Pith. sign in

REVIEW 4 major objections 6 minor 19 references

A small local language model can safely write V2X scheduler policies if a validator and a fast controller keep it out of the radio loop.

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 · grok-4.5

2026-07-11 20:19 UTC pith:X5XGM7Q7

load-bearing objection Solid hierarchical template for small-LLM V2X policy generation; validity is real, high-density win is directional and under-powered, and the xApp shield does a lot of the safety work. the 4 major comments →

arxiv 2607.04290 v1 pith:X5XGM7Q7 submitted 2026-07-05 cs.NI cs.AI

Agentic-V2X: Small Language Model Agents for Deadline-Aware V2X Scheduling in 5G/6G Networks

Gerasimos Papanikolaou-Ntais , Alexandros Kaloxylos , Athanasios Kanavos This is my paper
classification cs.NI cs.AI
keywords 5G NRV2XO-RANrAppxAppsmall language modelsdeadline-aware schedulingns-3
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.

Large language models are too slow and too unpredictable to run a near-real-time packet scheduler, especially for mixed critical and background vehicle traffic. This paper proposes Agentic-V2X: a small local model acts only as a slow policy writer (every 10 seconds), producing a structured YAML document of service weights, bounds, and safety limits. A validator repairs or rejects the document, and a deterministic controller then adapts scheduler weights every 100 ms inside those bounds. Across 126 matched simulation runs that include teleoperated driving, cooperative awareness, HD maps, and sensor sharing, every generated policy stayed valid and executable. The adaptive path improved mean critical reliability over plain proportional fair at the densest load and offered useful medium-density latency and throughput trade-offs, yet it did not beat the strongest static expert policies in aggregate. The claim is therefore architectural: small models can serve as safe policy assistants for deadline-aware V2X scheduling when the fast loop remains deterministic and bounded.

Core claim

When a small local language model is confined to slow, schema-constrained policy generation and its outputs are validated, repaired, and executed by a bounded deterministic controller, it produces fully executable V2X scheduler policies that stay competitive at selected operating points (notably high-density critical reliability) without ever becoming a universal scheduler replacement.

What carries the argument

Agentic-V2X hierarchical control: an rApp-inspired small-LLM policy agent that emits a constrained YAML weight-and-bound document every 10 s, paired with a validator-plus-runtime-shield and a 100 ms xApp-like weight adapter that never exceeds the validated limits.

Load-bearing premise

The single-cell simulator with fixed service rates, three vehicle densities, seven seeds, a 10-second refresh, and a hand-crafted YAML schema is representative enough that the observed validity and density-specific gains would still hold in multi-cell hardware and real O-RAN interfaces.

What would settle it

Re-run the identical service mix and seeds with a multi-cell handover-capable 5G-LENA setup or a real Non-RT/Near-RT RIC and A1/E2 path; if policy validity falls below 100 percent or the high-density critical-reliability edge over proportional fair disappears under the same paired analysis, the central architectural claim fails.

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

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 Agentic-V2X, a hierarchical control architecture for deadline-aware 5G NR V2X scheduling in which a small local LLM (qwen2.5:7b) acts as a slow, rApp-inspired policy generator (10 s) that emits constrained YAML policies (priorities, weight bounds, safety constraints), while a deterministic xApp-like controller enforces validated policies via PF scheduler-weight adaptation every 100 ms in ns-3/5G-LENA/SUMO/ns3-ai. A validator with repair/fallback plus a runtime performance shield (rollback and directional UL/DL guards) is intended to keep stochastic LLM outputs out of the near-real-time loop. Over a full factorial of 126 matched-seed runs (3 densities × 7 seeds × 6 methods), the authors report 100% policy acceptance (210/210 adaptive updates), density-specific competitiveness (e.g., higher mean critical DC-PRR than PF at 30 vehicles), and honest paired Wilcoxon/bootstrap analysis showing that the adaptive method is not the best aggregate critical-reliability method. The central claim is architectural: a safe, executable small-LLM policy-generation design rather than a universally dominant scheduler.

