “A Vision of 6G Wireless Systems: Applications, Trends, Technologies, and Open Research Problems,” IEEE Network, vol. 34, no. 3, pp. 134–142, May 2020

    2020

  6. [6]

    Convolutional Networks for Images, Speech, and Time Series,

    Y . LeCun and Y . Bengio, “Convolutional Networks for Images, Speech, and Time Series,” inHandbook of Brain Theory and Neural Networks, M. A. Arbib, Ed., MIT Press, 2003, pp. 255–258

    2003

  7. [7]

    Autoencoders,

    D. Bank, N. Koenigstein, and R. Giryes, “Autoencoders,” arXiv, cs.NI 2003.05991, Apr. 2021

    Pith/arXiv arXiv 2003

  8. [8]

    Causal Reasoning: Charting a Revolutionary Course for Next- Generation AI-Native Wireless Networks,

    C. K. Thomas, C. Chaccour, W. Saad, M. Debbah, and C. S. Hong, “Causal Reasoning: Charting a Revolutionary Course for Next- Generation AI-Native Wireless Networks,”IEEE Vehicular Technology Magazine, vol. 19, no. 1, pp. 16–31, Mar. 2024

    2024

  9. [9]

    R. S. Sutton and A. G. Barto,Reinforcement Learning: An Introduc- tion. Cambridge, MA: MIT Press, 1998, p. 322

    1998

  10. [10]

    An Intro- duction to Deep Reinforcement Learning,

    F.-L. Vincent, H. Peter, I. Riashat, G. B. Marc, and P. Joelle, “An Intro- duction to Deep Reinforcement Learning,”Foundations and Trends® in Machine Learning, vol. 11, no. 3–4, pp. 219–354, Dec. 2018

    2018

  11. [11]

    OnSlicing: Online End-to-End Network Slicing with Reinforcement Learning,

    Q. Liu, N. Choi, and T. Han, “OnSlicing: Online End-to-End Network Slicing with Reinforcement Learning,” in17th ACM International Conference on Emerging Networking Experiments and Technologies (CoNEXT 2021), Virtual Conference: ACM, Dec. 2021, pp. 141–153

    2021

  12. [12]

    ORANSlice: An Open Source 5G Network Slicing Platform for O-RAN,

    H. Cheng, S. D’Oro, R. Gangula, S. Velumani, D. Villa, L. Bonati, M. Polese, T. Melodia, G. Arrobo, and C. Maciocco, “ORANSlice: An Open Source 5G Network Slicing Platform for O-RAN,” in30th ACM International Conference on Mobile Computing and Networking (MobiCom 2024), Washington, D.C.: ACM, Nov. 2024, pp. 2297–2302

    2024

  13. [13]

    Explainable Deep Reinforcement Learning: State of the Art and Challenges,

    G. A. V ouros, “Explainable Deep Reinforcement Learning: State of the Art and Challenges,”ACM Computing Surveys, vol. 55, no. 5, pp. 1–39, Dec. 2022

    2022

  14. [14]

    Generalization- Enhanced DRL-Based Resource Allocation in Wireless Communica- tion Networks With Dynamic User Loads,

    X. Yang, Y . Shi, J. Liu, Z. Xie, M. Sheng, and J. Li, “Generalization- Enhanced DRL-Based Resource Allocation in Wireless Communica- tion Networks With Dynamic User Loads,”IEEE Wireless Communi- cations Letters, vol. 14, no. 12, pp. 3902–3906, Dec. 2025

    2025

  15. [15]

    AI Embodiment Through 6G: Shaping the Future of AGI,

    L. Bariah and M. Debbah, “AI Embodiment Through 6G: Shaping the Future of AGI,”IEEE Wireless Communications, vol. 31, no. 5, pp. 174–181, Oct. 2024

    2024

  16. [16]

    From artificial intelligence to active inference: the key to true AI and the 6G world brain [Invited],

    M. Maier, “From artificial intelligence to active inference: the key to true AI and the 6G world brain [Invited],”Journal of Optical Communications and Networking, vol. 18, no. 1, A28–A43, Jan. 2026

    2026

  17. [17]

    Active inference and learning,

    K. J. Friston, T. FitzGerald, F. Rigoli, P. Schwartenbeck, J. ODoherty, and G. Pezzulo, “Active inference and learning,”Neuroscience & Biobehavioral Reviews, vol. 68, pp. 862–879, Sep. 2016

    2016

  18. [18]

    Explainable Artificial Intelligence (XAI): What we know and what is left to attain Trustworthy Artificial Intelligence,

    S. Ali, T. Abuhmed, S. El-Sappagh, K. Muhammad, J. M. Alonso- Moral, R. Confalonieri, R. Guidotti, J. Del Ser, N. D ´ıaz-Rodr´ıguez, and F. Herrera, “Explainable Artificial Intelligence (XAI): What we know and what is left to attain Trustworthy Artificial Intelligence,” Elsevier Information Fusion, vol. 99, p. 101 805, Nov. 2023

    2023

  19. [19]

    Explainable Artificial Intelligence (XAI) 2.0: A manifesto of open challenges and interdisciplinary research directions,

    L. Longo, M. Brcic, F. Cabitza, J. Choi, R. Confalonieri, J. D. Ser, R. Guidotti, Y . Hayashi, F. Herrera, A. Holzinger, R. Jiang, H. Khosravi, F. Lecue, G. Malgieri, A. P ´aez, W. Samek, J. Schneider, T. Speith, and S. Stumpf, “Explainable Artificial Intelligence (XAI) 2.0: A manifesto of open challenges and interdisciplinary research directions,”Elsevie...

