{"id":"5a811171-331c-4284-9773-66d3779f6151","arxiv_id":"2505.06378","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A multi-agent reinforcement learning algorithm with pruning is proposed for migrating vehicular AI 'twins' between roadside units, but the equilibrium proof is flawed and experiments are not reproducible.","lead":"This paper proposes a learning-based system that decides how self-driving cars buy bandwidth from roadside computers when moving heavy AI tasks off the car. The authors add a game-theory model and a pruning trick to speed up the learning, but the core math has errors that undermine the main claim.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (10) misdifferentiates the follower utility, so the claimed unique Stackelberg equilibrium does not follow.","rationale":"The paper's central claim is a unique Stackelberg equilibrium and a DRL algorithm that approximates it. The most load-bearing step is Eq. (12), the follower's best response derived from Eq. (5). It is algebraically incorrect, and the error propagates through the leader's problem and Theorem 1. This is an internal inconsistency, not a matter of disagreement with prior work. The reader's weakest_assumption focused on the undefined DRL reward; that is a valid secondary concern, but the equilibrium proof fails independently of the reward specification. I therefore keep the reader's REJECT verdict unchanged. The concrete derivative check is inexpensive and would settle the matter; no code or data is needed to expose the error. Independent support in the paper is limited: no reproducibility artifacts, no formal verification, and the experiments report no error bars.","tokens_in":16918,"tokens_out":10880,"duration_ms":92925,"concrete_test":"Re-derive ∂U_F_v/∂b_rv from Eq. (5) using a symbolic differentiator (e.g., SymPy). If the result differs from Eq. (10) by the presence of the α_v β w_r prefactor on −p_r, compute the correct first-order condition and check whether it reduces to Eq. (12). A numeric instantiation (α_v=3, β=1, two RSUs with p_r=p_l=1, T_v=1) gives w_r=0.5; the correct best response is 1.5−e/3 ≈ 1.409, while Eq. (12) gives 1−e/3 ≈ 0.909. If the corrected best response changes the leader's maximizer in Eqs. (14)–(18), then Theorem 1's uniqueness conclusion is unsupported and the paper's theoretical contribution fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (10) miscomputes the partial derivative of U_F_v in Eq. (5) with respect to b_rv. With w_r=(1/p_r)/Σ_l(1/p_l), U_F_v=Σ_r[α_v β w_r ln(e+α_v b_rv/T_v) − p_r b_rv]. The correct derivative is Σ_r[α_v β w_r · (α_v/T_v)/(e+α_v b_rv/T_v) − p_r], but Eq. (10) writes Σ_r[α_v β w_r((α_v/T_v)/(e+α_v b_rv/T_v) − p_r)], applying the α_v β w_r prefactor to the −p_r term. Setting the correct derivative to zero yields b_rv = α_v β w_r/p_r − eT_v/α_v, not Eq. (12)'s b_hat_rv=1/p_r−eT_v/α_v. This error propagates into the leader's utility in Eq. (14) and the uniqueness proof of Theorem 1. Additionally, Eq. (17) sums per-follower solutions of a single quadratic rather than solving the aggregate first-order condition, so the leader best response is also not derived. The paper's central claim—a unique Stackelberg equilibrium under α_v≥e p_r T_max_v—is therefore not established. The reader's concern about the unspecified DRL reward is secondary: even a reward exactly equal to U_F and U_L would not recover an equilibrium whose defining equations are incorrect.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper models the migration of vehicular embodied AI agent twins (VEAATs) from autonomous vehicles to roadside units as a multi-leader multi-follower Stackelberg game over bandwidth pricing and bandwidth demand. The authors propose a Bi-LSTM multi-agent PPO algorithm (TMABLPPO) with an NTK-based Path eXclusion pruning scheme, claim a unique Stackelberg equilibrium under α_v ≥ e p_r T_max^v, and report simulation results showing improved rewards and RSU/AV utilities. The central theoretical claim is the existence of this equilibrium and the assertion that the DRL algorithm approximates it.","tokens_in":17226,"tokens_out":7481,"duration_ms":73448,"significance":"Should the equilibrium derivation and the reward design be correct, the paper would offer a useful integration of game-theoretic resource allocation with computation-aware DRL compression for