{"id":"3503fa96-f339-4154-96d4-582c86112a97","arxiv_id":"2611.16801","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An agentic control loop combining UPF telemetry, LLM-based candidate selection, deterministic validation, and PFCP actuation is implemented and evaluated for SLA-constrained network slicing.","lead":"This paper implements a control loop that uses a local LLM to select rate limits for 5G network slices during overload, then applies those limits via PFCP updates to a programmable UPF. The work demonstrates a working integration of telemetry, LLM inference, deterministic validation, and PFCP actuation, but does not measure delivered per-class QoS or the incremental benefit of the LLM selector.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"LLM benefit is untested and the evaluation cannot distinguish whether the LLM adds value over a deterministic rule-based selector with identical candidates, telemetry, and timing.","rationale":"The reader's weakest_assumption correctly identifies that the LLM benefit is untested. The paper's own limitations (b161, b164, b165) state this explicitly. My concern is that the paper's framing and title imply a novel contribution from the LLM, but the evaluation cannot support that. The strongest claim (b159) is about integration, which is supported, but the paper's novelty hinges on the LLM component. The reader's verdict is CONDITIONAL, and I agree that this is appropriate. The concrete test I propose would directly address this gap. I do not recommend changing the verdict because the paper honestly states the limitation, and the integration claim is supported. However, if the authors intend to claim LLM benefit, they need this comparison. The paper's own sentences (b161, b165) show they are aware of this, so the verdict remains CONDITIONAL with the same condition: the LLM benefit must be tested. I agree with the reader's assessment.","tokens_in":11147,"tokens_out":635,"duration_ms":6533,"concrete_test":"Implement a deterministic selector that, given the same telemetry deltas, candidate set, and phase label, selects the candidate that minimizes the absolute difference between delivered throughput and a target (e.g., 90% of offered load) while respecting monotonicity. Run the same 10 trials per phase. If the deterministic selector matches or exceeds E4b on any metric (throughput, oscillations, fallbacks), then the LLM provides no measurable benefit in this setting.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central architectural claim is that an LLM-based selector is part of a working agentic control loop. However, the evaluation explicitly does not compare against a deterministic selector that uses the same candidate set, telemetry, constraints, and timing (b161, b165). The only comparison is between E4 (unconstrained agentic) and E4b (proposed with monotone filtering and deterministic action derivation), but this conflates the effect of the LLM with the effect of the filtering/derivation. The paper acknowledges this gap (b161, b164, b165), but it means the core claim that the LLM provides any benefit—or even that the LLM is necessary—is unsupported. Without this comparison, the system could be replaced by a simple rule (e.g., 'reduce Bronze by 10% each window until plateau') and achieve the same or better results. The paper's own strongest claim (b159) is about integration, not about LLM advantage, but the title and framing imply agentic orchestration as a novel contribution. The load-bearing assumption is that the LLM selection is a meaningful component; this is untested and therefore the central claim is weaker than presented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper proposes an agentic control loop for SLA-constrained network slicing in programmable mobile data planes. It integrates UPF telemetry, LLM-based candidate selection, deterministic validation, and PFCP actuation on a real BESS-UPF testbed. The evaluation shows no oscillations or fallbacks and unchanged highest-priority rate limits, but acknowledges that delivered per-class QoS protection remains unmeasured and the incremental benefit of LLM selection is untested.","tokens_in":11447,"tokens_out":1478,"duration_ms":12123,"significance":"If the integration is robust and the approach can be extended to unannounced traffic transitions and per-class QoS measurements, it could contribute to practical network slicing enforcement. However, the current evaluation does not demonstrate that the LLM component provides any benefit over a deterministic selector, and the lack of per-class QoS measurements limits the significance of the results.","major_comments":[{"comment":"The paper explicitly states that the incremental benefit of LLM selection is untested and requires a deterministic selector with identical candidates, telemetry, constraints, and timing. This is a central claim of the paper (agentic orchestration with LLM), yet the evaluation does not isolate the LLM's contribution. Without this comparison, the