{"id":"4cffc30e-66b3-43e0-8574-de77c907b72b","arxiv_id":"2507.14230","paper_version":2,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"An LLM-based agentic loop translates user intents into RAN transmit power configurations and converges on energy efficiency targets in a simulated network.","lead":"This paper builds a control loop that uses a large language model (LLM) to translate high-level requests such as 'make the RAN more energy efficient' into specific transmit power settings. It demonstrates in a simulator that the loop lowers transmit power and raises measured energy efficiency over a few iterations, though without controls, error bars, or released code.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported closed-loop convergence likely depends on the self-admitted unreliable aggregated PM data, since Section IV-B concedes inaccurate PM can mislead the loop and the experiments appear to use aggregated PM.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the closed loop assumes the returned PEE.EnergyEfficiency values are accurate, yet the authors themselves state in Section IV-B that aggregated PM data often do not reflect the true network state and can create a misinformation feedback loop. My reading of the experimental setup (O1 PM at fixed intervals, no streaming alternative used) suggests the reported runs likely used aggregated PM, so the concern is not hypothetical. The paper offers no evidence that the specific PM values used in Figs. 3 and 4 were accurate, and the unit inconsistency in Table III (kWh vs. bits/joule in Fig. 3) reinforces the uncertainty about what was measured. A direct comparison between aggregated PM and simulator ground truth, or a rerun with streaming PM, would settle whether the reported convergence is genuine. This does not move the verdict from the reader's CONDITIONAL because that verdict already reflects the high correctness risk and the need for additional evidence; my stress-test confirms and sharpens the condition rather than changing it.","tokens_in":6417,"tokens_out":3424,"duration_ms":41406,"concrete_test":"Re-run the closed-loop experiment from Section III while logging both the aggregated PM values delivered over O1 (as used in Figs. 3 and 4) and the simulator's internal ground-truth PEE.EnergyEfficiency at each iteration. Compute the relative error between the two series. If the error exceeds a few percent at any iteration before the claimed convergence, the reported TxPower trajectory was driven by inaccurate feedback and the convergence claim is invalidated. Additionally, rerun the seven-attempt precision evaluation with streaming, non-aggregated PM (as suggested in Section IV-B); if iteration counts or final TxPower values differ materially, the published results are artifacts of aggregation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the LLM-orchestrated closed loop improves RAN energy efficiency by adjusting TxPower based on real-time PM feedback (Figs. 3 and 4). That claim is load-bearing on the accuracy of the PEE.EnergyEfficiency values delivered to the agents. Section IV-B states that aggregated PM values 'often do not accurately reflect the network's true state' and that inaccurate PM creates a 'feedback loop of misinformation' that can destabilize the network. The reported experiments use the O-RAN O1 interface with PM transmitted at fixed intervals and aggregated at the gNB; the authors offer streaming PM as future work, implying the current results used the questionable aggregated path. If the PM values used during the reported runs were inaccurate, then the convergence in Figs. 3 and 4 may reflect convergence to a distorted metric rather than a genuine improvement in energy efficiency. The paper provides no ground-truth comparison, no error analysis, and no replication, so this threat is not ruled out. The additional unit inconsistency (kWh in Table III versus bits/joule in Fig. 3) further obscures what quantity was actually optimized. Because the authors themselves identify PM inaccuracy as a critical limitation, the central claim remains conditional on PM fidelity that is not demonstrated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an intent-based network (IBN) architecture for RAN management in which large language models (LLMs), embedded in a Strategist Agent and a History Analyzer Agent orchestrated with LangGraph, translate formalized intents (JSON per 3GPP TS 28.312) into TxPower configurations via the O-RAN O1 interface. The authors introduce a five-section prompt structure and report a single simulator experiment (VIA VI AI RSG) in which TxPower decreases from ~30 dBm to ~11 dBm while a PEE.EnergyEfficiency metric rises to ~8×10^5 bits/joule, plus a closed-loop precision test with seven attempts and a boundary test in Table III. The paper claims that the network can automatically improve energy efficiency through closed-loop LLM-orchestrated adaptation, but it also acknowledges in Section IV that aggregated PM data may be inaccurate and can cause a feedback loop of misinformation.","tokens_in":6616,"tokens_out":4234,"duration_ms":51177,"significance":"If the claimed results hold, the work would provide a concrete, standard-interface-based demonstration of LLM-driven intent translation and closed-loop RAN optimization, with strengths including a clear agentic decomposition, a structured prompt design, use of the O-RAN O1 interface for CM/PM, and an explicit discussion of contextual and data-accuracy limitations. The paper does not provide code, data, or a reproducible experimental pipeline, and the validation