{"id":"631e2c8c-996a-4c15-bf69-d54bf9071771","arxiv_id":"2507.07418","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"An optimal Myerson-style auction is identified for single-slot joint advertising, and a neural network named BundleNet approximates it in single-slot tests and outperforms two existing baselines in most multi-slot tests.","lead":"This paper studies online ad auctions where a single slot is won by a pair of advertisers, a retailer and a supplier, rather than by one advertiser. It derives the revenue-optimal single-slot auction in closed form and proposes a neural-network mechanism, BundleNet, for the multi-slot case.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proof of the central optimality theorem is not sound as printed: Eq. (8) in Appendix B has the wrong sign on the IR constants, reversing the argument that sets low-type utility to zero. Without a corrected derivation, Theorem 4.3 is unproved.","rationale":"The central claim is Theorem 4.3, and its proof in Appendix B is the only support. The sign error in Eq. (8) is an internal inconsistency, not a scope issue: even under the paper's regularity and additivity assumptions, the displayed derivation does not yield the revenue-maximization conclusion. Because the intended formula is the standard Myerson one, I regard this as a correctable typo rather than evidence that the theorem is false. The reader's weakest_assumption about correlated or non-additive values is a legitimate external-validity caveat, but it is secondary; it does not identify this proof defect. The experimental anomalies in Table 1 (BundleNet revenue above the 'Optimal' baseline in U2) also deserve a check, but the proof gap is the more load-bearing concern for the paper's main claim. Since the theorem remains plausible and the fix is local, keeping the conditional-accept verdict is appropriate rather than accepting or rejecting outright.","tokens_in":16201,"tokens_out":25792,"duration_ms":308564,"concrete_test":"Re-derive Eq. (8) algebraically from Eq. (7) and Eq. (6). If the U_r(a_r,a_r) and U_s(a_s,a_s) terms acquire minus signs, correct Eq. (8) and confirm that setting them to zero maximizes revenue; if the sign is genuinely plus, the theorem's proof is invalid. Also add to Theorem 4.3 the explicit allocation rule 'allocate to the bundle maximizing c_r(v_r)+c_s(v_s), subject to v0' and verify that the stated step-function and critical-value conditions are equivalent to it.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is Lemma B.2/Theorem 4.3's reduction of expected revenue to virtual surplus. Substituting Eq. (7) into Eq. (6), the [U_r(a_r,a_r) + integral (1-F_r)x_r] terms carry a minus sign, so the correct decomposition is U0 = ∫ v0 f + ∑_e ∫ x_e(v_s+v_r-(1-F_r)/f_r-(1-F_s)/f_s-v0)f dv - ∑_r U_r(a_r) - ∑_s U_s(a_s). As printed, Eq. (8) instead has plus signs before the two U sums. With the printed sign, the text's statement that these non-negative terms should be set to zero to maximize revenue is backwards: minimizing them would maximize revenue. The conclusion that IR binds at the lowest type therefore does not follow from the displayed equations. Separately, Theorem 4.3 never states the allocation rule that assigns the slot to the highest-virtual-value bundle; Appendix B's 'This implicitly indicates...' is doing that equivalence without proof. Both defects are repairable, but as written the central claim is not demonstrated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies joint advertising auctions, in which an ad slot is allocated to a bundle consisting of a retailer and a supplier rather than to a single advertiser. The main theoretical contribution is Theorem 4.3, an 'if and only if' characterization of the revenue-maximizing mechanism for a single-slot joint advertisement under regularity, modeled as a Myerson-style single-parameter auction with additively combined virtual values. The paper then proposes BundleNet, a neural network architecture for multi-slot joint advertising, which uses bundle-level incentive-compatibility constraints and is evaluated against RVCG, JRegNet, and the single-slot optimal benchmark on synthetic data.","tokens_in":16446,"tokens_out":9174,"duration_ms":90604,"significance":"If Theorem 4.3 were properly established, it would provide a clean, exact Myerson-type solution for single-slot joint advertising, and it would serve as a useful theoretical benchmark for learning-based mechanisms. The BundleNet architecture is a reasonable extension of RegretNet to a bipartite bundle structure, and the idea of formulating IC constraints at the bundle level is motivated by the problem structure. The paper is useful in bringing these two threads together, but the printed proof of the central theorem contains a sign error in the revenue decomposition, and the 'if and only if' direction is not demonstrated. The empirical state-of-the-art claim is also stronger than the results in Table 5 support. The underlying ideas appear sound and repairable, but the manuscript as written does not yet establish its main claims.","major_comments":[{"comment":"The sign error in Eq. (8) breaks the argument that