edge-assisted vehicular AI migration. The system model is topical, the use of QoS-aware logarithmic utilities is reasonable, and the experiments are conducted on an embedded platform with a pruning method that is at least concretely specified. However, the mathematical analysis establishing the Stackelberg equilibrium contains a derivative error that propagates through the leader-level proof, and the DRL reward is never defined. These issues are load-bearing: without a correct best-response derivation and a reward that matches the utilities, the claimed approximation of the Stackelberg equilibrium is not established. The paper's main contribution is therefore conditional and, as written, not supportable.","major_comments":[{"comment":"Eq. (10) misdifferentiates the follower utility in Eq. (5). With θ_rv = (1/p_r) / Σ_l (1/p_l), the derivative of U_F_v with respect to b_rv is Σ_r [α_v β θ_rv (α_v/T_max_v) / (e + α_v b_rv/T_max_v) − p_r], but Eq. (10) applies the prefactor α_v β θ_rv to both the logarithmic term and the −p_r term. The resulting first-order condition therefore cannot yield Eq. (12); the correct best response would be b_rv = α_v β θ_rv / p_r − e T_max_v / α_v, not b̂_rv = 1/p_r − e T_max_v / α_v. Because Eq. (12) is substituted into Eq. (14) and used in the proof of Theorem 1, the claimed uniqueness of the Stackelberg equilibrium does not follow.","section":"Section IV-A, Eqs. (10) and (12)"},{"comment":"Eq. (17) is not the solution of the leader's aggregate first-order condition. Eq. (15) sets ∂U_L^r/∂y_r = Σ_v [ ... ] = 0, which requires solving a sum over all followers; Eq. (17) instead writes ŷ_j as a sum over v of the individual quadratic solutions. These operations are not equivalent unless the cross-terms vanish, and no justification is given. Consequently, the leader best-response function G(Y), the standard-function argument, and the claimed unique Nash equilibrium of the leader-level game are not established.","section":"Section IV-B, Eq. (17)"},{"comment":"The reward function for the MADRL agents is never defined. Section V-B only states that RSUs and AVs generate 'rewards that combine immediate operational outcomes and long-term sustainability metrics,' and Algorithm 1 invokes R_t^r and R_t^v without equations. The paper's claim that TMABLPPO approximates the Stackelberg equilibrium requires the training reward to coincide with U_F and U_L from Eqs. (5) and (6), or to be a proxy whose optimum provably coincides with the equilibrium. With no reward specification, the numerical comparisons in Section VI demonstrate only that one DRL variant obtains a higher total of an unstated objective; they cannot validate equilibrium approximation.","section":"Section V-B and Algorithm 1"}],"minor_comments":[{"comment":"The logarithmic term in Eq. (5) uses α_r b_rv, while the task importance parameter is defined as α_v; the subscript appears to be a typo and should be α_v (or α_r must be defined).","section":"Section III-C1, Eq. (5)"},{"comment":"The latency constraint Σ_{j∈R} T_rv ≤ T_max^i mixes the indices r, v, and i; it should be a per-pair constraint such as T_rv ≤ T_max^v, and the sum over RSUs is unclear because T_rv already depends on r.","section":"Section III-C2, Eq. (7)"},{"comment":"The best response in Eq. (18) includes a zero branch, yet the subsequent standard-function proof asserts positivity of G(Y); the paper should clarify the domain on which positivity holds and correct the strict/weak inequality in the threshold condition.","section":"Section IV-B, Eq. (18)"},{"comment":"The saliency expression ∂R/(∂θ^2) ⊙ θ^2 is ill-formed; it should presumably be (∂R/∂θ)^2 ⊙ θ^2. The dimension statement for the NTK in Eq. (29), written as an NK×NK matrix, is also confusing and should be clarified.","section":"Section V-C, Eq. (28)"},{"comment":"The numerical section omits key experimental parameters (numbers of AVs and RSUs, α_v, β, c_r, p_max, and the simulator's transition dynamics) and reports no error bars or multiple seeds, which limits reproducibility and makes it difficult to interpret the reward comparisons in Figs. 3–8.","section":"Section VI"}],"recommendation":"reject","confidential_remarks":"To the editor: the manuscript's theoretical novelty rests on a Stackelberg equilibrium analysis that contains a fundamental derivative error, and the DRL reward is never