evidence is currently too limited to support the general claim of automatic energy-efficiency improvement. The authors honestly flag a critical vulnerability in their own data source, but they do not show that the reported runs were unaffected by that vulnerability.","major_comments":[{"comment":"The authors state in Section IV-B that aggregated PM values \"often do not accurately reflect the network's true state\" and can trigger a \"feedback loop of misinformation,\" while Section II-D describes retrieving PM via the O1 interface with fixed-interval transmission and aggregation at the gNB, and Section III gives no indication that streaming PM was used in the reported runs. Since the convergence shown in Figs. 3 and 4 is judged on exactly these PM values, the central demonstration is valid only if the PM data in those runs were accurate. Please report the actual aggregation settings and time windows used, validate the PM values against the simulator's ground-truth state, or repeat the experiments with streaming PM to rule out the self-identified failure mode.","section":"Section IV-B vs. Sections II-D and III"},{"comment":"Figure 3 reports a single trajectory from one un-replicated simulator run, with no error bars, no multiple seeds, and no baseline comparison (e.g., a rule-based policy, a simple hill-climbing controller, or a fixed-power scheme). One run cannot support the abstract's claim that \"the network can automatically improve its energy efficiency\" through the proposed method. The authors should provide multiple independent runs with variance, and compare the LLM-driven loop against at least one non-LLM baseline to demonstrate that the agentic architecture is responsible for the improvement.","section":"Section III-A, Fig. 3"},{"comment":"The energy-efficiency metric is inconsistent across the experimental sections: Fig. 3 and the intent example use bits/joule with a target around 800,000, and Section III-B refers to targets 800, 803, and 805 without units, while Table III labels PEE.EnergyEfficiency in \"KWh\" with values 765.4, 750, and 803. As written, it is unclear whether Table III and the Fig. 4 targets describe the same quantity on the same scale, and the \"KWh\" unit is not appropriate for an energy-efficiency metric. Please unify the units, rescale consistently, and restate the boundary-test results in the same metric used in the optimization.","section":"Table III and Figs. 3-4"},{"comment":"The paper's primary contribution is described as enhancing intent translation through a structured prompt, but no quantitative metric for translation accuracy is reported (e.g., parse success rate, semantic correctness of generated strategies, or alignment between the LLM output and a ground-truth configuration). The presented loop evaluation only checks whether the metric reaches the target value that was already provided in the prompt, so it conflates control performance with translation quality. Please add a direct evaluation of translation quality and an ablation that isolates the effect of the proposed prompt structure.","section":"Sections II-B and III"}],"minor_comments":[{"comment":"The manuscript contains several grammatical and typographical issues; for example, Section III-A says \"for iteration(s)\" and Section III-B says \"finally is 805.\" A careful language revision is needed.","section":"General"},{"comment":"Figure 1 appears to be a low-resolution screenshot with small text; it should be redrawn as a vector diagram with readable labels so that the agentic architecture and the O1 interface connections are clearly visible.","section":"Figure 1"},{"comment":"The description of one-shot prompting would benefit from showing the actual one-shot example in the prompt, rather than only describing the five-section structure, so that the structured technique can be reproduced.","section":"Section II-B"},{"comment":"The text says the first five attempts target 800, the sixth targets 803, and the seventh targets 805, but it does not state whether these values have the same units as the 800,000 bits/joule in the example intent; please clarify the scale and the precision of the comparison.","section":"Section III-B"},{"comment":"The baseline value in Table III is described as \"recorded initial simulation data,\" but the text does not state how many samples or over what time window; please specify the measurement procedure and the reported value's units.","section":"Section III-C"}],"recommendation":"major_revision","confidential_remarks":"The paper is a short demonstration of an interesting idea, but the experimental evidence is too thin for a standard journal paper: the central claim rests on a single un-replicated run and on PM data that the authors themselves admit may be unreliable. The unit inconsistency between Table III and Fig. 3 makes it hard to interpret the boundary test. I would not reject the manuscript outright, since the architecture and prompt design have merit and the limitations are explicitly acknowledged, but the authors need to substantially strengthen the validation, add baselines and repetitions, and clarify the PM data path before this can be considered for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a proof-of-concept for using a two-agent LLM loop over O-RAN O1 to translate intents into RAN parameter changes, with energy efficiency as the test case. What's genuinely new is the concrete integration: a Strategist agent and a History Analyzer agent, a five-section prompt, and NETCONF for configuration with PM feedback. That combination is not in the cited prior work. The authors also deserve credit for Section IV, where they openly admit the PM aggregation issue and the risk of a 'feedback loop of misinformation.'