{"id":"c8b92e1f-2944-4ad0-80d1-64ddee1e73da","arxiv_id":"2608.06789","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Reinforcement-learning fine-tuning of a compact LLM actor inside an O-RAN controller outperforms heuristics, a convex-optimization baseline, and much larger cloud LLMs in simulated IAB power allocation.","lead":"This paper proposes a system called EvoRIC that fine-tunes a small language model with reinforcement learning to control radio access networks, and tests it on wireless backhaul power allocation. A compact fine-tuned model is reported to beat much larger cloud models and a model-based optimizer, suggesting that small specialized models could run near-real-time network control.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed generalization across topologies in Figure 3 is not verifiable because the paper never specifies which topologies were used for PPO training; the results may be in-distribution rather than evidence of generalization.","rationale":"The reader's weakest-assumption analysis identifies the ambiguous training-topology protocol as the central issue, and my reading of the paper converges on the same point. The paper's own framing makes generalization the key novelty: abstract says traditional ML suffers from limited generalization, and Section I lists 'generalization across diverse network scenarios' as a key contribution. Yet the experimental section provides no explicit statement of which topologies were used for PPO training rollouts, nor any holdout designation. Figure 2's training curve uses M=3, N=5 (Scenario 2), making it plausible but not certain that Scenario 2 was the training topology; Figure 3 then includes Scenario 2 alongside two other topologies. This leaves open two very different interpretations: (i) EvoRIC-L2 was trained on Scenario 2 and evaluated on two unseen topologies, which would support a generalization claim; or (ii) training included all three scenarios, in which case Figure 3 only shows in-distribution fitting and the headline novelty evaporates. The paper must disambiguate this before the central claim can be assessed. I also note the absence of error bars on the Figure 3 averages; with 500 test examples, confidence intervals are cheap to compute and would materially affect the strength of the claim, especially in Scenario 3 where the reported headroom is small. These concerns do not require rejecting the framework—the architecture and KL-stabilized PPO design are reasonable—but they do require making the evaluation protocol explicit and providing statistical support before accepting the empirical claims as established. Since the reader already issued a CONDITIONAL verdict for essentially the same reason, my stress-test does not change the recommended verdict.","tokens_in":10298,"tokens_out":2779,"duration_ms":28922,"concrete_test":"Require the authors to state explicitly which topologies were used for PPO training and which were held out during evaluation. A decisive check: retrain EvoRIC-L2 under three protocols—(A) training only on Scenario 2, (B) training on Scenarios 1, 2, and 3 jointly, and (C) separate per-scenario models—and report the Figure 3 throughput for each protocol. If protocol B matches or exceeds protocol A on all scenarios, the claimed generalization is not supported; if protocol A preserves the Scenario 1 and 3 margins, the generalization claim holds. Additionally, report 95% confidence intervals for all Figure 3 points to assess statistical significance, particularly in Scenario 3.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central contribution stated in the abstract and Section I is that EvoRIC 'facilitates generalization across diverse network scenarios (e.g., varying numbers of BSs or users) without the need for architecture reconstruction.' The only direct evidence for this is Figure 3, which reports EvoRIC-L2 throughput on Scenario 1 (M=3, N=6), Scenario 2 (M=3, N=5), and Scenario 3 (M=2, N=3). However, Section III-C never states which topology (or topologies) the PPO rollouts were collected on. The training curve in Figure 2 is explicitly for M=3, N=5, which is Scenario 2; if EvoRIC-L2 was trained on Scenario 2, then that scenario is in-distribution and only Scenarios 1 and 3 are out-of-distribution candidates. The paper also does not rule out the possibility that the model was fine-tuned on all three scenarios, in which case Figure 3 demonstrates in-distribution performance only and the stated generalization contribution is unsupported. The absence of error bars or confidence intervals on the Figure 3 averages further weakens the claim, especially in Scenario 3 where the reported headroom over EPA is only 16.6% and the margin over baselines may not be statistically significant. Because the generalization claim is load-bearing for the paper's novelty and is not verifiable from the manuscript, the empirical results must be treated as conditional until the training/evaluation protocol is specified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces EvoRIC, a hierarchical O-RAN control framework in which a compact LLM (Llama-3.2-3B) is