paper's title and framing overstate the role of the LLM. The authors should either provide such a comparison or adjust the claims to reflect that the LLM is a replaceable component whose benefit is not demonstrated.","section":"§V, b161, b164, b165","authors_acknowledge":"The incremental benefit of LLM selection requires a deterministic selector with identical candidates, telemetry, constraints, and timing.","authors_acknowledge_p":0.95},{"comment":"The paper repeatedly states that delivered per-class QoS protection is unmeasured. However, the problem statement (Figure 1) and the motivation rely on protecting Gold traffic during overload. Without per-class throughput, loss, or latency measurements, the paper cannot support the claim that the proposed controller preserves priority differentiation. The aggregate throughput results (Table III) show that agentic controllers deliver less throughput than Static and Greedy, but without per-class data, this cannot be attributed to service protection. The authors should either provide per-class measurements or explicitly reframe the contribution as configuration preservation only, not QoS protection.","section":"§V, b7, b50, b56, b118","authors_acknowledge":"Without per-class measurements, the lower aggregate throughput cannot be attributed to a service-protection benefit.","authors_acknowledge_p":0.96},{"comment":"The evaluation uses a known traffic schedule with pre-labeled phases (A, B, C, D). The paper acknowledges that autonomous phase detection is not assessed. This is a significant limitation because the controller relies on the phase label to identify the affected class and to apply monotone filtering. In a real deployment, the controller would need to detect overload autonomously. The authors should discuss how the system could be extended to detect phases without external labels, or at least acknowledge that the current evaluation does not test this capability.","section":"§V, b122, b164","authors_acknowledge":"The evidence concerns configuration preservation and operational integration; delivered per-class QoS, autonomous phase detection, and the incremental benefit of LLM selection remain open.","authors_acknowledge_p":0.64},{"comment":"The paper reports that E4b (proposed) has zero oscillations and zero fallbacks, while E4 (unconstrained) has 80 oscillations and 50 fallbacks. However, the comparison between E4 and E4b conflates two changes: monotone candidate filtering and deterministic action derivation. The paper acknowledges this (b110) but does not disentangle their effects. It is unclear whether the improvement is due to the filtering alone, the action derivation alone, or both. The authors should design an ablation study to isolate the contribution of each component.","section":"§V, Table IV, b146","authors_acknowledge":"Their comparison therefore evaluates both changes together.","authors_acknowledge_p":0.81},{"comment":"The paper states that traffic continuity between windows and overload-to-enforcement delay are not reported. This is a critical metric for a real-time control loop. Without knowing the delay from overload detection to MBR update, it is impossible to assess whether the controller can react quickly enough to prevent QoS degradation. The authors should measure and report this delay, or at least discuss the implications of the observed latencies (median 2.72 s, 95th percentile 3.53 s) on the control loop's effectiveness.","section":"§V, b158","authors_acknowledge":"Traffic continuity between windows and overload-to-enforcement delay are not reported.","authors_acknowledge_p":0.91}],"minor_comments":[{"comment":"The description of candidate construction uses the term 'phase label supplied by the experimental schedule.' This should be clarified: the controller uses the known schedule to identify the affected class. In a real deployment, this would not be available. The authors should note this assumption explicitly.","section":"§III, Step 2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The table lists initial MBR values and settings for E2 and E3, but it does not show the MBR values for E4 and E4b. It would be helpful to include the initial configuration used by all controllers for clarity.","section":"§IV, Table II","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The figure shows aggregate throughput and policing ratio, but the axes are not clearly labeled. The y-axis for policing ratio is labeled 'Policing Ratio (app-QER-red ÷ received)' but the formula uses 'received' while the text defines ρ = r/(r+p). This inconsistency should be resolved.","section":"§V, Figure 4","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The Mann-Whitney U test p-value is reported as 1.83×10^{-4} for all six comparisons between E4 and conventional policies. The paper notes that identical rank statistics yield identical p-values despite different throughput gaps. This is correct but could be misinterpreted. The authors should clarify that the test only