Significance. If the result holds, the work is a useful and timely contribution to LLM-for-RAN research: it argues against putting LLMs in the near-real-time scheduling loop and instead demonstrates a practical O-RAN-inspired timescale separation with explicit validation, repair, fallback, and runtime shielding. Strengths that should be credited include the full 126-run matched-seed campaign, explicit policy-validity and safety-intervention metrics, paired statistical analysis that refuses aggregate dominance, and an auditable appendix (system/user prompts, validated YAML schema, and an accepted example policy). These make the paper more reproducible and more carefully framed than much of the adjacent LLM-agent networking literature. The significance is primarily architectural and methodological rather than a new scheduling algorithm or a claim of state-of-the-art reliability.

major comments (4)
  1. Abstract and §VI-B/Table VIII highlight that adaptive LLM-rApp/xApp obtains improved mean critical reliability over PF at 30 vehicles (65.4% vs 59.8%). Table XVIII shows this is a non-significant n=7 paired difference (+5.6 pp, Wilcoxon p=0.156, 95% bootstrap CI includes negative values, win rate 71.4%). Aggregate paired tests (Table XVI) are negative versus PF, static LLM, and static expert. The high-density mean is therefore an operating-point observation, not a statistically supported superiority claim. The Abstract and contribution language should be revised so that this point is not presented as established improvement without the paired-test caveat already present later in the paper.
  2. Table XIV reports 229.3 rollback rows per run plus ~84 directional UL/DL interventions per run for the adaptive method, while schema-level rejections are zero (Table VII). Algorithm 1 and §IV-D make clear that much of the safety and corrective behaviour is supplied by the deterministic runtime shield and xApp bounds rather than by the content of the LLM policy. The paper’s claim that the design is a “small-LLM-assisted” competitive architecture is still defensible, but §VI-E/§VII currently under-analyse how often and under what conditions the LLM policy is overridden. A load-bearing clarification is needed: either quantify the fraction of control time spent under the original LLM policy versus rolled-back/fallback policies, or temper language that attributes density-specific gains primarily to LLM policy generation.
  3. The evaluation isolates static LLM vs adaptive LLM-rApp/xApp and heuristic xApp vs adaptive, but does not cleanly ablate (i) LLM-generated bounds/baselines with the runtime shield disabled, or (ii) the same adaptive xApp rules driven by a fixed expert policy rather than LLM updates. Without that, it remains hard to attribute the 30-vehicle critical-DC-PRR mean and the 25-vehicle throughput/latency operating points to the LLM versus the shared deterministic machinery. Given that the central claim is architectural usefulness of the LLM policy layer, §V-C/§VI-E should either add a minimal ablation or explicitly state that current evidence supports the full stack more than the LLM content alone.
  4. Limitations (§VIII) correctly note single-cell simulation, three densities, seven seeds, one model, and fixed 10 s/100 ms cadences. These constraints are acceptable for a systems paper, but they interact with the weakest transfer assumption: that validity and density-specific competitiveness under a hand-crafted YAML schema transfer toward real multi-cell V2X and A1/E2 O-RAN deployments. The Discussion should more sharply separate what is demonstrated (safe executable policy generation in this ns-3 setup) from what is hypothesized (deployable rApp/xApp pattern), so the architectural claim is not over-read as deployment readiness.
minor comments (6)
  1. Table XIV footnote notes that static-LLM generation latency was not separately logged; reporting it would complete the practicality comparison with the adaptive path.
  2. Figure 4/5/6 captions and axis labels should state units and whether error bars are SEM over seeds (as claimed in the text) so the plots are self-contained.
  3. Terminology oscillates between “rApp-inspired/xApp-like” and shorter “LLM-rApp/xApp”; a single consistent short form after first definition would improve readability.
  4. In Table II and the service mix (Table VI), briefly justify why ToD UL is 16 Mbps while ToD DL is 0.048 Mbps for readers less familiar with 5GAA teleoperated-driving models.
  5. Appendix Listing A.4 shows ToD bounds [4,8] while safety_constraints.min_critical_weight is 5; a one-line note on how the validator reconciles baseline/bounds with min_critical_weight would avoid confusion.
  6. A few typos and hyphenation inconsistencies appear (e.g., “Y AML” spacing, “RELATEDWORK” heading concatenation in the source); a copy-edit pass is needed.

Circularity Check

0 steps flagged

No significant circularity: empirical architecture evaluation against independent external baselines and metrics, with no derivation that reduces to its own inputs by construction.

full rationale

This is a systems/architecture paper whose load-bearing claims are empirical simulation outcomes (critical DC-PRR, ToD UL DC-PRR, background throughput, p95 latency, policy acceptance counts) measured in ns-3/5G-LENA against external deterministic baselines (PF, static expert, balanced expert, heuristic xApp). Metrics are defined independently of LLM outputs (deadline-constrained reception ratios, violation rates as 1-DC-PRR, Jain fairness, throughput). The LLM produces structured YAML that is validated/repaired by a separate deterministic layer; 100% acceptance is an observed outcome under a constrained schema, not a fitted parameter later re-presented as a prediction, nor a self-definitional identity. There are no uniqueness theorems, no self-citation chains that force the central result, no ansatz smuggled via prior author work, and no renaming of a known empirical pattern as a first-principles derivation. Aggregate and density-specific comparisons are paired statistical tests over matched seeds; the paper itself states the adaptive method is not the best aggregate method. The evaluation is therefore self-contained against external benchmarks and does not exhibit the listed circularity patterns.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 2 invented entities