\n\nThe soft spot is the validation, which is very thin. The central demonstration is a single simulator run—no repeated trials, no error bars, no baseline comparison, no released code or exact prompts. The unit inconsistency is telling: Fig. 3 uses bits/joule, Table III uses kWh. That makes it hard to know what quantity was actually optimized. And the authors' own admission in Section IV-B undermines the load-bearing assumption: if the aggregated PM data don't reflect true network state, the reported convergence could be convergence to a distorted metric. They offer streaming PM as future work, which confirms the current results used the questionable aggregated path. There's also no quantitative measure of intent translation accuracy; the LLM is essentially given the target in the prompt and judged by whether it reaches that target, which is circular in a practical sense.\n\nThat said, the central claim isn't refuted. The architecture is plausible and the direction is timely. But as it stands, the evidence doesn't support the abstract's assertion that the network 'automatically improves its energy efficiency' in a robust sense. This is a workshop-grade prototype paper, not a rigorous journal result.\n\nWho should read it: anyone working on intent-based networking, LLM agents for O-RAN, or RAN automation. It's a useful reference for the architectural sketch and the honest limitations. I'd send it to peer review rather than desk reject—the topic is important and a serious referee could push the authors to fill the gaps—but the expectation should be major revision, with replication, baselines, and a resolution of the PM accuracy question. For my own work, I wouldn't cite it as evidence yet; it's a pointer to a direction, not a verified result.","headline":"A plausible proof-of-concept for LLM-driven RAN intent management, but the validation is a single un-replicated simulator run and the authors' own PM-accuracy admission undermines the central claim as stated.","tokens_in":7169,"tokens_out":2843,"would_cite":false,"duration_ms":30659,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An LLM-orchestrated agent loop translates a formalized network intent into transmit-power settings that move simulated RAN energy efficiency toward the requested target, converging in two to four iterations and correctly declaring…","keywords":["intent-based networking","large language models","radio access network","closed-loop control","energy efficiency","O-RAN","agentic AI","transmit power optimization"],"falsifier":"Re-run the seven-attempt precision test while injecting a small offset (for example, +3 percent) into the energy-efficiency values delivered to the agent, leaving the simulator's true state unchanged. If the loop still converges to the true target, the demonstration is robust to measurement error; if it oscillates, diverges, or marks attempts as failures prematurely, the convergence claim depends on trustworthy measurements.","tokens_in":6233,"feed_emoji":"📡","tokens_out":7861,"duration_ms":82961,"temperature":0.7,"pith_summary":"This paper tries to establish that a large language model can act as the translation and decision layer of an intent-based radio access network: a high-level goal such as raising energy efficiency above 800,000 bits per joule for a specific cell is converted into concrete configuration actions, and a closed loop keeps adjusting until the goal is met. The demonstration case is energy efficiency on a simulated two-cell RAN, where the LLM-orchestrated loop lowered transmit power from about $30$ dBm toward $11$ dBm and raised PEE.EnergyEfficiency from roughly $7.7\\times 10^5$ to about $8\\times 10^5$ bits per joule. Target values set near this baseline were reached in two to four iterations, while a target beyond the reachable power range was correctly declared a failure when the adjustment plateau was reached. If the claim is right, LLMs can close the intent-translation gap that has limited intent-based networking, turning formalized objectives into continuously re-tuned RAN configurations.","feed_headline":"AI agents tune RAN power to hit energy targets","feed_subtitle":"A closed-loop intent system reached energy-efficiency targets in two to four iterations on a simulated two-cell network.","key_machinery":"The load-bearing mechanism is the structured prompt that forces the LLM to treat intent translation as a constrained decision: five named sections (Instruction, Intent, Current Observation, Configuration Constraints, Output Format) turn a natural-language goal plus current network measurements into one parseable configuration action per iteration. Around that prompt, a two-agent orchestration flow uses past strategy attempts as context and closes the loop by reading performance-management data after each NETCONF configuration edit, so the next prompt sees the effect of the previous action. The important property is that the prompt converts an open-ended language model into an incremental controller that lowers or raises transmit power by small steps, which is what makes the loop converge instead of drift.","core_discovery":"The central claim is that the network can automatically improve its energy efficiency by dynamically optimizing critical RAN parameters through a closed-loop mechanism driven by LLM-orchestrated agents. In the reported experiment, a Strategist Agent parses a formalized JSON intent modeled on 3GPP TS 28.312, reads