specified. The numerical experiments therefore cannot support the central claim. The paper also shares substantial methodological structure with the authors' earlier IoT-J paper [26] (multi-leader multi-follower Stackelberg game plus tiny multi-agent DRL); the present contribution would need a corrected equilibrium derivation and a fully specified reward model before it could be evaluated fairly. I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the VEAAT migration paper. The headline is that the central analytic claim does not hold up: the derivative in Eq. (10) is not the derivative of the utility in Eq. (5). With w_r = (1/p_r)/Σ_l(1/p_l), U_F_v = Σ_r [α_v β w_r ln(e+α_v b_rv/T_v) − p_r b_rv]. The correct partial is α_v β w_r · (α_v/T_v)/(e+α_v b_rv/T_v) − p_r, but Eq. (10) applies the prefactor to the −p_r term as well. That makes the closed-form best response in Eq. (12) wrong (the correct root is b_rv = α_v β w_r/p_r − eT_v/α_v, not 1/p_r − eT_v/α_v), and the condition α_v ≥ e p_r T_max_v in Eq. (13)/Theorem 1 does not follow. The leader-level analysis and uniqueness proof are built on that same incorrect substitution, and Eq. (17) appears to sum per-follower solutions of a quadratic rather than solving the aggregate first-order condition. So the unique SE is not established.\n\nWhat the paper does well: the system model is thoughtfully assembled—latency model with channel estimation, a Stackelberg formulation with QoS-weighted utility, and a reasonable plan to personalize pruning rates by computation tier. The idea of adapting the PX pruning saliency to actor networks and tuning sparsity per vehicle class is a sensible engineering contribution, and the presentation is mostly clear. If someone ignores the game-theoretic proof, the DRL-with-pruning recipe could be a useful starting point.\n\nThe soft spots beyond the algebra: the DRL reward is never specified (Section V.B gives a vague sentence, Algorithm 1 computes R_t but no equation), so the claim that TMABLPPO approximates the SE is unverifiable. The experiments are all on a simulator built from the same game model, with no error bars, no ablations on random seeds, no code or data, and only reward curves. The pruning comparisons are interesting but also internal. The novelty is incremental—the framework closely mirrors the authors' earlier UAV twin migration work, and the core components are off-the-shelf—though the pruning-tier adaptation is new.\n\nVerdict: this is a reject as-is. The paper is for readers working on vehicular edge resource allocation who want a template for combining Stackelberg games with MADRL and pruning; they can use the system model and pruning idea, but should not rely on the equilibrium analysis. I would not send it to an external referee in its current state—the wrong derivative is a load-bearing flaw that a serious editor can catch directly. If the authors fix the utility/derivative inconsistency, redo the equilibrium analysis, and provide concrete reward definitions and reproducible experiments, it becomes a borderline workshop paper. As it stands, I would desk-reject.","headline":"The paper's central equilibrium proof is wrong—Eq. (10) misdifferentiates the follower utility—so the claimed unique Stackelberg equilibrium does not follow; the DRL and pruning pieces are standard and the experiments are not independently verifiable.","tokens_in":17776,"tokens_out":3374,"would_cite":false,"duration_ms":29914,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91A65"],"pacs":[],"model":"deepseek-v4-flash","headline":"In vehicular embodied AI networks, offloading an AI agent to a roadside unit has a unique Stackelberg equilibrium when the task's importance exceeds the product of bandwidth price and delay tolerance, and a Bi-LSTM multi-agent learner…","keywords":["vehicular embodied AI networks","AI twin migration","Stackelberg game","multi-agent deep reinforcement learning","bidirectional LSTM","network pruning","Path eXclusion","bandwidth allocation"],"falsifier":"Run TMABLPPO in the described environment, record the converged bandwidth requests and RSU prices, and compare them with the analytic best responses $b^\\star_{rv} = 1/p_r - e T^{\\max}_v/\\alpha_v$ and the leader's first-order condition across the regime $\\alpha_v \\ge e p_r T^{\\max}_v$; a systematic gap between learned and analytic actions would