The work is an empirical systems architecture paper. Its load-bearing premises are standard simulation and O-RAN timescale assumptions plus several design choices (model, periods, schema bounds, service deadlines) that are fixed by the authors rather than derived. No new physical entities are postulated; the 'Agentic-V2X' name and the YAML policy schema are engineering constructs whose only evidence is the reported simulation campaign.

free parameters (6)
  • LLM update period (10 s) and xApp control period (100 ms)
    Fixed design choices that define the hierarchical split; results are not shown to be robust to other cadences.
  • Scheduler weight domain [1,10] and default critical min weight 5
    Hand-chosen bounds that constrain every generated policy and the action space of the xApp.
  • Service deadlines (ToD 20 ms, Awareness 100 ms, Sensor 500 ms, HDMap 1000 ms)
    Scenario design parameters used to define DC-PRR; not measured outcomes.
  • Vehicle densities 20/25/30 and fixed per-flow rates (e.g., ToD UL 16 Mbps)
    Chosen operating points that determine contention; aggregate load scales only by vehicle count.
  • qwen2.5:7b model selection and prompt/schema wording
    Single representative small model; validity and performance may depend on this choice and on the constrained prompt.
  • Runtime shield thresholds (3 consecutive degradation windows, 0.05 UL/DL DC-PRR gap)
    Hand-set safety parameters that trigger rollback and directional boosts.
axioms (4)
  • domain assumption O-RAN-style separation of non-real-time policy reasoning (seconds) from near-real-time actuation (tens–hundreds of ms) is a useful control architecture for RAN.
    Invoked throughout Introduction and Related Work via citation to Polese et al.; the implementation does not use real A1/E2.
  • domain assumption A PF-based scheduler with configurable per-UE UL/DL weights is an adequate substrate for evaluating service-level priority policies.
    System Model and Experimental Methodology; no alternative scheduling disciplines are compared.
  • domain assumption ns-3/5G-LENA + SUMO Manhattan grid + ns3-ai sufficiently captures the radio and mobility dynamics needed to rank the compared methods.
    Stated as the evaluation environment; Limitations explicitly notes transfer to hardware is unproven.
  • ad hoc to paper Constrained YAML schema plus deterministic repair/fallback and runtime performance shield make stochastic LLM outputs safe enough for the policy layer.
    Core safety argument of Sections IV-D and VII-E; supported by 100% acceptance in this campaign but not proven for open-ended prompts or other models.
invented entities (2)
  • Agentic-V2X hierarchical architecture (LLM-rApp policy agent + validator + xApp-like executor) no independent evidence
    purpose: To keep LLM inference out of the near-real-time scheduling loop while still using language models for service-objective reasoning.
    The named system and its two-timescale control loop are the paper's central engineering contribution; evidence is the simulation campaign only.
  • Structured rApp-inspired YAML policy schema with runtime_safety block no independent evidence
    purpose: To force LLM outputs into a bounded, machine-checkable action space for the deterministic controller.
    Defined in Table III and Appendix; validity is measured only inside this schema.

pith-pipeline@v1.1.0-grok45 · 24359 in / 3572 out tokens · 36642 ms · 2026-07-11T20:19:02.813254+00:00 · methodology

0 comments
read the original abstract

Large Language Models (LLMs) are proposed as control interfaces for next-generation networks, but their latency, hallucinations, and lack of control guarantees make them unsuitable for near-real-time packet schedulers, especially in dynamic V2X environments. This paper introduces Agentic-V2X, an architecture where a small, locally deployed language model acts as a periodic non-real-time rApp-inspired policy creator, while a lightweight xApp-like controller executes validated policies at intervals suitable for scheduling. The framework targets deadline-aware 5G NR V2X scheduling with heterogeneous services (teleoperated driving, cooperative awareness, HD map sharing, and sensor sharing). Given a scenario summary, service objective, and telemetry, the LLM generates a structured policy containing service priorities, weight bounds, and safety constraints. A validator checks and repairs the policy before the controller enforces it via scheduler-weight adaptation in ns-3/ns3-ai. The evaluation compares proportional fair scheduling, static expert policies, a heuristic xApp, static LLM policies, and adaptive LLM-rApp policies over 126 completed runs. Metrics include deadline-constrained packet reception ratio, tail latency, deadline violations, throughput, fairness, policy validity, and safety interventions. Results show that the adaptive LLM-rApp/xApp design generates valid and executable policies and remains competitive at several operating points, including improved mean critical reliability over PF at the highest density. However, paired statistical analysis shows that the adaptive method is not the best aggregate method and remains below the strongest static policies overall. These results support Agentic-V2X as a safe, executable small-LLM policy-generation architecture rather than a universally dominant scheduler.