fine-tuned via proximal policy optimization (PPO) in the non-RT RIC and then deployed in the near-RT RIC for power allocation in IAB networks. The framework serializes network state into structured prompts, uses a composite reward with format penalty and KL regularization, and closes the loop through O1/A1/E2 interfaces. Experiments compare EvoRIC-L2 against EPA, Llama3B-NoFT, DeepSeek-V3, Gemini, and an SCA baseline, reporting throughput gains across three topologies, intent adaptation, and robustness to telemetry corruption.","tokens_in":10609,"tokens_out":5931,"duration_ms":56907,"significance":"If the results hold, the paper demonstrates a practically attractive combination: a 3B-parameter LLM fine-tuned with reinforcement learning can outperform both a conventional heuristic and a model-based solver for IAB power allocation while keeping inference latency below one second, and it can transfer to varying topology sizes without retraining. The closed-loop collect-update-deploy design and the explicit handling of format compliance and safety fallback are well motivated. The paper also includes a limitation statement for the telemetry robustness experiments, which is commendable. However, the empirical claims rest on an under-specified evaluation protocol, and the central generalization claim is not yet demonstrated.","major_comments":[{"comment":"The paper's central generalization claim is not verifiable because the training topology distribution is never specified. The training curve in Figure 2 is reported for M=3, N=5, and the evaluation scenarios in Figure 3 include M=3,N=6; M=3,N=5; and M=2,N=3. If the PPO rollouts used any of these scenarios, the corresponding result is in-distribution. Please state exactly which topologies were used for training, whether any evaluation scenarios were held out, and how the reported averages were computed.","section":"Section III-C, Figure 3"},{"comment":"All numerical results are averages over 500 episodes with no confidence intervals, standard deviations, or number of random seeds. The claimed margin in Scenario 3 (16.6% over EPA) may be within sampling noise; without variance bars it is impossible to judge statistical significance. Please report per-scenario confidence intervals or box plots and repeat with at least three seeds.","section":"Section III-C, Figure 3 and Table 1"},{"comment":"The safety verification module falls back to EPA whenever the LLM output is invalid, so the throughput averages in Figure 3 and Table 1 mix successful LLM actions with EPA fallback actions. The paper does not report the fraction of fallback episodes or the throughput conditional on a valid generation. This confound should be disentangled to assess the true contribution of the learned policy.","section":"Section III-B-4 and Section III-C"},{"comment":"The SCA-based baseline is under-specified; no details are given on the convex approximation scheme, initialization, iteration count, or convergence tolerance. Given that EvoRIC reportedly outperforms SCA by 10.6% on the weighted-throughput intent, please document the SCA implementation and verify that both methods use the same information and solve the same optimization problem.","section":"Section III-C, Table 1"},{"comment":"The simulator setup is not described sufficiently for reproduction. The paper omits the simulation platform, path-loss exponents, LoS probability parameters, Nakagami-m parameter, number of users per SBS, mobility parameters, episode length, and initial-state distribution. Please provide these details or release the simulator code.","section":"Section III-C and Section IV"}],"minor_comments":[{"comment":"The caption contains an apparent text-encoding artifact (e.g., '/s48/s49/s48/s48/s48 /s50/...'); please repair the caption and ensure the figure is legible.","section":"Figure 2 caption"},{"comment":"The phrase 'an fail-safe protocol' should read 'a fail-safe protocol'.","section":"Section III-B-4"},{"comment":"The notation 'top p' should be made consistent with 'top-p sampling' at first use; also define the PPO hyperparameters (clip ratio, GAE lambda, number of epochs) for reproducibility.","section":"Section III-C"},{"comment":"The sentence 'the solver improves upon EPA by only 16.6%' is ambiguous; specify which solver (SCA or EvoRIC) is referred to.","section":"Section III-C, Scenario 3 discussion"},{"comment":"The inference latency of 0.56 s is reported without specifying the inference hardware, quantization, batch size, or whether prompt parsing is included; please state the test conditions or add a caveat.","section":"Section III-C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central claim rests on the unstated training/evaluation protocol; without that, the generalization and comparison results are conditional. The garbled text in Figure 2 also suggests the camera-ready version was not proofread. If the authors supply the missing protocol and variance information, the paper could be suitable for publication; in the current form, the empirical evidence is insufficient."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on EvoRIC.