{"id":"6cb7cd4a-11c3-45b5-b62c-61aae09c5790","arxiv_id":"2501.01141","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"A proposed framework uses LLAVA to convert vehicle camera images into text and GAE-PPO to adapt transmission, claiming up to 36% higher QoE than DDPG and over 90% data reduction, though the QoE metric is unvalidated.","lead":"This paper combines a large vision-language model (LLAVA) with a reinforcement learning method (GAE-PPO) to shrink images from vehicles into short text descriptions and then choose how to transmit them. The authors report large bandwidth savings and higher quality-of-experience than standard baselines, but the quality metric is defined by the authors and not validated.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (13)'s QoE increases as semantic similarity decreases, so the headline QoE gains are measured on an internally inverted objective.","rationale":"The reader correctly identified the QoE metric in Eq. (13) as the load-bearing premise. My stress-test sharpens this: the metric is not merely unvalidated against external QoE standards; it is internally inconsistent with the paper's own definition of quality. Since Θ_a(x)=x and the argument is (1−ξ), QoE decreases as semantic similarity increases, so the optimization and reward push toward the constraint boundary ξth rather than toward high semantic fidelity. This inversion alone invalidates the headline quantitative claims, because both the 36% gain over DDPG and the 61.4% scalability gain are improvements measured on this objective. The proposed concrete test is computational and would settle the issue immediately. In good faith, the paper's component-level observations (LLAVA text extraction and attention maps, GAE-PPO convergence behavior) may survive, but the central claim as stated does not. I therefore keep the reader's REJECT verdict, with the concern reframed from an external-validity objection to an internal-correctness objection.","tokens_in":21414,"tokens_out":3420,"duration_ms":34513,"concrete_test":"Evaluate Eq. (13) for a single active link with fixed positive power at ξ=1 and ξ=0: QoE(ξ=1)=0 while QoE(ξ=0)=sigmoid(P)>0, demonstrating that the metric rewards complete semantic mismatch. Then rerun the Fig. 8 comparison with the corrected quality function QoE'=Σ Θ_a(ξ)Θ_b(...) (or a properly derived Weber-Fechner form) under the same MDP; if GAE-PPO no longer shows the 36% and 61.4% gains, those headline numbers are artifacts of the inverted reward. Also record the learned ξ values to check whether the policy drives them to the ξth=0.3 boundary.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equations (13) and (27) define QoE (and reward) as Σ_w Σ_q Θ_a(1−ξ_q[w]) Θ_b(b_q[w] P_q^{V2V}[w]), with Θ_a(x)=x and Θ_b(x)=1/(1+e^{−x}) (Sec. 5.1.1). For any active link with positive power, this is an increasing function of (1−ξ), i.e., decreasing in the semantic similarity ξ that the paper treats as the quality measure (Eq. (12), constraint (14f)). Perfect reconstruction (ξ=1) contributes 0; a completely dissimilar message (ξ=0) contributes the maximum sigmoid(power). Thus the DRL reward actively pushes semantic similarity down to the lower boundary ξth=0.3, and the 'up to 36% QoE improvement' compares policies under an objective that rewards semantic loss. Moreover, because the sum over q includes links with b_q[w]=0, inactive links contribute Θ_b(0)=0.5·(1−ξ), further uncoupling the metric from actual transmission quality. This is an internal correctness problem, not just a missing external validation: even if the Weber-Fechner claim were accepted, Eq. (13) does not express the stated objective. The later substitution of SSIM for ξ in Fig. 10 compounds the ambiguity but is secondary to this inversion.