    2024

  20. [20]

    Explainable Artificial Intelligence for 6G: Improving Trust between Human and Machine,

    W. Guo, “Explainable Artificial Intelligence for 6G: Improving Trust between Human and Machine,”IEEE Communications Magazine, vol. 58, no. 6, pp. 39–45, Jun. 2020

    2020

  21. [21]

    A Unified Approach to Interpreting Model Predictions,

    S. M. Lundberg and S.-I. Lee, “A Unified Approach to Interpreting Model Predictions,” in31st International Conference on Neural In- formation Processing Systems (NIPS 2017), Long Beach, CA: Curran Associates Inc., Dec. 2017, pp. 4768–4777

    2017

  22. [22]

    “Why Should I Trust You?

    M. T. Ribeiro, S. Singh, and C. Guestrin, ““Why Should I Trust You?”: Explaining the Predictions of Any Classifier,” in22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA: ACM, Aug. 2016, pp. 1135 –1144

    2016

  23. [23]

    XAInomaly: Explainable, Interpretable and Trustworthy AI for xURLLC in 6G Open-RAN,

    O. T. Basaran and F. Dressler, “XAInomaly: Explainable, Interpretable and Trustworthy AI for xURLLC in 6G Open-RAN,” in3rd Inter- national Conference on 6G Networking (6GNet 2024), Paris, France: IEEE, Oct. 2024, pp. 93–101

    2024

  24. [24]

    EXPLORA: AI/ML EXPLainability for the Open RAN,

    C. Fiandrino, L. Bonati, S. D’Oro, M. Polese, T. Melodia, and J. Widmer, “EXPLORA: AI/ML EXPLainability for the Open RAN,” Proceedings of the ACM on Networking, vol. 1, pp. 1–26, Nov. 2023

    2023

  25. [25]

    Towards Explainable AI for Channel Estimation in Wireless Communications,

    A. K. Gizzini, Y . Medjahdi, A. J. Ghandour, and L. Clavier, “Towards Explainable AI for Channel Estimation in Wireless Communications,” IEEE Transactions on Vehicular Technology, vol. 73, no. 5, pp. 7389– 7394, May 2024

    2024

  26. [26]

    ChronoProf: Profiling Time Series Forecasters and Classifiers in Mobile Networks with Explainable AI,

    P. F. P ´erez, I. Bravo, A. Kamath, C. Fiandrino, and J. Widmer, “ChronoProf: Profiling Time Series Forecasters and Classifiers in Mobile Networks with Explainable AI,” in26th IEEE International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM 2025), Fort Worth, TX: IEEE, May 2025, pp. 41–50

    2025

  27. [27]

    AIChronoLens: Advancing Explainability for Time Series AI Forecasting in Mobile Networks,

    C. Fiandrino, E. P. G ´omez, P. F. P ´erez, H. Mohammadalizadeh, M. Fiore, and J. Widmer, “AIChronoLens: Advancing Explainability for Time Series AI Forecasting in Mobile Networks,” in43rd IEEE International Conference on Computer Communications (INFOCOM 2024), Vancouver, Canada: IEEE, May 2024, pp. 1521–1530

    2024

  28. [28]

    SYMBXRL: Symbolic Explainable Deep Reinforcement Learning for Mobile Networks,

    A. Duttagupta, M. Jabbari, C. Fiandrino, M. Fiore, and J. Widmer, “SYMBXRL: Symbolic Explainable Deep Reinforcement Learning for Mobile Networks,” in44th IEEE International Conference on Com- puter Communications (INFOCOM 2025), London, United Kingdom: IEEE, May 2025, pp. 1–10

    2025

  29. [29]

    Active inference on discrete state-spaces: A synthesis,

    L. Da Costa, T. Parr, N. Sajid, S. Veselic, V . Neacsu, and K. J. Friston, “Active inference on discrete state-spaces: A synthesis,”Journal of Mathematical Psychology, vol. 99, p. 102 447, Dec. 2020

    2020

  30. [30]

    Reinforce- ment Learning through Active Inference,

    A. Tschantz, B. Millidge, A. K. Seth, and C. L. Buckley, “Reinforce- ment Learning through Active Inference,” arXiv, cs.LG, Feb. 2020

    2020

  31. [31]

    An Empirical Study of Active Inference on a Humanoid Robot,

    G. Oliver, P. Lanillos, and G. Cheng, “An Empirical Study of Active Inference on a Humanoid Robot,”IEEE Transactions on Cognitive and Developmental Systems, vol. 14, no. 2, pp. 462–471, Jun. 2022