refute the claim that the algorithm approximates the equilibrium.","tokens_in":16731,"feed_emoji":"🚗","tokens_out":5799,"duration_ms":53895,"temperature":0.7,"pith_summary":"The paper argues that migrating vehicular embodied AI agents to roadside units can be treated as a multi-leader multi-follower Stackelberg game in which roadside units set bandwidth prices and vehicles respond with bandwidth requests. Its main theoretical claim is that this game has a unique Stackelberg equilibrium once each task's importance is high enough relative to the price and the vehicle's delay tolerance. On top of that equilibrium analysis the paper builds a multi-agent reinforcement learning algorithm using bidirectional LSTMs, and a computation-aware pruning step that shrinks each vehicle's actor network according to its onboard computing power. The aim is to keep roadside-unit workloads balanced while keeping migration latency low, and the reported simulations on embedded hardware show the pruned learner outperforming baselines in cumulative reward.","feed_headline":"Stackelberg theorem fixes price and bandwidth for vehicle AI migration","feed_subtitle":"A unique equilibrium balances roadside-unit load and latency; a pruned multi-agent learner reaches it on embedded hardware.","key_machinery":"The central object is the MLMF Stackelberg game with QoS-weighted utilities: AV utility is a logarithmic Weber–Fechner revenue term minus bandwidth cost, and RSU utility is bandwidth revenue minus base cost. The argument is carried by the closed-form follower best response and by a standard-function proof for the leader best response, which together establish the unique equilibrium condition $\\alpha_v \\ge e p_r T^{\\max}_v$. To reach that equilibrium in a partially observable setting, the algorithm uses Bi-LSTM actors that ingest the previous $L$ rounds of prices and demands, a PPO-style clipped surrogate loss, and a Path eXclusion pruning step that scores weights by a path-activation by path-weights product so the pruned subnetworks keep the task-critical pathways.","core_discovery":"Under the condition $\\alpha_v \\ge e p_r T^{\\max}_v$, the paper derives a unique Stackelberg equilibrium $(\\mathbf{B}^\\star,\\mathbf{P}^\\star)$ for the AV–RSU bandwidth market. The follower-level best response is closed form: vehicle $v$ requests $b^\\star_{rv} = 1/p_r - e T^{\\max}_v/\\alpha_v$ from RSU $r$ when the condition holds, and requests zero bandwidth otherwise. The leader-level best response is shown to be a standard function, which implies a unique Nash equilibrium among RSUs and therefore a unique Stackelberg equilibrium. The TMABLPPO algorithm is then proposed to approximate this equilibrium through decentralized coordination, and the Path eXclusion pruning method compresses the actor networks so that vehicles with different computation tiers execute the learned policies within latency constraints.","pith_inferences":["The analytic equilibrium could be used directly as a reward-shaping term or a curriculum prior for the DRL training, which would make the claimed approximation testable independently of the simulation's reward design.","The same leader–follower structure applies to other digital-twin migration markets, such as UAV or metaverse service migration, so the uniqueness theorem transfers wherever utility has the same log-revenue-minus-cost form.","One could replace the fixed density tiers with an online latency–accuracy Pareto search per vehicle, since the paper's three-tier thresholds are a heuristic discretization of a continuous trade-off.","The PX saliency's reliance on the neural tangent kernel suggests a principled connection between pruning and trainability: weights on high-curvature paths are kept, which may explain why 90%-density pruning does not degrade reward."],"forward_implications":["Roadside units can use the equilibrium condition as an admission rule: vehicles whose task importance falls below $e p_r T^{\\max}_v$ should buy no bandwidth from that RSU, which offloads the cheapest traffic automatically.","The closed-form best response gives a direct benchmark for any learned policy; a DRL agent that does not converge to $b^\\star_{rv}$ in the stated regime is not solving the game as modeled.","Bi-LSTM temporal context shortens training to reach optimal strategies, so the same reward signal yields