assesses whether the distributions differ, not the magnitude of the difference.","section":"§V, b120","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The figure shows a strong correlation (r=0.976) between token volume and LLM request latency. The authors should discuss whether this correlation is expected and whether it has implications for the control loop's timing predictability.","section":"§V, Figure 10","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The conclusion lists future work items, but it would be helpful to prioritize them. The most critical gap is the lack of per-class QoS measurements, which should be addressed before any claims about service protection can be made.","section":"§VI, b165","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper uses the term 'agentic' extensively but does not define it clearly. A brief definition or reference to the agentic AI literature would help readers unfamiliar with the term.","section":"General","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Colleague,\n\nThis paper does what it says on the tin: it wires together UPF telemetry, an LLM-based candidate selector, deterministic validation, and PFCP actuation on a real BESS-UPF data plane, and shows that the resulting loop runs without oscillations or fallbacks while preserving configured Gold MBR limits. That integration is real, and the authors are unusually honest about what they have and haven't shown. The paper is worth a read if you work on programmable data planes or 5G/6G QoS enforcement and want to see one concrete way to close a telemetry-to-actuation loop.\n\nWhat is actually new is the specific assembly: SLA-monotone candidate filtering, deterministic action derivation that decouples the model's output from the PFCP command, and the audit trail. The individual pieces are off-the-shelf (BESS-UPF, Qwen2.5-7B, PFCP), but the end-to-end integration on a two-site FABRIC testbed is not something I've seen before. The evaluation is clean: ten independent runs per controller, Mann-Whitney U tests, clear event counting.\n\nNow the soft spots, which the paper flags itself. The big one: the LLM's incremental benefit over a deterministic rule-based selector is completely untested (b161, b165). The only comparison that involves the LLM is E4 (unconstrained agentic) vs E4b (proposed with monotone filtering and deterministic derivation), but that conflates the effect of the LLM with the effect of the filtering/derivation. The paper could be replaced by a simple rule like 'reduce Bronze by 10% each window until plateau' and we wouldn't know if the result would be the same or better. The authors acknowledge this openly, but it means the title's 'agentic orchestration' is doing more rhetorical work than the evidence supports. Second, delivered per-class QoS is unmeasured (b7, b50, b56, b118). The paper only reports aggregate throughput and policing ratio, so we cannot tell whether the control loop actually protects Gold/Silver/Bronze service, which is the whole point of the exercise. Third, the evaluation uses a known traffic schedule with pre-labeled phases (b164), so autonomous overload detection is not assessed. These are not fatal flaws—the paper is honest about them—but they limit what can be concluded.\n\nWho is this for? Researchers building programmable UPF control loops who want a reference implementation and a clear statement of what integration looks like on real hardware. It is not for someone looking for evidence that LLMs improve QoS enforcement, because that evidence is absent.\n\nRecommendation: I would send this to peer review. The integration is real, the implementation is available, and the paper is honest about its limitations. A good reviewer will ask for the missing comparison and per-class measurements, but the paper as a snapshot of a working system is worth publishing as a demo or short paper. I would not cite it for claims about LLM benefit, but I would cite it for the integration architecture and the SLA-monotone candidate policy.","headline":"Honest integration paper that delivers exactly what it claims—a working control loop—but the LLM's role is decorative, not load-bearing, and the authors say so themselves.","tokens_in":12033,"tokens_out":740,"duration_ms":8200},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"Agentic loop adjusts 5G slice rates without oscillations","keywords":["5G network slicing","QoS enforcement","User Plane Function (UPF)","PFCP","agentic AI","LLM-based control","BESS-UPF","SLA-monotone candidate policy"],"falsifier":"Run the same experiment with a deterministic selector that always chooses the most restrictive admissible candidate during overload (e.g., the lowest MBR value for the affected class that satisfies the rate budget) and compare the number of control events, oscillations, fallbacks, and aggregate throughput. If the deterministic selector matches or outperforms the LLM-based selector, the claim that LLM selection provides incremental benefit is