\n\nThe framework itself is clear and sensible: a non-RT RIC runs PPO fine-tuning on a 3B LLM, the near-RT RIC does inference with a safety-verification fallback. The reward design with format penalty and KL regularization is standard but applied cleanly. I like the intent-adaptation and telemetry-robustness experiments; those go beyond what most papers of this type bother with.\n\nWhat's actually new is the specific application: a PPO-fine-tuned Llama-3.2-3B as the RL actor for IAB power allocation inside an O-RAN-style split. The baselines are reasonable, and the sub-second inference claim is relevant for the near-RT RIC.\n\nThe soft spots are real, and one is load-bearing. The paper's central contribution is cross-topology generalization, but the training topology set is never specified. Figure 2 shows training on M=3, N=5, which is Scenario 2, yet the text never says whether the final model was trained only on that scenario or on all three. If it was trained on all three, Figure 3 demonstrates in-distribution performance and the generalization claim collapses. That is not a minor omission; it is the key claim. The missing error bars and seeds on Figure 3 make it worse, especially in Scenario 3 where the margin over EPA is only 16.6%. The comparison to DeepSeek and Gemini is a bit apples-to-oranges since those are zero-shot via API, but the authors acknowledge the trade-off, so I don't weight that heavily. The latency numbers also come from different hardware, a practical point rather than a scientific one.\n\nIf the authors specify the protocol and report variance across seeds, this could become a solid paper. As it stands, the quantitative results are conditional. The framework idea is not wrong, and the architecture is described well enough that a referee can push for the missing details.\n\nMy call: send it to peer review rather than desk reject. It deserves referee time, but it needs major revision before acceptance. I would not cite the specific numerical gains until the training protocol is clarified.","headline":"EvoRIC is a well-structured framework paper whose headline generalization claim is unverified because the paper never specifies the PPO training topologies; the quantitative results are conditional.","tokens_in":11170,"tokens_out":2757,"would_cite":false,"duration_ms":26418,"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":"A fine-tuned 3B LLM can outperform optimization solvers and far larger models for RAN power allocation.","keywords":["open radio access network","RAN intelligent controller","large language model","reinforcement learning fine-tuning","proximal policy optimization","integrated access and backhaul","network topology generalization","near-real-time inference"],"falsifier":"Train EvoRIC-L2 on a held-out topology set that excludes the three reported scenarios, then evaluate on those scenarios: if the throughput advantage over EPA disappears or drops to in-distribution levels, the generalization claim is unsupported.","tokens_in":10108,"feed_emoji":"📡","tokens_out":6833,"duration_ms":54702,"temperature":0.7,"pith_summary":"This paper proposes EvoRIC, a hierarchical O-RAN control architecture in which a compact language model acts as the reinforcement-learning policy and is fine-tuned in the non-real-time RIC, then deployed to near-real-time RICs for fast inference. The claim is that this collect-update-deploy loop gives the LLM wireless-domain decision-making ability without expert-labeled data. Evaluated on integrated access and backhaul power allocation, the fine-tuned 3B model is reported to outperform equal-power allocation, the model-based SCA algorithm, and much larger general-purpose LLMs, while completing inference in 0.56 seconds. If correct, this would make LLM-driven, edge-deployable RAN control practical and reduce the need for topology-specific retraining.","feed_headline":"Small fine-tuned LLM beats giant models at RAN power control","feed_subtitle":"PPO fine-tuning lets a 3B Llama edge model outdo DeepSeek and Gemini for IAB allocation in 0.56 seconds.","key_machinery":"The machinery is an LLM as the actor inside a PPO reinforcement-learning agent, regularized by a frozen reference model through a KL-divergence penalty. The non-RT RIC aggregates interaction tuples (state prompt, generated action, reward) via the O1 interface, runs PPO updates on the last L transformer blocks, and redeploys via A1. The near-RT RIC serializes network state into a prompt with explicit output-format directives, parses the LLM's text into a power vector, verifies it against the power budget, and falls back to equal power allocation if validation fails. A composite reward adds a format penalty to the throughput objective, which keeps the model from producing unexecutable outputs.","core_discovery":"On its own terms, the paper's central discovery is that a 3B-parameter Llama-3.2-Instruct model, fine-tuned with PPO using only environmental reward, becomes a better power-allocation policy for IAB networks than the equal-power heuristic, the successive convex approximation solver, the un-fine-tuned Llama, and