    2022

  32. [32]

    Real-World Robot Control Based on Contrastive Deep Active Inference With Demonstrations,

    K. Fujii, T. Isomura, and S. Murata, “Real-World Robot Control Based on Contrastive Deep Active Inference With Demonstrations,”IEEE Access, vol. 12, pp. 172 343–172 357, 2024

    2024

  33. [33]

    Designing explainable arti- ficial intelligence with active inference: A framework for transparent introspection and decision-making,

    M. Albarracin, I. Hip ´olito, S. E. Tremblay, J. G. Fox, G. Ren ´e, K. J. Friston, and M. J. D. Ramstead, “Designing explainable arti- ficial intelligence with active inference: A framework for transparent introspection and decision-making,” arXiv, cs.AI, Jun. 2023

    2023

  34. [34]

    T. Parr, G. Pezzulo, and K. J. Friston,Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. The MIT Press, 2022

    2022

  35. [35]

    The free-energy principle: a unified brain theory?

    K. J. Friston, “The free-energy principle: a unified brain theory?” Nature Reviews Neuroscience, vol. 11, no. 2, pp. 127–138, Jan. 2010

    2010

  36. [36]

    Aerial: A GPU Hyperconverged Platform for 5G,

    A. Kelkar and C. Dick, “Aerial: A GPU Hyperconverged Platform for 5G,” inACM SIGCOMM 2021, Demo Session, Virtual Conference: ACM, Aug. 2021, pp. 79–81

    2021

  37. [37]

    NVIDIA Aerial GPU Hosted AI-on-5G,

    A. Kelkar and C. Dick, “NVIDIA Aerial GPU Hosted AI-on-5G,” in 4th IEEE 5G World Forum (5GWF 2021), Montr ´eal, Canada: IEEE, Oct. 2021, pp. 64–69

    2021

  38. [38]

    NVIDIA Aerial SDK

    NVIDIA. “NVIDIA Aerial SDK. ”[Online]. Available: https : / / developer.nvidia.com/aerial-sdk (accessed: 15.01.2026)

    2026

  39. [39]

    OpenAirInterface: Democratizing innovation in the 5G Era,

    F. Kaltenberger, A. P. Silva, A. Gosain, L. Wang, and T.-T. Nguyen, “OpenAirInterface: Democratizing innovation in the 5G Era,”Elsevier Computer Networks, vol. 176, p. 107 284, Jul. 2020

    2020

  40. [40]

    Foxconn RPQN

    Foxconn. “Foxconn RPQN. ”[Online]. Available: https://fcc.report/ FCC-ID/2AQ68RPQN7801/5573870.pdf (accessed: 15.01.2026)

    2026

  41. [41]

    OSC Near Realtime RIC

    O-RAN SC. “OSC Near Realtime RIC. ”[Online]. Available: https: //wiki.o-ran-sc.org/display/RICP/2022-05-24+Release+E (accessed: 15.01.2026)

    2022

  42. [42]

    ColO- RAN: Developing Machine Learning-Based xApps for Open RAN Closed-Loop Control on Programmable Experimental Platforms,

    M. Polese, L. Bonati, S. D’Oro, S. Basagni, and T. Melodia, “ColO- RAN: Developing Machine Learning-Based xApps for Open RAN Closed-Loop Control on Programmable Experimental Platforms,” IEEE Transactions on Mobile Computing, vol. 22, no. 10, pp. 5787– 5800, Oct. 2023

    2023

  43. [43]

    Playing Atari with Deep Reinforcement Learning,

    V . Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wierstra, and M. Riedmiller, “Playing Atari with Deep Reinforcement Learning,” arXiv, cs.LG 1312.5602, Dec. 2013

    Pith/arXiv arXiv 2013

  44. [44]

    Deep Reinforcement Learning with Double Q-learning,

    H. van Hasselt, A. Guez, and D. Silver, “Deep Reinforcement Learning with Double Q-learning,” in30th AAAI Conference on Artificial Intelligence (AAAI 2016), Phoenix, AZ: AAAI Press, Feb. 2016, pp. 2094–2100. 13

    2016

  45. [45]

    Simple statistical gradient-following algorithms for connectionist reinforcement learning,

    R. J. Williams, “Simple statistical gradient-following algorithms for connectionist reinforcement learning,”Machine Learning, vol. 8, no. 3–4, pp. 229–256, May 1992

    1992

  46. [46]

    Proximal Policy Optimization Algorithms,

    J. Schulman, R. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal Policy Optimization Algorithms,” arXiv, cs.LG, Jul. 2017, pp. 1–12

    2017

  47. [47]

    Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor,

    T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, “Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor,” arXiv, cs.LG 1801.01290, Aug. 2018. Osman Tugay Basaranis a 6G/AI Research Scien- tist at Prof. Dr.-Ing. Falko Dressler’s Telecommuni- cations Networks Group (TKN), School of Electrical Engineering and Comput...

    Pith/arXiv arXiv 2018