individual rationality binds at the lowest type. Substituting Eq. (7) into Eq. (6) yields U0 = ∫ v0 f + ∑_e ∫ x_e(v_s + v_r − (1−F_r)/f_r − (1−F_s)/f_s − v0) f dv − ∑_r U_r(a_r,a_r) − ∑_s U_s(a_s,a_s). As printed, Eq. (8) has plus signs before the two U sums. With the printed plus signs, the statement that these non-negative terms should be set to zero to maximize revenue is backwards; with the correct minus signs, the conclusion follows. Since this step is the origin of the payment rule in Eq. (9), Theorem 4.3 is not supported by the printed derivation.","section":"Appendix B, Eq. (8)"},{"comment":"The proof of the 'if and only if' claim is incomplete. After writing the virtual surplus objective, the text says 'This implicitly indicates that the slot will be allocated to the bundle with the highest virtual value,' but no formal equivalence is proved between the virtual-surplus maximizer and the step-function allocation with critical-value payments stated in Theorem 4.3. The necessity direction is absent, tie cases at v_i = v̂_i are not discussed, and the payment calculation is only summarized as 'we integrate the allocation rule.' In addition, the theorem does not explicitly state a bundle-level allocation rule, so the relation between x_i^M(v_i,v_−i) and the identity of the winning bundle is left informal.","section":"Appendix B, proof of Theorem 4.3"},{"comment":"The abstract and Section 6.2 claim that BundleNet achieves state-of-the-art performance, but Table 5 shows that in the N5×5 setting JRegNet's revenue (2.2071) exceeds BundleNet's (2.1393). The claim should be qualified as 'in most settings' or the state-of-the-art statement should be based on a defined aggregation across configurations. As written, the empirical claim is stronger than the data support.","section":"Table 5 (Appendix D.2)"}],"minor_comments":[{"comment":"The allocation rule is first defined as xe : V → 2^M, a set-valued mapping, but later xe(v) is treated as a scalar or a vector; this notation should be made consistent.","section":"Definition 3.1"},{"comment":"Eq. (4) sums over ℓ = 0 to L while the rest of the paper uses ℓ = 1 to L, and the Lagrange multipliers are denoted inconsistently as λ, µ, and μ across Section 5.3 and Algorithm 1.","section":"Eq. (4) and Algorithm 1"},{"comment":"In the last payment equation of Appendix B, the winner's payment is written as z_r(v_−r) without any definition of z_r; it should be v̂_r(v_−r) as in Theorem 4.3.","section":"Appendix B, final displayed equations"},{"comment":"The doubly stochastic construction uses matrices of size (n+1) × (m+1), but Lemma 5.2 is stated for n × m matrices; the role of the dummy row and column and the treatment of the resulting allocation probabilities should be clarified.","section":"Section 5.2"}],"recommendation":"major_revision","confidential_remarks":"The sign error in Eq. (8) is best viewed as a typo in a repairable derivation, but because it sits at the load-bearing step of the central theorem, the revision must include a fully corrected proof. The empirical state-of-the-art claim should also be softened in light of Table 5. I see no ethical or citation concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What's actually new: the paper gives the first explicit Myerson-style characterization of the revenue-optimal mechanism for single-slot joint advertising, and the bundle-level IC relaxation in Lemma 5.1 is a reasonable idea that hasn't appeared in JRegNet or JAMA. The BundleNet visualizations show sensible boundary behavior, roughly tracking the theoretical allocation. The debt to Myerson is direct but the adaptation to two-bidder bundles is non-trivial, so the theoretical core is a legitimate contribution.\n\nSoft spots, in proportion.\n\nFirst, the proof as printed is not sound. I checked Appendix B: Equation (8) has plus signs before the IR constants U_r(a_r) and U_s(a_s). Substituting Equation (7) into Equation (6) gives minus signs. With the printed plus signs, the text's claim that setting these non-negative terms to zero maximizes revenue is backwards; minimizing them would maximize revenue. The intended argument is clear and the fix is trivial, but as written the proof of Theorem 4.3 does not go through. Along the same lines, the appendix never explicitly states that the slot goes to the highest-virtual-value bundle; it says this is \"implicitly indicated\" without proving the equivalence, and ties are not handled in the step-function characterization. For an \"if and only if\" theorem, that's a gap, though repairable.\n\nSecond, the experiments undercut the claims. In Table 1, U2, JRegNet revenue is 0.5622 while the \"Optimal\" baseline is 0.5247. Either the baseline is miscomputed or the experiment doesn't match the theorem's assumptions; a nearly IC mechanism cannot beat the true optimum. Table 3, N2 shows the same pattern. And the multi-slot \"state-of-the-art\" claim is contradicted by Table 5, N5x5, where JRegNet revenue (2.2071) exceeds BundleNet's (2.1393). No error bars, no code, and the experimental details are sparse. These are not fatal to the theoretical contribution, but they mean the empirical section should not be trusted until redone.