{"id":"61f05e2d-50a1-4d89-9161-8b0b32b6b734","arxiv_id":"2412.18200","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A frozen Llama2-7B with a CNN encoder and LoRA adapters is claimed to select congestion control algorithms that improve WiFi fairness and prevent starvation, but the experiment does not isolate the LLM's contribution.","lead":"TCP-LLM is a framework that lets Llama2-7B choose WiFi congestion control algorithms from numerical TCP metrics, using a CNN encoder and low-rank fine-tuning. It matters because it asks whether one pre-trained language model can replace hand-tuned network controllers, potentially cutting engineering effort across TCP tasks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The empirical section never tests whether Llama2's pre-trained weights cause the reported gains, so the central claim that TCP-LLM leverages pre-trained knowledge is unsupported.","rationale":"The paper's own framing in the abstract and Section VII.A makes the causal role of pre-trained knowledge the central claim, not just the availability of an ML-based CCA selector. Since the transformer weights are frozen (Section IV.C), the only parameters that can adapt to the TCP task are the encoder, projection, LoRA matrices, and head. A random frozen transformer, or even no transformer, could in principle supply the same nonlinear feature map. Without an ablation, the reported near-100% accuracy and throughput improvements are equally consistent with the hypothesis that a small supervised or offline-RL classifier is doing all the work. That would reduce the contribution to a known result, exactly the alternative the reader identifies. The contradictory Scenario 1 numbers (42 Mbps in Section VI.C vs 45/40 Mbps in Section VII.C for the same Figure 10a) compound the uncertainty, because they prevent even a faithful reader from pinning down which numbers are being compared. I do not see a way to repair this without new experiments or released artifacts; a revised paper with the random-init and no-LLM ablations and consistent numbers could become a legitimate empirical study. The framework itself is described in enough detail to be reproduced, and the LoRA memory figures are not internally inconsistent, but the decisive comparison is missing. For those reasons the reader's REJECT verdict is appropriate, and my stress test does not change it.","tokens_in":17052,"tokens_out":3680,"duration_ms":33549,"concrete_test":"Rerun the identical TCP-LLM pipeline (same encoder, LoRA rank, linear head, data splits, hyperparameters, and reward definition) three ways: (1) frozen Llama2-7B as in the paper; (2) the same model with Llama2 weights replaced by a randomly initialized transformer of the same shape, also frozen; (3) no transformer at all, i.e., CNN features go directly to the linear head. Compare CCA-selection accuracy and the Scenario 1-3 throughput/loss/RTT distributions. If (2) and (3) match (1) within run-to-run variability, the pretrained LLM contributes nothing and the central claim is refuted; if (1) clearly dominates, the claim survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a frozen Llama2-7B, combined with a CNN encoder, linear projection, LoRA matrices, and a linear head, exploits pre-trained knowledge to outperform DRL and fixed CCAs on CCA selection. The only comparisons reported are against DRL and fixed CCAs (Figures 10-15); there is no ablation in which the frozen transformer weights are randomly initialized, and no non-LLM model with the same encoder and head. Because all trainable components (CNN, projection, LoRA A/B, head) sit outside the frozen transformer, the observed accuracy curves in Figures 8 and 9 and the throughput/loss/RTT results in Figures 10-15 could be produced entirely by these small learned modules; the frozen LLM may be inert. Section VII.A's assertion that TCP-LLM 'leverages pre-trained capabilities' is therefore an attribution that the experimental design cannot support. An additional internal inconsistency weakens the same evidence: Section VI.C Scenario 1 reports TCP-LLM throughput as 42 Mbps for the BBR to Cubic transition, while Section VII.C states Figure 10a balances Cubic at 45 Mbps and BBR at 40 Mbps; the paper does not reconcile these numbers.