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an embodied-AI framework for vehicular networks in which LLAVA converts camera images into compact textual semantic messages and a GAE-PPO agent selects V2V power, subband, and semantic-symbol allocations. The authors formulate an optimization problem over a QoE metric that they state is inspired by the Weber-Fechner law, and they report that the proposed transmission strategy improves QoE by up to 36% over DDPG, converges faster than pure PPO, and reduces transmitted data size by more than 90% via LLAVA. The paper also compares LLAVA-1.5-7b against four vision-language baselines on parameter count, inference time, and visual-question-answering accuracy, and analyzes convergence, QoE scaling, and the effect of semantic sentence length.","tokens_in":1753,"tokens_out":2181,"duration_ms":89435,"significance":"If validated, the framework would address a real problem: reducing raw image transmission in bandwidth-limited vehicular networks while retaining task-relevant semantic content. The LLAVA compression figures and the systematic comparison of vision-LLM baselines are useful empirical data points, and using GAE to stabilize PPO in this setting is a sensible design choice. However, the central quantitative claims rest on Eq. (13), which is both unvalidated against any external QoE or perceptual standard and internally inconsistent, since it rewards decreasing semantic similarity. The absence of error bars and the undefined fine-tuning loss further prevent the reported improvements from being accepted as stated. The paper therefore does not currently establish its main claims.","major_comments":[{"comment":"The QoE and reward are defined as the double sum of Theta_a(1 - xi_q[w]) Theta_b(b_q[w] P_q^V2V[w]), with Theta_a(x) = x and Theta_b(x) = 1 / (1 + e^{-x}) set in Sec. 5.1.1. Since xi = 1 is perfect similarity and xi = 0 is no similarity (Eq. (12)), this quantity is strictly increasing in (1 - xi) for any active link with positive power: perfect reconstruction contributes zero, and completely dissimilar content contributes the maximum sigmoid(power). The DRL agent is therefore rewarded for driving xi down to the constraint boundary xi_th = 0.3, and the headline 36% QoE gain compares policies under an objective that penalizes semantic fidelity. Moreover, because the sum includes links with b_q[w] = 0, inactive links contribute Theta_b(0) = 0.5 times (1 - xi), so the metric is nonzero even when nothing is transmitted. The metric should be a non-decreasing function of xi, or the paper must provide a derivation that justifies the inversion; as written, the central claim is internally inconsistent.","section":"Sec. 3.3, Eq. (13); Sec. 4.2.2, Eq. (27)"},{"comment":"The paper states that Eq. (13) is inspired by the Weber-Fechner law, but it never derives Eq. (13) from that law, and the formula is not in the logarithmic Weber-Fechner form. The shaping functions Theta_a and Theta_b are declared in Sec. 5.1.1 without justification or sensitivity analysis, and no comparison against human-perception data or a standard QoE dataset is provided. Consequently, the reported 36% and 61.4% improvements are improvements on an unvalidated, author-defined objective; the citation to psychological QoE work does not establish validity.","section":"Sec. 3.3; Sec. 5.1.1"},{"comment":"The LLAVA fine-tuning loss is written as min over theta of sum_i ||M_i - M_hat_i||^2, but M_i and M_hat_i are natural-language semantic messages (text descriptions) and no embedding, tokenizer, or vector representation is specified in the paper. An L2 norm on raw text is therefore undefined, making the fine-tuning procedure non-reproducible. The authors should specify the representation (for example, BERT embeddings) and the exact loss used in the implementation.","section":"Sec. 4.1.2, Eq. (18)"},{"comment":"The simulations are reported as single trajectories with no error bars, confidence intervals, or seeds, so the claimed gains (61% accumulated-return gain, 36% QoE gain, up to 47% fewer convergence steps, and 61.4% scaling improvement) cannot be distinguished from random variation. In addition, the abstract and contributions claim improved decision accuracy, but no experiment measures a downstream decision-making task; Fig. 6 only reports visual-question-answering output correctness. The authors should define and measure decision accuracy, or remove that claim.","section":"Sec. 5, Figs. 7-10"},{"comment":"The threshold xi_th is described as an SSIM threshold and Fig. 10 plots SSIM versus SINR and sentence length, whereas Eq. (12) defines xi as BERT cosine similarity between text messages. SSIM is an image-quality metric and cannot be applied directly to the text-based semantic similarity used in Problem (14); the relationship shown in Fig. 10 therefore does not validate constraint (14f). The metric terminology should be unified and the actual quantity plotted should match the quantity used in the optimization.","section":"Sec. 5.1.1, Fig. 