the much larger DeepSeek-V3 and Gemini 1.5 Flash models. EvoRIC-L2, the version with two final transformer blocks trained, achieves these results across three network topologies, adapts to a changed objective through prompt modification alone, and loses almost nothing when one telemetry stream is corrupted. The paper also reports that the KL-regularized PPO update is necessary for stable training, since without it the policy drifts and training becomes numerically unstable.","pith_inferences":["The paper never states whether the three evaluation topologies were part of the PPO rollout buffer; if they were, the reported cross-topology generalization is in-distribution and a held-out topology evaluation is needed.","The 0.56 s inference figure is measured on an RTX 4090 workstation; real edge hardware will likely be slower, so the sub-second claim should be re-verified on target COTS edge servers.","The comparison against DeepSeek and Gemini is zero-shot with prompt engineering; fine-tuning those large models might narrow the gap, though their size would still make near-RT deployment impractical.","The composite reward's format penalty and fail-safe fallback mean part of the measured gain comes from suppressing parse failures; ablating the fallback would quantify how much."],"forward_implications":["A single fine-tuned compact LLM could replace topology-specific RL models, cutting the retraining burden in heterogeneous RAN deployments.","Sub-second inference (0.56 s) falls inside the near-RT RIC control window, so LLM-based control is not limited to non-real-time planning.","KL regularization keeps the fine-tuned model close to its pretrained knowledge, which should reduce catastrophic forgetting during continued learning.","Prompt-level intent changes let the same model switch objectives, such as priority-weighted throughput, without parameter updates.","The safety-verification and fallback layer ensures that even hallucinated outputs degrade to a valid baseline, a necessary condition for deployment in critical infrastructure."],"supporting_citations":[{"why":"Defines the O-RAN RIC architecture, interfaces, and timescales that EvoRIC is built on.","marker":"[1]"},{"why":"Prior reinforced in-context learning for power control that the paper positions against.","marker":"[4]"},{"why":"Supplies the mechanism for RL fine-tuning of LLMs as PPO actors.","marker":"[10]"},{"why":"Motivates the KL-divergence penalty used to regularize PPO updates.","marker":"[11]"},{"why":"Provides the Llama-3.2-3B backbone model that EvoRIC fine-tunes.","marker":"[12]"},{"why":"One of the large cloud-based general-purpose LLM baselines compared in the experiments.","marker":"[13]"},{"why":"The other large cloud-based general-purpose LLM baseline compared via API.","marker":"[14]"},{"why":"Provides the convex-optimization solver used for the SCA baseline.","marker":"[15]"}],"fun_headline_variants":["3B LLM fine-tuned with PPO wins RAN power control","Small LLM outplays giants in O-RAN power allocation","PPO-tuned Llama 3B tops DeepSeek and Gemini for IAB","EvoRIC: RL fine-tuning makes 3B LLM a RAN control champ","RL-tuned 3B Llama beats huge models in RAN control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The cross-topology generalization claim assumes the test scenarios were not seen during fine-tuning, but the paper does not specify which topologies were in the training rollout buffer.","fun_headline_variants_meta":{"raw":{"variants":["3B LLM fine-tuned with PPO wins RAN power control","Small LLM outplays giants in O-RAN power allocation","PPO-tuned Llama 3B tops DeepSeek and Gemini for IAB","EvoRIC: RL fine-tuning makes 3B LLM a RAN control champ","RL-tuned 3B Llama beats huge models in RAN control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000681,"raw_usage":{"total_tokens":3097,"prompt_tokens":953,"completion_tokens":2144,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":2040}},"tokens_in":569,"tokens_out":2144,"duration_ms":14533,"temperature":1.0,"reasoning_tokens":2040,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:49:22.998885+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train EvoRIC-L2 on a held-out topology set that excludes the three reported scenarios, then evaluate on those scenarios: if the throughput advantage over EPA disappears or drops to in-distribution levels, the generalization claim is unsupported.","supporting_citations":[{"cited_title":"Un- derstanding O-RAN: Architecture, interfaces, algorithms , security, and research challenges,","cited_arxiv_id":null,"evidence_quote":"Defines the O-RAN RIC architecture, interfaces, and timescales that EvoRIC is built on."},{"cited_title":"Prompting Wireless Networks: Reinforced In-Context Learning for Power Control","cited_arxiv_id":"2506.06526","evidence_quote":"Prior reinforced in-context learning for power control that the paper positions against."},{"cited_title":"CVXPY: A Python-embedded model ing lan- guage for convex optimization,","cited_arxiv_id":null,"evidence_quote":"Provides the convex-optimization solver used for the SCA baseline."}],"review_version":1}