falsified.","tokens_in":14362,"feed_emoji":"🔁","tokens_out":1614,"duration_ms":23307,"temperature":0.7,"pith_summary":"This paper tries to establish that an agentic control loop using a local LLM to select among precomputed rate candidates can adjust per-session bandwidth limits in a 5G user-plane data plane during overload, while preserving the configured limits of the highest-priority slices and avoiding oscillations. The authors demonstrate a working integration of telemetry, LLM-based selection, deterministic validation, and PFCP actuation on a real BESS-UPF DPDK pipeline. The core claim is that the proposed architecture, which combines SLA-monotone candidate generation with deterministic action derivation, can reduce control events and eliminate oscillations and fallbacks compared to an unconstrained agentic variant. However, the paper explicitly acknowledges that delivered per-class QoS protection remains unmeasured, that autonomous phase detection is not assessed, and that the incremental benefit of LLM selection over a deterministic selector with identical candidates is untested. The significance, if true, is that it provides a path toward more adaptable and safer QoS enforcement in programmable mobile data planes, combining the flexibility of LLM-based reasoning with the reliability of deterministic constraints, but the evidence is limited to configuration preservation and operational integration under a known traffic schedule.","feed_headline":"Agentic loop adjusts 5G slice rates without oscillations","feed_subtitle":"LLM selects bandwidth candidates; deterministic rules preserve priority limits and eliminate fallbacks in real UPF tests.","key_machinery":"The SLA-monotone candidate policy is the central mechanism. It defines rules for constructing the set of admissible MBR vectors during each control window: (i) Gold slices are pinned to their initial MBR values, (ii) during overload, candidates may only maintain or reduce the affected class MBR within the current phase (monotone tightening), and (iii) during recovery, the candidate set collapses to the initial MBR vector. This policy, combined with deterministic validation that checks the LLM-selected candidate against the rate budget and monotonicity rules, ensures that the controller cannot increase the affected class MBR during overload or violate the priority ordering. The paper shows that this filtering alone eliminates the oscillations and fallbacks observed in the unconstrained variant E4, independent of the selector's quality.","core_discovery":"The paper demonstrates a working integration of a telemetry-aware agentic control loop for adapting configured per-session MBR limits on a real BESS-UPF data plane. The controller combines UPF telemetry with a locally hosted LLM (Qwen2.5-7B-Instruct) that selects among a finite set of SLA-monotone MBR candidates. The key innovation is the separation of candidate selection from deterministic validation and action derivation: the LLM chooses a candidate identifier, but the controller independently validates the choice against the rate budget and monotonicity rules, and derives the PFCP update. In the evaluated runs with known traffic phases, this approach preserved Gold slice limits across all 480 windows, eliminated oscillations and fallbacks that occurred in the unconstrained agentic variant (E4), and reduced the number of MBR modifications from 124 to 32. The paper claims this demonstrates configuration preservation and operational integration, but explicitly states that delivered per-class QoS protection, autonomous phase detection, and the incremental benefit of LLM selection over a deterministic selector remain unmeasured and open.","pith_inferences":["The paper's strongest evidence is for configuration preservation and operational integration, not for service-level QoS protection. A reader should infer that the claimed benefits are about stability and reliability of the control loop itself, not about improved throughput or latency for any traffic class. The aggregate throughput measurements do not show a clear advantage over simpler policies li","The lack of comparison with a deterministic selector using the same candidates, telemetry, and timing means the paper does not establish that the LLM adds value over a rule-based selector. The results could be replicated or even improved by a simple lookup table or heuristic that selects the most restrictive admissible candidate during overload. This is the most critical open question for the arch","The experiment uses a known traffic schedule with pre-labeled phases, so the controller never needs to autonomously detect overload or transition between phases. The paper's results say nothing about the system's ability to operate under unannounced traffic changes, which is the realistic operational condition. Future work would need to demonstrate autonomous phase detection or at least robustness","The median control interval of 44.4 seconds, which includes time outside the traffic