10"}],"minor_comments":[{"comment":"The phrase 'To adopt LLAVA' should be 'To adapt LLAVA', and the Algorithm 1 title contains a stray space in 'V ehicular'.","section":"Sec. 4.1.2; Algorithm 1"},{"comment":"The V2V power is described as continuous in [-100, 23] dBm in one place and as the discrete set {-100, 5, 15, 23} dBm in Table 1; please reconcile whether power control is continuous or discrete.","section":"Sec. 5.1.1; Table 1"},{"comment":"The y-axes are labelled 'Accumulated return' and 'Achievable QoE' with no units or confidence bands; report the mean and standard deviation over multiple random seeds.","section":"Fig. 7; Fig. 8"},{"comment":"The text says the target value uses n-step returns, but the displayed formula V_tar(s_t) = r(s_t,a_t) + gamma V_theta_C(s_{t+1}) is a one-step bootstrap; align the description with the implementation.","section":"Sec. 4.2.1, Eq. (24)"},{"comment":"The text reports 'up to 47%' fewer convergence steps, while Fig. 9 lists absolute step reductions of 10, 23, and 54; report percentages together with the baseline totals so the comparison is unambiguous.","section":"Sec. 5.2, Fig. 9"}],"recommendation":"reject","confidential_remarks":"The major issue is an internal inconsistency rather than a matter of taste or consensus: Eq. (13) directly contradicts the paper's own definition of xi as semantic quality. Because the headline QoE numbers are produced by optimizing this inverted objective, I do not see a local edit that would make the current claims stand; the numerical study would need to be redone with a correctly specified and externally validated metric, and the decision-accuracy claim is unmeasured. I would therefore not invite a minor revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the thing to know before you spend time on this: the central QoE metric in Eq. (13) is backwards. With Θ_a(x)=x and Θ_b(x)=1/(1+e^{-x}), the term Θ_a(1−ξ) equals 1−ξ, so QoE increases as semantic similarity ξ decreases. Perfect reconstruction scores zero; a meaninglessly different message scores the maximum. The penalty in Eq. (28) only keeps ξ above the 0.3 threshold, so the RL agent learns to deliver the worst allowed quality, and the reported 36% QoE improvement over DDPG is an improvement on a metric that rewards destroying semantic content. That is an internal correctness problem, not a missing external validation. What is genuinely new: the integration of LLAVA for image-to-text semantic extraction with GAE-PPO for transmission control in a vehicular setting is, as far as I can tell, not in the prior work. The benchmark of five vision-language models on parking-lot queries is a small but real empirical contribution, and the raw data-size reduction (614 KB image to 12.1 KB text) is a concrete, plausibly reproducible result. The convergence-speed gain of GAE-PPO over pure PPO is standard but cleanly demonstrated. The soft spots beyond the inverted metric: Eq. (18) uses an L2 norm on text messages, which is not a vector space—that loss is undefined. Section 5.1.1 calls ξ_th the SSIM threshold even though ξ was defined as BERT cosine similarity, and Fig. 10 plots SSIM as if it were the same quantity. No code or data are released, and the simulation figures have no error bars, so the magnitude claims (47%, 61.4%) rest entirely on the authors' own runs. I would not send this to peer review in its present form. The central performance measure is self-contradictory, and even a fix to the sign would require re-running the entire experiment. The framework idea has some merit, and a cleaned version with a correctly defined QoE could become a modest extension worth a second look. But as it stands, the headline conclusion is not supported.","headline":"The paper's QoE metric is inverted—it rewards lower semantic similarity—so the headline 36% gain measures degradation, not quality.","tokens_in":621,"tokens_out":2497,"would_cite":false,"duration_ms":48040,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An LLAVA-plus-GAE-PPO framework shrinks vehicular image transmission by over 90% while improving QoE by up to 36%.","keywords":["embodied AI","vehicular networks","semantic communication","LLAVA","deep reinforcement learning","quality of experience","Weber-Fechner law","GAE-PPO"],"falsifier":"Replace the BERT cosine similarity in Eq. (12) with a task-level measure: have the receiver run an obstacle-detection or lane-change decision