Figures

Figures reproduced from arXiv: 2607.04290 by Alexandros Kaloxylos, Athanasios Kanavos, Gerasimos Papanikolaou-Ntais.

Figure 1
Figure 1. Figure 1: Proposed Agentic-V2X architecture. The system establishes strict timescale separation, isolating the stochastic small language model inside a slow [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Manhattan-style Urban grid seeds. The offered load per vehicle is fixed and compliant to the V2X 5GAA application definitions, so increasing ve￾hicle density increases aggregate offered load and channel contention because more vehicles generate traffic according to the same fixed service models. B. V2X Service Classes and Deadlines Four V2X services are evaluated, with two critical and two background class… view at source ↗
Figure 3
Figure 3. Figure 3: Control timescales in Agentic-V2X. The LLM policy agent operates [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Reliability trends by density. The left panel reports critical DC-PRR, while the right panel isolates ToD UL DC-PRR. Error bars show the standard [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overall critical-service DC-PRR components across all densities and seeds. Bars show means and error bars show the standard error of the mean [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Selected operating points where adaptive LLM-rApp/xApp is strongest or especially competitive: high-density critical reliability, high-density ToD [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Aggregate performance trade-offs across all densities and seeds. Higher is better for critical DC-PRR and background throughput, while lower is [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Aggregate paired critical DC-PRR differences for adaptive LLM [PITH_FULL_IMAGE:figures/full_fig_p011_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

19 extracted references · 3 linked inside Pith

  1. [1]

    Under- standing O-RAN: Architecture, interfaces, algorithms, security, and re- search challenges,

    M. Polese, L. Bonati, S. D’Oro, S. Basagni, and T. Melodia, “Under- standing O-RAN: Architecture, interfaces, algorithms, security, and re- search challenges,”IEEE Communications Surveys & Tutorials, vol. 25, no. 2, pp. 1376–1411, 2023

    2023

  2. [2]

    Comprehensive survey of radio resource allocation schemes for 5G V2X communications,

    T. T. T. Le and S. Moh, “Comprehensive survey of radio resource allocation schemes for 5G V2X communications,”IEEE Access, vol. 9, pp. 123 117–123 133, 2021

    2021

  3. [3]

    Challenges and solutions for cellular based V2X communications,

    S. Gyawali, S. Xu, Y . Qian, and R. Q. Hu, “Challenges and solutions for cellular based V2X communications,”IEEE Communications Surveys & Tutorials, vol. 23, no. 1, pp. 222–255, 2021

    2021

  4. [4]

    Large language model (LLM) for telecommunications: A comprehensive sur- vey on principles, key techniques, and opportunities,

    H. Zhou, C. Hu, Y . Yuan, Y . Cui, Y . Jin, C. Chen, H. Wu, D. Yuan, L. Jiang, D. Wu, X. Liu, C. Zhang, X. Wang, and J. Liu, “Large language model (LLM) for telecommunications: A comprehensive sur- vey on principles, key techniques, and opportunities,”arXiv preprint arXiv:2405.10825, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Intent-based network for RAN management with large language models,

    F. A. Bimo, M. A. C. Galdon, C.-K. Lai, R.-G. Cheng, and E. K. P. Chong, “Intent-based network for RAN management with large language models,”arXiv preprint arXiv:2507.14230, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Deep reinforcement learning for resource allocation in V2V communications,

    H. Ye and G. Y . Li, “Deep reinforcement learning for resource allocation in V2V communications,” inProc. IEEE International Conference on Communications (ICC), 2018, pp. 1–6

    2018

  7. [7]

    Deep reinforcement learning based resource allocation for V2V communications,

    H. Ye, G. Y . Li, and B.-H. F. Juang, “Deep reinforcement learning based resource allocation for V2V communications,”IEEE Transactions on Vehicular Technology, vol. 68, no. 4, pp. 3163–3173, 2019

    2019

  8. [8]