\n\nWho this is for: anyone working on neural auction design in sponsored search, or on joint-advertising mechanisms specifically. The single-slot theorem is worth having, after a corrected proof. The multi-slot heuristic is plausible but currently oversold.\n\nRecommendation: send to peer review, but only with a request for major revisions. The sign error and the baseline miscalculation must be fixed, and the SOTA claim should be qualified or removed. A serious referee should check the corrected proof before acceptance.","headline":"Plausible and novel single-slot optimality result for joint ads, but the appendix proof has a sign error and the experiments overclaim state-of-the-art; worth a serious referee, not acceptance as-is.","tokens_in":16978,"tokens_out":5989,"would_cite":false,"duration_ms":66949,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91B26","91B03","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"For one ad slot, the revenue-maximizing joint-auction mechanism is a step function that awards the slot to the bundle with the highest combined virtual value and charges each winner the critical bid.","keywords":["joint advertising","auction mechanism design","revenue maximization","virtual values","critical value payments","neural auction design","incentive compatibility","individual rationality"],"falsifier":"Take a single-slot joint-ad environment whose values are sampled from a distribution with non-monotone virtual values or correlated retailer and supplier values, compute the expected revenue of the step-function critical-payment mechanism, and compare it with an unrestricted mechanism on the same draw distribution; if the unrestricted mechanism reliably earns more, or if truthful bidding is not a dominant strategy under the step-function rule, the optimality claim is false.","tokens_in":15989,"feed_emoji":"💰","tokens_out":8004,"duration_ms":83971,"temperature":0.7,"pith_summary":"This paper sets out to find the revenue-maximizing auction rule for joint advertising, in which a single ad slot is awarded to a bundle made of one retailer and one supplier who bid together. It claims that in the single-slot case the optimal deterministic mechanism is a step function: award the slot to the bundle whose combined virtual values clear the auctioneer's reserve and beat every competing bundle, and charge each winning bidder the critical bid at which that bundle would just lose. For the multi-slot case, where no characterization is claimed, it introduces BundleNet, a neural network that learns allocation and payment rules from bundle-level regret constraints, and reports that it tracks the theoretical optimum on single-slot problems and raises revenue over earlier mechanisms on multi-slot problems while keeping incentive-compatibility violations near zero. The result matters because joint advertising is a real revenue format on large platforms, and no previous mechanism for it had an optimality guarantee.","feed_headline":"Optimal joint-ad auction: highest combined virtual value wins","feed_subtitle":"A step-function rule with critical-value payments is revenue-optimal for a single joint ad slot.","key_machinery":"The carrying object is the bundle virtual value, defined as the sum of the two partners' virtual values, $c_e(v_r,v_s)=c_r(v_r)+c_s(v_s)$, where each bidder's virtual value is the standard single-parameter transform $c_i(v)=v-(1-F_i(v))/f_i(v)$. Because the slot goes to the bundle with the highest such sum above the reserve, each bidder only needs to track the neighboring partner with the highest virtual value, which reduces the two-sided joint-bid problem to a one-dimensional step-function allocation with a critical-value payment. In the multi-slot neural mechanism, the carrying device is a bundle-level regret constraint whose sum is shown to upper-bound the sum of per-bidder regret, so driving bundle regrets to zero pushes the whole mechanism toward dominant-strategy incentive compatibility.","core_discovery":"The central discovery is a necessary-and-sufficient characterization of optimal joint auctions with a single slot (Theorem 4.3). Under regular value distributions, a deterministic joint auction maximizes expected revenue if and only if its allocation is a step function that gives the slot to the bundle with the highest sum of the two partners' virtual values, provided that sum exceeds the auctioneer's reserve value $v_0$, and its payments charge each winner the critical value, defined as the lowest bid at which the winner's best bundle would still clear the reserve and beat all bundles that exclude the winner. The experiments then show that BundleNet's learned mechanism approximates this step-function allocation in the single-slot setting and delivers the reported revenue improvements in the multi-slot setting against prior learned and incentive-compatible baselines.","pith_inferences":["Our inference: if the single-slot characterization carries over to multiple slots, the optimum should be a weighted virtual-value rank using the slot click-through rates $\\lambda_k$; checking BundleNet's learned allocations against such a weighted