from the recovered text and compare success rates against decisions made from the original image. If a panel or automated benchmark shows no better-than-baseline task accuracy while Eq. (13) reports high QoE, the metric and the claimed gains are falsified.","tokens_in":21205,"feed_emoji":"🚗","tokens_out":5830,"duration_ms":49061,"temperature":0.7,"pith_summary":"The paper proposes an embodied-AI framework for vehicular networks in which each vehicle uses the LLAVA vision-language model to turn camera images into short structured text messages, and a generalized-advantage-estimation PPO agent decides transmission power, spectrum sharing, and semantic symbol length. The central claim is that this combination keeps decision-relevant content while shrinking transmitted data by more than 90%, and that the learned transmission policy improves quality of experience by up to 36% over DDPG while converging in fewer steps than pure PPO. The optimization target is a Weber-Fechner-law-inspired QoE metric that balances semantic similarity (BERT cosine) against bandwidth use. The work's significance is showing that LLM-based semantic compression and DRL-based resource control can be jointly trained in the same loop, which would matter if 6G vehicular networks must carry massive sensor data through limited spectrum.","feed_headline":"LLM semantic extraction cuts vehicular data by ~90%","feed_subtitle":"A vision-language model plus reinforcement learning lifts QoE by up to 36%.","key_machinery":"The load-bearing object is the Weber-Fechner-inspired QoE objective, $\\text{QoE} = \\sum_{w=1}^{W}\\sum_{q=1}^{Q} \\Theta_a(1-\\xi_q[w])\\,\\Theta_b\\!\\left(b_q[w]P_q^{\\mathrm{V2V}}[w]\\right)$, with $\\Theta_a(x)=x$ and $\\Theta_b(x)=1/(1+e^{-x})$, where $\\xi$ is the BERT cosine similarity between transmitted and recovered messages. This scalar metric converts semantic fidelity and transmit power into a single reward, making the non-convex resource allocation problem (14) amenable to reinforcement learning. Around it, LLAVA acts as the semantic encoder that compresses images to text, and GAE-PPO is the policy optimizer that maximizes cumulative QoE under SINR and similarity thresholds.","core_discovery":"On the paper's own terms, the discovery is that LLAVA-1.5-7b-hf compresses a 614 KB camera frame into a 12.1 KB text message while preserving the facts needed for driving decisions (vehicle counts, parking occupancy, lane incidents), and that a GAE-PPO agent optimizing the QoE objective reaches higher accumulated return and faster convergence than pure PPO or DDPG baselines. The quantitative claims are a QoE gain up to 36% over DDPG at eight vehicles, a 61.4% QoE improvement when scaling from four to eight vehicles, and convergence-step reductions of 10, 23, and 54 for three of four simulated vehicles compared with pure PPO. The authors present these as evidence that semantic communication plus reinforcement learning can jointly optimize bandwidth use and decision accuracy in dynamic vehicular environments.","pith_inferences":["The QoE metric itself is the least-tested link: a subjective or task-based validation study would likely be needed before the exact 36% number can be trusted.","The 90% compression claim presumes text retains all decision-critical information; a natural extension is to compare end-to-end driving-task accuracy under text-only versus raw-image transmission.","The same LLM-plus-DRL architecture could be transferred to roadside units, drone relays, or other embodied agents, since the framework does not depend on vehicle-specific physics.","The attention-map analysis hints at a further bandwidth saving: transmitting only the highlighted image regions, rather than full frames, could push compression beyond 90%."],"forward_implications":["If the framework's results hold, V2I links can transmit semantic text rather than raw video, cutting spectrum load by an order of magnitude while retaining the content that matters for driving decisions.","GAE-PPO's faster, more stable convergence than pure PPO would make DRL-based transmission control feasible in real time, where training episodes are cheap but decision latency matters.","The reported 61.4% QoE gain from four to eight vehicles suggests the method scales as network load grows, which is exactly the regime where spectrum is scarce.","Adapting semantic symbol or sentence length to SINR would