    Spectrum sharing in vehicular networks based on multi-agent reinforcement learning,

    L. Liang, H. Ye, and G. Y . Li, “Spectrum sharing in vehicular networks based on multi-agent reinforcement learning,”IEEE Journal on Selected Areas in Communications, vol. 37, no. 10, pp. 2282–2292, 2019

    2019

  9. [9]

    Understanding telecom language through large language models,

    L. Bariah, H. Zou, Q. Zhao, B. Mouhouche, F. Bader, and M. Debbah, “Understanding telecom language through large language models,” in Proc. IEEE Global Communications Conference (GLOBECOM), 2023, pp. 6542–6547

    2023

  10. [10]

    LLM-based policy generation for intent-based management of applications,

    K. Dzeparoska, J. Lin, A. Tizghadam, and A. Leon-Garcia, “LLM-based policy generation for intent-based management of applications,” inProc. 19th International Conference on Network and Service Management (CNSM), 2023, pp. 1–7

    2023

  11. [11]

    LLM-guided open RAN: Empowering hierarchical RAN intelligent control,

    L. Bao, S. Yun, J. Lee, and T. Q. S. Quek, “LLM-guided open RAN: Empowering hierarchical RAN intelligent control,”arXiv preprint arXiv:2504.18062, 2025

    arXiv 2025

  12. [12]

    When large language model agents meet 6G networks: Perception, grounding, and alignment,

    M. Xu, D. Niyato, J. Kang, Z. Xiong, S. Mao, Z. Han, D. I. Kim, and K. B. Letaief, “When large language model agents meet 6G networks: Perception, grounding, and alignment,”arXiv preprint arXiv:2401.07764, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    A survey on large language model based autonomous agents,

    L. Wang, C. Ma, X. Feng, Z. Zhang, H. Yang, J. Zhang, Z. Chen, J. Tang, X. Chen, Y . Lin, W. X. Zhao, Z. Wei, and J.-R. Wen, “A survey on large language model based autonomous agents,”Frontiers of Computer Science, vol. 18, no. 6, p. 186345, 2024

    2024

  14. [14]

    Mi- croscopic traffic simulation using SUMO,

    P. A. L ´opez, M. Behrisch, L. Bieker-Walz, J. Erdmann, Y .-P. Fl ¨otter¨od, R. Hilbrich, L. L ¨ucken, J. Rummel, P. Wagner, and E. Wießner, “Mi- croscopic traffic simulation using SUMO,” inProc. 21st International Conference on Intelligent Transportation Systems (ITSC), 2018, pp. 2575–2582

    2018

  15. [15]

    An E2E simulator for 5G NR networks,

    N. Patriciello, S. Lag ´en, B. Bojovi´c, and L. Giupponi, “An E2E simulator for 5G NR networks,”Simulation Modelling Practice and Theory, vol. 96, p. 101933, 2019

    2019

  16. [16]

    Calibration of the 5G- LENA system level simulator in 3GPP reference scenarios,

    K. Koutlia, B. Bojovic, Z. Ali, and S. Lagen, “Calibration of the 5G- LENA system level simulator in 3GPP reference scenarios,”Simulation Modelling Practice and Theory, vol. 119, p. 102580, 2022

    2022

  17. [17]

    ns3-ai: Fostering artificial intelligence algorithms for networking research,

    H. Yin, P. Liu, K. Liu, L. Cao, L. Zhang, Y . Gao, and X. Hei, “ns3-ai: Fostering artificial intelligence algorithms for networking research,” in Proc. 2020 Workshop on ns-3 (WNS3), 2020, pp. 57–64. APPENDIX This appendix reports the main artifacts used by the rApp-inspired LLM policy generator. It includes the prompt structure, the constrained Y AML poli...

    2020

  18. [18]

    >", ">=",

    System Prompt:Listing A.1: System prompt used by the rApp-inspired LLM policy generator. You are a 5G NR V2X rApp policy generator. Your task is to generate one structured scheduler policy for a deterministic xApp controller. You are not the online scheduler. You do not output immediate runtime actions. You only output a policy that the xApp can validate ...

  19. [19]

    >" value: 15 action: service: ToD type: increase_weight value: 1 - name: awareness_violation_guard condition: metric: awareness_violation_rate operator:

    User Prompt Template:The user prompt provides the scenario-specific telemetry and identifiers. The telemetry block is generated by the simulator-side logging pipeline and summarises the recent network state available to the slow policy-generation loop. Listing A.2: User prompt template used to instantiate a scenario-specific LLM-rApp policy request. Creat...