rule is a direct test.","Our inference: the additive bundle-value assumption is the first point of failure in practice; a testable extension would re-derive the critical values under correlated retailer-supplier values, and the theorem's step-function rule should not be trusted under such correlation.","Our inference: the mechanism assumes the auctioneer knows the true distributions, so a practical deployment would need estimated distributions; measuring the revenue loss of the step-function rule under distributional misspecification would quantify how much robustness costs."],"forward_implications":["A single-slot joint-ad auction has an exactly characterized revenue-optimal rule, so platforms can test any deterministic mechanism they run against the step-function and critical-value prescription.","Winning payments have a clean operational meaning: each winning partner pays the lowest bid that keeps their bundle ahead of the auctioneer's reserve and all rival bundles.","The bundle-level regret constraint gives a reusable substitute for per-bidder incentive constraints in multi-slot joint auctions, because it dominates the per-bidder constraints by Lemma 5.1.","The reported experiments imply that encoding the bundle structure explicitly, rather than treating partners as independent bidders, is what lets a learned mechanism approach the theoretical optimum."],"supporting_citations":[{"why":"Supplies the single-parameter feasibility lemma and the revenue identity that Theorem 4.3 builds on.","marker":"(Myerson, 1981)"},{"why":"Defines the joint-advertising model and provides the Revised VCG incentive-compatible baseline used in the experiments.","marker":"(Ma et al., 2024)"},{"why":"Provides the previous neural joint-ad mechanism and the multi-slot experimental setting that BundleNet is compared against.","marker":"(Zhang et al., 2024)"},{"why":"Contributes the RegretNet-style learned-mechanism framework and the doubly stochastic allocation construction used by BundleNet's allocation network.","marker":"(D¨utting et al., 2024)"},{"why":"Studies repeated joint ads from a regret-minimization angle, delineating the single-round revenue-setting claim of this paper.","marker":"(Aggarwal et al., 2024)"},{"why":"Define the classic incentive-compatible benchmark that the joint-ad mechanisms are compared with.","marker":"(Vickrey, 1961; Clarke, 1971; Groves, 1973)"}],"fun_headline_variants":["Optimal single-slot joint ad auction uses step-function allocation","BundleNet: neural approach for near-optimal multi-slot joint ads","Critical-value payments make joint ad auctions incentive-compatible","Highest sum of virtual values is revenue-optimal for single-slot joint ads"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theorem assumes each advertiser's value is drawn independently from a known regular distribution and that a bundle's value is exactly the sum of the two partners' values, so if real joint-ad values are correlated, non-additive, or non-regular, the step-function rule need not be revenue-optimal.","fun_headline_variants_meta":{"raw":{"variants":["Optimal single-slot joint ad auction uses step-function allocation","BundleNet: neural approach for near-optimal multi-slot joint ads","Critical-value payments make joint ad auctions incentive-compatible","Highest sum of virtual values is revenue-optimal for single-slot joint ads"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000756,"raw_usage":{"total_tokens":3308,"prompt_tokens":838,"completion_tokens":2470,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":454,"completion_tokens_details":{"reasoning_tokens":2408}},"tokens_in":454,"tokens_out":2470,"duration_ms":22865,"temperature":1.0,"reasoning_tokens":2408,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:42:43.951048+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a single-slot joint-ad environment whose values are sampled from a distribution with non-monotone virtual values or correlated retailer and supplier values, compute the expected revenue of the step-function critical-payment mechanism, and compare it with an unrestricted mechanism on the same draw distribution; if the unrestricted mechanism reliably earns more, or if truthful bidding is not a dominant strategy under the step-function rule, the optimality claim is false.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the single-parameter feasibility lemma and the revenue identity that Theorem 4.3 builds on."},{"cited_title":"Joint auction in the online advertising market","cited_arxiv_id":null,"evidence_quote":"Provides the previous neural joint-ad mechanism and the multi-slot experimental setting that BundleNet is compared against."},{"cited_title":"Selling joint ads: A regret minimization perspective","cited_arxiv_id":null,"evidence_quote":"Studies repeated joint ads from a regret-minimization angle, delineating the single-round revenue-setting claim of this paper."},{"cited_title":"Counterspeculation, auctions, and competitive sealed tenders","cited_arxiv_id":null,"evidence_quote":"Define the classic incentive-compatible benchmark that the joint-ad mechanisms are compared with."}],"review_version":1}