give operators a tunable knob between fidelity and bandwidth, as Figure 10 indicates."],"supporting_citations":[{"why":"Supplies the Weber-Fechner law basis for the QoE metric in Eq. (13).","marker":"[36]"},{"why":"Provides the BERT-embedding cosine similarity in Eq. (12) that measures semantic fidelity.","marker":"[35]"},{"why":"Defines LLAVA, the vision-language model used for semantic extraction and fine-tuning.","marker":"[38]"},{"why":"Sets the vehicular-network scenario and the DRL baseline that the work extends.","marker":"[13]"},{"why":"Supplies the semantic encoder/decoder structure that transmits structured messages.","marker":"[25]"},{"why":"Provides the text corpus used in the semantic transmission evaluation.","marker":"[44]"},{"why":"Cited as the basis for generalized advantage estimation in the PPO updates.","marker":"[40]"},{"why":"Provides the driving-scenario image dataset used to fine-tune LLAVA.","marker":"[45]"}],"fun_headline_variants":["LLM vision cuts vehicular data 90%, RL boosts QoE 36%","Semantic LLM + RL: 90% less data, 36% better QoE","Vehicular AI: LLM shrinks data 90%, RL improves QoE 36%","LLM+RL for vehicles: 90% data cut, 36% QoE gain","Vision-language model + RL: 90% data drop, faster convergence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire performance comparison stands or falls on the assumption that the QoE formula in Eq. (13) captures real driving-related quality of experience; the paper asserts this via the Weber-Fechner law but does not derive the formula from it or validate it against human perception.","fun_headline_variants_meta":{"raw":{"variants":["LLM vision cuts vehicular data 90%, RL boosts QoE 36%","Semantic LLM + RL: 90% less data, 36% better QoE","Vehicular AI: LLM shrinks data 90%, RL improves QoE 36%","LLM+RL for vehicles: 90% data cut, 36% QoE gain","Vision-language model + RL: 90% data drop, faster convergence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000253,"raw_usage":{"total_tokens":1578,"prompt_tokens":972,"completion_tokens":606,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":497}},"tokens_in":588,"tokens_out":606,"duration_ms":4944,"temperature":1.0,"reasoning_tokens":497,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:34:50.812841+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the BERT cosine similarity in Eq. (12) with a task-level measure: have the receiver run an obstacle-detection or lane-change decision from the recovered text and compare success rates against decisions made from the original image. If a panel or automated benchmark shows no better-than-baseline task accuracy while Eq. (13) reports high QoE, the metric and the claimed gains are falsified.","supporting_citations":[{"cited_title":"Psyqoe: Improving quality-of- experience assessment with psychological effects in video stream- ing,","cited_arxiv_id":null,"evidence_quote":"Supplies the Weber-Fechner law basis for the QoE metric in Eq. (13)."},{"cited_title":"A wireless AI-generated content (AIGC) provisioning framework empowered by semant ic communication,","cited_arxiv_id":null,"evidence_quote":"Provides the BERT-embedding cosine similarity in Eq. (12) that measures semantic fidelity."},{"cited_title":"Visual instruction tuning,","cited_arxiv_id":null,"evidence_quote":"Defines LLAVA, the vision-language model used for semantic extraction and fine-tuning."},{"cited_title":"Semantic-Aware Spectrum Sharing in Internet of Vehicles Based on Deep Reinforcement Learning","cited_arxiv_id":"2406.07213","evidence_quote":"Sets the vehicular-network scenario and the DRL baseline that the work extends."},{"cited_title":"Contrastive learning-based se- mantic communications,","cited_arxiv_id":null,"evidence_quote":"Supplies the semantic encoder/decoder structure that transmits structured messages."},{"cited_title":"Europarl: A parallel corpus for statistical machine translation,","cited_arxiv_id":null,"evidence_quote":"Provides the text corpus used in the semantic transmission evaluation."},{"cited_title":"Constrained update projection approach to safe policy optimization,","cited_arxiv_id":null,"evidence_quote":"Cited as the basis for generalized advantage estimation in the PPO updates."},{"cited_title":"Improved baselines with visual instruc- tion tuning,","cited_arxiv_id":null,"evidence_quote":"Provides the driving-scenario image dataset used to fine-tune LLAVA."}],"review_version":1}