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TCP-LLM, a framework that adapts a frozen Llama2-7B large language model to TCP control tasks by prepending a trainable 1D-CNN encoder, linear projection, LoRA adapters, and a linear classification head. The authors evaluate it on three tasks: flow fairness, starvation prevention, and CCA selection, using offline learning from datasets collected on a WiFi testbed. They report that TCP-LLM achieves faster convergence, lower loss, higher accuracy, and better throughput/loss/RTT trade-offs than a DRL baseline and than fixed CCAs (Cubic, BBR, PCC) in three transition scenarios.","tokens_in":17248,"tokens_out":3443,"duration_ms":33688,"significance":"If the empirical claims were fully supported, TCP-LLM would be a useful demonstration that a frozen decoder-only LLM can be adapted to CCA selection with a small trainable interface, potentially reducing engineering effort compared with bespoke DRL designs. The paper has concrete strengths: the architecture is described in sufficient detail to be reproduced, the parameter-efficiency claim (99% reduction in trainable parameters, 28.23 GB GPU memory) is quantified, and the single-step inference latency of 0.015 s is a clear and relevant number. The comparison against DRL and classical CCAs is a reasonable first benchmark set. However, the manuscript does not currently provide the ablation, statistical detail, or data/code release needed to verify the central attribution of the gains to the LLM's pre-trained knowledge.","major_comments":[{"comment":"The central claim that TCP-LLM 'leverages pre-trained capabilities' is not supported by the experimental design. All trainable components (the CNN encoder, linear projection, LoRA matrices A and B, and the linear TCP-LLM head) operate outside the frozen Llama2-7B weights. The reported accuracy and performance gains in Figures 8-15 could therefore be produced entirely by these small trainable modules, with the frozen LLM contributing little or nothing. The paper does not compare against a randomly initialized transformer with the same encoder/head, nor against a non-LLM model with the same encoder/head. Without such an ablation, the claim in Section VII.A that TCP-LLM 'leverages pre-trained capabilities to dynamically adapt to unseen conditions' is an attribution that the experiments cannot establish. This is a load-bearing issue because the paper's stated novelty is the use of pre-trained knowledge, not merely the addition of a small supervised network.","section":"Sections IV, V, VII.A; Figures 8-9"},{"comment":"There is an internal inconsistency in the reported TCP-LLM throughput for the Cubic/BBR transition. Section VI.C, Scenario 1 states that TCP-LLM achieves 'even higher throughput (42 Mbps; Figure 10a, right)' for the BBR-to-Cubic transition, whereas Section VII.C states that 'TCP-LLM balanced Cubic and BBR flows at 45 Mbps and 40 Mbps (Figure 10a)'. The paper does not reconcile these two numbers. In addition, the surrounding text says Cubic achieves a median throughput of 80 Mbps, which makes the statement that 42 Mbps is 'higher' than the traditional setup confusing unless the comparison is only to BBR's 7 Mbps. These discrepancies undermine the precision of the empirical evidence and need to be corrected and clarified.","section":"Section VI.C (Scenario 1) and Section VII.C"},{"comment":"The empirical evaluation lacks the statistical information needed to support the claim of 'significant improvements'. Figures 8 and 9 show training/test loss and accuracy curves without error bars or multiple runs, and Figures 10-15 present box plots and CDFs without stating the number of independent repetitions, seeds, or confidence intervals. The paper also does not report whether the DRL baseline was retrained under exactly the same data collection and testbed conditions as TCP-LLM, or whether the DRL results are taken from the authors' prior work. Without this information, the comparisons cannot be quantitatively assessed. The authors should provide per-run results with error bars or CIs, specify the number of runs, and clearly describe the DRL training setup.","section":"Figures 8-15 and Section VI"},{"comment":"The paper does not release code, training datasets, or the evaluation scripts, and the description of the training data is incomplete (e.g., no dataset size, no sampling procedure, no train/test split details beyond '80 epochs'). Since the paper's conclusions depend entirely on the reported empirical comparisons, the absence of reproducibility artifacts makes it impossible for a reviewer or reader to verify