the current configuration and measurement data over an O-RAN O1 interface, and issues an incremental transmit-power edit through NETCONF; the resulting performance measurement is then fed back into the next decision. Across attempts with target values of $800$, $803$, and $805$ thousand bits per joule, the loop converged in $2$--$4$ iterations, and a seventh attempt stopped with a \"Fail\" outcome once the allowed transmit-power range was exhausted. The discovery is that an LLM, constrained by a structured prompt and supplied with history, can behave as a stable incremental optimizer for RAN configuration rather than a one-shot generator of settings.","pith_inferences":["The paper leaves implicit that caching the history agent's successful strategy trajectories could warm-start later attempts at the same intent, likely reducing iteration counts below the reported two to four.","A testable extension suggested by the authors' measurement-accuracy concern is to feed the LLM raw, unaggregated measurement streams; the expected effect is fewer misinformation loops but larger prompt sizes.","Because the experiment uses static user equipment and two cells, an unstated next question is whether the same five-section prompt remains stable when mobility and inter-cell interference corrupt the measurement-to-configuration mapping."],"forward_implications":["If the central claim holds, catalog-based intent translation is not required: an LLM can interpret a formalized intent and emit actionable RAN configuration changes directly.","A convergence time of two to four iterations means an intent can be fulfilled and verified within a few performance-measurement cycles, making closed-loop assurance practical.","When the transmit-power range is exhausted, the system stops and reports the intent as unfulfillable, so an autonomous controller can avoid endless or destabilizing retries.","For targets already satisfied by the current state, the system performs monitoring and verification only, functioning as an intent-assurance check as well as an optimizer."],"supporting_citations":[{"why":"defines the five-component closed-loop IBN lifecycle that the proposed agentic loop instantiates","marker":"[2]"},{"why":"supplies the 3GPP TS 28.312 format that the formalized JSON intent is parsed from","marker":"[14]"},{"why":"provides the O-RAN experimental platform whose cell configuration and measurement functions the system uses","marker":"[12]"},{"why":"supplies the agentic orchestration flow that sequences the LLM agents' actions","marker":"[13]"},{"why":"represents the prior generative-AI intent management approach that lacks the closed-loop mechanism this paper adds","marker":"[10]"}],"fun_headline_variants":["LLM agents tune RAN power to hit energy targets in steps","Closed-loop LLM optimizes RAN energy in 2-4 iterations","Agentic LLM adjusts RAN power for energy efficiency","LLM-driven RAN management converges on energy targets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The loop assumes the measurement values returned by the simulator are accurate enough to guide control; the authors report that aggregated measurement values often do not reflect the network's true state and that inaccurate values can create a feedback loop of misinformation.","fun_headline_variants_meta":{"raw":{"variants":["LLM agents tune RAN power to hit energy targets in steps","Closed-loop LLM optimizes RAN energy in 2-4 iterations","Agentic LLM adjusts RAN power for energy efficiency","LLM-driven RAN management converges on energy targets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000218,"raw_usage":{"total_tokens":1391,"prompt_tokens":851,"completion_tokens":540,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":467,"completion_tokens_details":{"reasoning_tokens":468}},"tokens_in":467,"tokens_out":540,"duration_ms":5604,"temperature":1.0,"reasoning_tokens":468,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:37:54.862305+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the seven-attempt precision test while injecting a small offset (for example, +3 percent) into the energy-efficiency values delivered to the agent, leaving the simulator's true state unchanged. If the loop still converges to the true target, the demonstration is robust to measurement error; if it oscillates, diverges, or marks attempts as failures prematurely, the convergence claim depends on trustworthy measurements.","supporting_citations":[{"cited_title":"A survey on intent-based networking,","cited_arxiv_id":null,"evidence_quote":"defines the five-component closed-loop IBN lifecycle that the proposed agentic loop instantiates"},{"cited_title":"Lte; 5g; management and orchestration; intent driven man- agement services for mobile networks (3gpp ts 28.312 version 18.6.0 release 18),","cited_arxiv_id":null,"evidence_quote":"supplies the 3GPP TS 28.312 format that the formalized JSON intent is parsed from"},{"cited_title":"Design and implementation of next-generation research platforms,","cited_arxiv_id":null,"evidence_quote":"provides the O-RAN experimental platform whose cell configuration and measurement functions the system uses"},{"cited_title":"langgraph","cited_arxiv_id":null,"evidence_quote":"supplies the agentic orchestration flow that sequences the LLM agents' actions"},{"cited_title":"Towards intent-based network management for the 6g system adopting multimodal generative ai,","cited_arxiv_id":null,"evidence_quote":"represents the prior generative-AI intent management approach that lacks the closed-loop mechanism this paper adds"}],"review_version":1}