faster convergence than one-step observation baselines.","Pruning at 90% density preserves or improves cumulative reward, so the latency gain from compression does not cost task performance in the tested regime.","At 33% density the model keeps about 68% of its original reward, indicating graceful degradation under aggressive compression."],"supporting_citations":[{"why":"Supplies the Stackelberg-game-plus-MADRL template for vehicular twin migration that this paper extends with QoS utilities.","marker":"[11]"},{"why":"Defines the Path eXclusion pruning algorithm that the computation-aware pruning step is built on.","marker":"[16]"},{"why":"Provides the Bi-LSTM architecture used to capture bidirectional temporal dependencies in the actor networks.","marker":"[15]"},{"why":"Supplies the standard-function conditions that the leader-level uniqueness proof relies on.","marker":"[37]"},{"why":"Supplies the NTK-based weight-importance perspective used in the PX saliency derivation.","marker":"[38]"},{"why":"The MAPPO baseline whose PPO formulation underlies the actor-critic loss.","marker":"[39]"},{"why":"The asynchronous actor-critic baseline (MAA3C) used for comparison in the experiments.","marker":"[40]"},{"why":"The Weber–Fechner law motivates the logarithmic QoS revenue term in the AV utility function.","marker":"[34]"}],"fun_headline_variants":["Pruned DRL approximates unique Stackelberg equilibrium for AV migration","Unique Stackelberg equilibrium balances AV-RSU bandwidth and latency","Stackelberg equilibrium for AV bandwidth market via pruned DRL","Computation-aware pruning enables vehicle AI migration via Stackelberg equilibrium"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reward the simulator gives to the DRL agents equals the AV and RSU utility functions used in the equilibrium analysis; the paper never states that reward, so the connection between the learned policies and the claimed Stackelberg equilibrium rests on this unstated matching.","fun_headline_variants_meta":{"raw":{"variants":["Pruned DRL approximates unique Stackelberg equilibrium for AV migration","Unique Stackelberg equilibrium balances AV-RSU bandwidth and latency","Stackelberg equilibrium for AV bandwidth market via pruned DRL","Computation-aware pruning enables vehicle AI migration via Stackelberg equilibrium"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0016,"raw_usage":{"total_tokens":6397,"prompt_tokens":991,"completion_tokens":5406,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":5331}},"tokens_in":607,"tokens_out":5406,"duration_ms":36355,"temperature":1.0,"reasoning_tokens":5331,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:45:18.852392+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run TMABLPPO in the described environment, record the converged bandwidth requests and RSU prices, and compare them with the analytic best responses $b^\\star_{rv} = 1/p_r - e T^{\\max}_v/\\alpha_v$ and the leader's first-order condition across the regime $\\alpha_v \\ge e p_r T^{\\max}_v$; a systematic gap between learned and analytic actions would refute the claim that the algorithm approximates the equilibrium.","supporting_citations":[{"cited_title":"When metaverses meet vehicle road cooperation: Multi-agent DRL-based stackelberg game for vehicular twins migration,","cited_arxiv_id":null,"evidence_quote":"Supplies the Stackelberg-game-plus-MADRL template for vehicular twin migration that this paper extends with QoS utilities."},{"cited_title":"Finding lottery tickets in vision models via data-driven spectral foresight pruning,","cited_arxiv_id":null,"evidence_quote":"Defines the Path eXclusion pruning algorithm that the computation-aware pruning step is built on."},{"cited_title":"Privacy- preserving incentive mechanism for multi-leader multi-follower IoT- edge computing market: A reinforcement learning approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the standard-function conditions that the leader-level uniqueness proof relies on."},{"cited_title":"Ntk-guided few-shot class incremental learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the NTK-based weight-importance perspective used in the PX saliency derivation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The Weber–Fechner law motivates the logarithmic QoS revenue term in the AV utility function."}],"review_version":1}