the claims. Releasing the code and data, or at least a detailed data card and training configuration, should be a condition for publication.","section":"Section VI (Experimental Setup) and overall reproducibility"}],"minor_comments":[{"comment":"The caption contains a typo: 'Answer Validiy' should be 'Answer Validity'.","section":"Figure 1 caption"},{"comment":"The caption is grammatically unclear: 'which is evident by the 1% parameters with a GPU of 28.23 GB' should be reworded, for example to 'which is evident from the 1% trainable parameters and the 28.23 GB GPU memory requirement'.","section":"Figure 2 caption"},{"comment":"The sentence 'The auto-regressive nature of LLMs' response generation, illustrated in presents significant challenges' is missing a noun after 'in'; it should refer to a specific figure or subsection.","section":"Section III"},{"comment":"The reward formula 'reward = Throughput / Latency + 1 - LossRate' is ambiguous; parentheses should clarify whether it is Throughput/(Latency+1) - LossRate or (Throughput/Latency) + 1 - LossRate, and the units of the quantities should be specified.","section":"Section IV.C.1"},{"comment":"The phrase 'see Figures 1b and 1a Additionally' is missing a period and should be split into two sentences; Figure 1a appears to show invalid outputs rather than accuracy, so the reference should be checked.","section":"Section VII.B"},{"comment":"The algorithms are presented in an informal pseudocode style and contain inconsistencies (e.g., Algorithm 3's teacher-forcing branch is not defined, and Algorithm 4's input mentions a 'TCP-related model' while the text says the pre-trained model is Llama2). Clarify the formal forward pass and the exact role of teacher forcing.","section":"Algorithms 1-4"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable proof-of-concept, but the central novelty claim is not yet backed by the experimental design. The missing random-init/non-LLM ablation and the internal throughput inconsistency are load-bearing and must be addressed before the paper can be considered for publication. The authors should also consider whether the DRL baseline from their prior work is a fair comparator and should make artifacts available."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a clean assembly of known pieces—frozen Llama2-7B, a CNN encoder, LoRA, and a linear head—for TCP congestion control algorithm selection over WiFi. What's genuinely new is the specific recipe: nobody has combined exactly these components for CCA fairness. The authors also report real testbed runs across three CCA-pair scenarios, not just simulations, and they show a 99% reduction in trainable parameters with a 0.015s single-step inference. Those are legitimate engineering contributions, and the writing is readable.\n\nThe soft spot is the one the reader flagged, and it's load-bearing: the paper claims the gains come from the LLM's pre-trained knowledge, but it never tests that. All trainable components—CNN, linear projections, LoRA matrices, the classification head—sit outside the frozen transformer. A randomly initialized transformer or a non-LLM model with the same encoder/head would isolate the effect, and that comparison is absent. So the evidence supports 'a small supervised network can imitate CCA selection,' which is already known, not 'the LLM's pre-training helps.' That's not a circular argument; the predictions are checked against the testbed. It's an attribution problem, and it's fixable.\n\nThere's also a small internal inconsistency: Scenario 1 reports TCP-LLM throughput as 42 Mbps for the BBR→Cubic transition, while Section VII.C says it balanced Cubic at 45 Mbps and BBR at 40 Mbps. That's the kind of number mismatch that makes a referee lose trust, even if it's a typo. No code or data are released, and there are no error bars, so the confidence intervals on the claimed improvements are unknown.\n\nThe citation pattern is fine—they cite NetLLM and other relevant work—so the issue isn't that they're ignoring prior art. It's that the evaluation doesn't support the headline claim.