window, suggests the loop operates at a timescale of tens of seconds, not milliseconds. This makes it suitable for slow-varying overload conditions but likely inadequate for burst-scale congestion control. The paper does not report overload-to-enforcement delay, which would be critical for assessing responsiveness"],"forward_implications":["If the agentic control loop works as demonstrated, it provides a template for combining LLM-based decision making with hard safety constraints in network control loops, potentially applicable beyond 5G slicing to any scenario requiring runtime adaptation with priority preservation.","The reduction of control events from 124 to 32 and the elimination of oscillations and fallbacks suggest that SLA-monotone candidate filtering can substantially improve the stability and predictability of LLM-driven network control, even without improvements in the selector itself.","The working integration of telemetry, LLM inference, deterministic validation, and PFCP actuation on a real DPDK pipeline demonstrates that such agentic loops are feasible with current hardware and software, not just simulations.","The explicit identification of gaps (unmeasured per-class QoS, no autonomous phase detection, untested LLM benefit over deterministic selector) sets a clear agenda for follow-up work to establish whether this architecture provides actual service-level benefits."],"supporting_citations":[{"why":"Supplies the 3GPP TS 23.501 definition of MBR enforcement at the UPF and QERs, which is the problem context the paper addresses.","marker":"[5]"},{"why":"Provides the PFCP protocol basis for session modification and the AGORA architecture that the paper extends for agentic control.","marker":"[9]"},{"why":"Describes the FABRIC testbed infrastructure used for the two-site experimental deployment.","marker":"[22]"},{"why":"Supplies the precedent of invalid action masking in policy gradient algorithms, which the paper cites as related to its action masking approach.","marker":"[20]"},{"why":"Supplies the precedent of safe reinforcement learning via shielding, cited as related to the deterministic validation layer.","marker":"[21]"}],"fun_headline_variants":[],"cache_read_input_tokens":0,"weakest_assumption_plain":"The paper assumes that having an LLM select among precomputed MBR candidates provides a benefit over a deterministic selector (e.g., a simple rule or lookup table) that uses the same candidates, telemetry, constraints, and timing, but this assumption is explicitly untested and identified as an open issue.","fun_headline_variants_meta":{"raw":{"variants":[]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000816,"raw_usage":{"cost":0.0008160152,"is_byok":false,"cost_details":{"upstream_inference_cost":0.0008160152,"upstream_inference_prompt_cost":7.25452e-05,"upstream_inference_completions_cost":0.00074347},"total_tokens":5087,"prompt_tokens":1002,"completion_tokens":4085,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":0}},"tokens_in":1002,"tokens_out":4085,"cost_source":"openrouter_usage","duration_ms":24137,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":0,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-13T12:34:28.292232+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Run the same experiment with a deterministic selector that always chooses the most restrictive admissible candidate during overload (e.g., the lowest MBR value for the affected class that satisfies the rate budget) and compare the number of control events, oscillations, fallbacks, and aggregate throughput. If the deterministic selector matches or outperforms the LLM-based selector, the claim that LLM selection provides incremental benefit is falsified.","supporting_citations":[{"cited_title":"System architecture for the 5G system (5GS),","cited_arxiv_id":null,"evidence_quote":"Supplies the 3GPP TS 23.501 definition of MBR enforcement at the UPF and QERs, which is the problem context the paper addresses."},{"cited_title":"AGORA: Agentic Green Orchestration Architecture for Beyond 5G Networks,","cited_arxiv_id":null,"evidence_quote":"Provides the PFCP protocol basis for session modification and the AGORA architecture that the paper extends for agentic control."},{"cited_title":"FABRIC: A National-Scale Programmable Experimental Network Infrastructure,","cited_arxiv_id":null,"evidence_quote":"Describes the FABRIC testbed infrastructure used for the two-site experimental deployment."},{"cited_title":"A closer look at invalid action masking in policy gradient algorithms,","cited_arxiv_id":null,"evidence_quote":"Supplies the precedent of invalid action masking in policy gradient algorithms, which the paper cites as related to its action masking approach."},{"cited_title":"Safe reinforcement learning via shielding,","cited_arxiv_id":null,"evidence_quote":"Supplies the precedent of safe reinforcement learning via shielding, cited as related to the deterministic validation layer."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}