\n\nWho should read this: people working on LLMs for network control will find the recipe worth trying, and the paper gives enough detail to replicate. It doesn't deserve a desk reject; it deserves a serious referee who will ask for the ablation and the artifacts. I'd send it to review with the expectation of major revision.","headline":"A clean LLM-for-TCP recipe whose central claim about pre-training is untested; send to review but require an ablation and artifacts.","tokens_in":17829,"tokens_out":2193,"would_cite":false,"duration_ms":19592,"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 frozen Llama2-7B with a CNN encoder and a linear head can outperform DRL and standard CCAs at TCP fairness over WiFi by dynamically switching among Cubic, BBR, and PCC.","keywords":["TCP fairness","large language models","congestion control algorithm selection","WiFi networks","low-rank adaptation","starvation prevention","flow unfairness","deep reinforcement learning"],"falsifier":"Run TCP-LLM with a randomly initialized transformer of identical architecture, using the same CNN encoder, linear head, LoRA matrices, data, and training procedure, and compare accuracy plus the WiFi fairness metrics against the frozen-Llama2 version; if the random-initialization version performs as well, the paper's attribution of performance to pre-trained knowledge is not supported.","tokens_in":16789,"feed_emoji":"📶","tokens_out":6806,"duration_ms":55899,"temperature":0.7,"pith_summary":"The paper argues that a large language model can be repurposed, with only small trainable add-ons, to make real-time TCP decisions that fixed congestion-control algorithms and a deep-reinforcement-learning baseline handle poorly over WiFi. It claims that a frozen Llama2-7B, fronted by a CNN encoder that converts throughput, loss, RTT, and sending-rate histories into token-like embeddings, and steered by a linear head plus low-rank adapters, can pick among Cubic, BBR, and PCC to keep competing flows fair, prevent starvation, and avoid CCA incompatibility. If true, network stacks could gain adaptive congestion control without per-task neural-network design or costly full retraining, because the same LLM backbone serves several TCP tasks with lightweight fine-tuning.","feed_headline":"Frozen Llama2 beats deep RL at TCP fairness","feed_subtitle":"A CNN-encoded, LoRA-tuned Llama2-7B balances WiFi flows in 0.015 seconds, beating fixed CCAs and deep RL.","key_machinery":"The mechanism is an integrated encoder plus a frozen LLM plus lightweight adapters. A 1D-CNN extracts temporal patterns from a four-metric time series (throughput, loss rate, RTT, sending rate); linear projections convert those features into token-like embeddings; layer normalization stabilizes them; the frozen Llama2-7B processes the embeddings; a trainable linear head outputs a probability distribution over the available CCAs in one step; and low-rank matrices update the weights as $W = W_0 + AB$ during fine-tuning. This design bridges the text-to-numbers modality gap, avoids autoregressive token-by-token generation, and keeps trainable parameters at about 1%.","core_discovery":"The paper's central claim is that TCP-LLM outperforms state-of-the-art CCAs and a DRL baseline on WiFi fairness, starvation prevention, and CCA selection. In a lab testbed with a 100 Mbps bottleneck and a real WiFi access point, the frozen LLM with CNN encoder, linear head, and LoRA produced more balanced throughput, lower loss, and lower RTT than both fixed algorithms and the DRL model across three CCA-mixing scenarios. The paper also reports that TCP-LLM reaches near-zero loss and near-perfect accuracy within about 40 epochs, and that it makes decisions in a single 0.015-second inference step, while the DRL baseline takes over 50 epochs to stabilize and shows larger oscillations.","pith_inferences":["Because the paper never compares against a randomly initialized transformer or a non-LLM model with the same CNN encoder and linear head, a clean attribution test would be to run that control and see whether pre-training actually explains the gains.","The reported numbers come from one WiFi testbed with a 100 Mbps bottleneck; repeating the framework on 5G, satellite, or multi-hop paths would test whether the generalization claim extends beyond this topology.","The encoder-plus-frozen-LLM pattern may transfer to other numerical time-series control problems in networking, such as adaptive bitrate selection or multipath scheduling, but the paper does not demonstrate that.","The fair-sharing results on a single access point leave open how the approach would behave with many competing flows, mobile clients, or interference-driven loss, which are natural next experiments."],"forward_implications":["A single frozen LLM can be adapted to several TCP tasks by swapping the linear head and low-rank matrices, removing the need to design a new neural network per task.","CCA switching can run in real time because the linear head emits a decision in one forward pass at 0.015 seconds, well below the autoregressive latency the paper measures.","Fairness and starvation problems caused by BBR, Cubic, and PCC mixing can be reduced without retraining per scenario, since the paper reports near-perfect accuracy after about 40 epochs.","Adapting a 7-billion-parameter LLM to TCP becomes affordable: 99% of parameters are frozen and GPU memory drops from 65.88 GB to 28.23 GB.","The same framework can be applied to prediction tasks as well as decision-making tasks, because both use the same encoder and low-rank adaptation pipeline with different loss functions."],"supporting_citations":[{"why":"It supplies the frozen Llama2-7B backbone whose pre-trained knowledge TCP-LLM is designed to exploit.","marker":"[27]"},{"why":"It provides the low-rank adaptation method that keeps trainable parameters near 1% and enables task-specific fine-tuning of the frozen model.","marker":"[9]"},{"why":"It is the deep-reinforcement-learning baseline the paper extends and compares against for fairness over WiFi.","marker":"[26]"},{"why":"It defines BBR, one of the congestion-control algorithms TCP-LLM must coexist with and choose among.","marker":"[5]"},{"why":"It defines Cubic, the other main baseline CCA in the fairness experiments.","marker":"[8]"},{"why":"It defines PCC, the third candidate CCA used in the switching scenarios.","marker":"[7]"},{"why":"It supports the data-driven offline reinforcement-learning pipeline TCP-LLM uses for decision-making tasks.","marker":"[18]"},{"why":"It supplies the Transformer architecture on which the LLM backbone and its attention-based processing depend.","marker":"[28]"}],"fun_headline_variants":["Frozen LLM tames WiFi TCP unfairness in 15 ms","LoRA-tuned Llama2 beats deep RL at TCP fairness","LLM outperforms deep RL on WiFi flow fairness","One LLM inference fixes WiFi TCP fairness"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the frozen Llama2-7B's pre-trained knowledge is what produces the reported gains, rather than the small trainable CNN encoder, linear projection, and head; the paper never tests a randomly initialized transformer or a non-LLM control, so if pre-training contributes little, the central claim collapses into ordinary supervised CCA imitation.","fun_headline_variants_meta":{"raw":{"variants":["Frozen LLM tames WiFi TCP unfairness in 15 ms","LoRA-tuned Llama2 beats deep RL at TCP fairness","LLM outperforms deep RL on WiFi flow fairness","One LLM inference fixes WiFi TCP fairness"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000135,"raw_usage":{"total_tokens":1070,"prompt_tokens":801,"completion_tokens":269,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":417,"completion_tokens_details":{"reasoning_tokens":201}},"tokens_in":417,"tokens_out":269,"duration_ms":2964,"temperature":1.0,"reasoning_tokens":201,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:56:32.106858+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run TCP-LLM with a randomly initialized transformer of identical architecture, using the same CNN encoder, linear head, LoRA matrices, data, and training procedure, and compare accuracy plus the WiFi fairness metrics against the frozen-Llama2 version; if the random-initialization version performs as well, the paper's attribution of performance to pre-trained knowledge is not supported.","supporting_citations":[{"cited_title":"On the fairness of internet congestion control over wifi with deep reinforcement learning","cited_arxiv_id":null,"evidence_quote":"It is the deep-reinforcement-learning baseline the paper extends and compares against for fairness over WiFi."},{"cited_title":"Bbr: congestion-based congestion control","cited_arxiv_id":null,"evidence_quote":"It defines BBR, one of the congestion-control algorithms TCP-LLM must coexist with and choose among."},{"cited_title":"Cubic: a new tcp-friendly high- speed tcp variant","cited_arxiv_id":null,"evidence_quote":"It defines Cubic, the other main baseline CCA in the fairness experiments."},{"cited_title":"{PCC} vivace:{Online- Learning} congestion control","cited_arxiv_id":null,"evidence_quote":"It defines PCC, the third candidate CCA used in the switching scenarios."}],"review_version":1}