{"id":"4f00a371-efb9-4cd8-a4b0-0cac50c20f79","arxiv_id":"2507.07711","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"HRegNet learns revenue-maximizing hybrid auctions that mix independent-store ads and store-brand bundle ads, reporting higher revenue than existing mechanisms in experiments.","lead":"The paper proposes a hybrid advertising auction for e-commerce, where each ad slot can show either a single store or a store-brand bundle, and trains a neural network, HRegNet, to set allocations and payments. The authors report that HRegNet raises platform revenue versus baseline auctions on synthetic and real-world data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Near-DSIC is asserted on the basis of finite-restart gradient-ascent regret estimates that can miss the optimal misreport; the revenue claim is not secure without an exact or exhaustive check.","rationale":"The paper's central contribution is an empirical one: HRegNet generates hybrid auction mechanisms that significantly improve revenue while maintaining near-DSIC and IR. The revenue numbers in Tables 1–3 are computed under truthful value reports, so the comparison to baselines is only meaningful if the near-DSIC claim holds to the reported tolerance. The paper's evidence for near-DSIC is an average empirical regret estimate obtained by gradient ascent over misreports, which is a reasonable heuristic but not a certificate. Because the misreport utility is a nonconcave function of a neural network's outputs, finite-restart gradient ascent can miss the true maximizer, and the reported regret bound could be optimistic. This is the single most load-bearing concern: if regret is underestimated, agents would misreport, allocations and payments would change, and the reported revenue gain could disappear. I do not consider the Reader's IR concern to be the decisive issue. Although Section 3.1 contains a false justification for IR, the mechanism itself satisfies IR because p_i = ptilde_i * v_i * g_i with ptilde_i in [0,1], so truthful utility is nonnegative even when g_i > 1. The paper's proof just needs a corrected argument. The DSIC concern, by contrast, cannot be repaired by a one-line fix; it requires either a more exhaustive search over misreports or a formal guarantee. The proposed dense-grid test on small instances would settle whether the concern lands: if the grid-based regret is also below 0.001, the near-DSIC claim is credible; if not, the revenue claim is unsupported. Since the Reader already issued a conditional verdict and my analysis reinforces the main condition (with a different emphasis on IR), I recommend no change to the verdict: the paper should be accepted only after the DSIC validation is strengthened or the underlying mechanisms are independently verified.","tokens_in":15259,"tokens_out":14490,"duration_ms":182189,"concrete_test":"Fix a trained HRegNet for Setting A (2 stores, 2 brands, 1 slot) and Setting B (3 stores, 4 brands, 3 slots). On 100 held-out value profiles, replace the 100-start gradient-ascent regret estimator with an exhaustive dense grid over each advertiser's one-dimensional misreport domain, e.g., 10,000 equally spaced points in [0,1], plus 1,000 random restarts of the paper's gradient ascent. Report the maximum per-profile regret and the mean regret. If the maximum regret remains below 0.001, the near-DSIC claim is validated; if any profile exceeds 0.001, the claim is falsified and the revenue tables must be recomputed under equilibrium misreports. Repeat the grid check on a subsample of real-data samples using observed bid ranges and the ru metric.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that HRegNet's mechanisms are near-DSIC, because the reported revenue is computed under truthful reports. The only support is the empirical regret in Tables 1–3, obtained by initializing 100 misreports and running 200 gradient-ascent steps (Section 4.1). For a neural allocation and payment rule, misreport utility is nonconcave and can have multiple isolated local maxima, so a finite number of gradient-ascent restarts can underestimate the maximum regret even when the misreport is a single scalar per advertiser. Since the same regret quantity is used as the training penalty (Equations (2)–(4), Algorithm 1), the test-time regret is not an independent certificate of DSIC. If a better misreport exists for some value profile, the mechanism is manipulable and the reported 'truthful' revenue is not an achievable outcome, undermining the headline revenue improvement. The IR concern raised by the Reader is a proof gap rather than a violation: Section 3.1's claim that payments 'cannot exceed the values (i.e., bids)' is literally false when alpha_i > 1 and a store wins both an independent and a bundle slot, since the expected allocated value can exceed 1; however, p_i = ptilde_i * v_i * g_i with ptilde_i in [0,1] gives truthful utility (1 - ptilde_i) * v_i * g_i >= 0, so IR holds by construction. The load-bearing uncertainty is therefore DSIC, not IR.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new sponsored-search advertising model, \"hybrid advertising,\" in which each ad slot can be allocated either to an independent store or to a store-brand bundle, with at most C winning bundles. To compute a revenue-maximizing mechanism in this model, the authors introduce HRegNet, a neural-network architecture whose allocation network enforces the per-slot, per-bidder, and bundle-cap constraints, and whose payment network scales expected allocated values by sigmoid parameters to enforce individual rationality. The training objective is an augmented Lagrangian of negative revenue plus per-advertiser empirical regret penalties (Section 3.2). Experiments on synthetic settings (Tables 1–2) and on real e-commerce log data (Table 3) report that HRegNet achieves significantly higher revenue than RegretNet, JRegNet, VCG, and a revised JRegNet baseline, while reporting regret below 0.001 in synthetic tests and a regret/utility ratio below 0.065 in real-data tests. The paper's central claims are that HRegNet generates near-DSIC and IR mechanisms with superior revenue.","tokens_in":15567,"tokens_out":8038,"duration_ms":101101,"significance":"If the empirical claims withstand scrutiny, the paper makes a useful applied contribution: it defines a practical auction format that is already close to industry practice, encodes nontrivial allocation constraints in a differentiable way, and provides real-data evidence that the hybrid format can outperform both store-only and bundle-only designs. The inclusion of multiple baselines, especially the revised JRegNet, is a strength, and the real-world evaluation on e-commerce log data adds credibility. However, the paper does not provide code or an independent verification protocol, and its theoretical guarantees are only as strong as the regret estimates used during training and testing. The claimed incentive guarantees therefore need additional support before the revenue-improvement headline can be fully relied upon.","major_comments":[{"comment":"The only evidence for the central \"near-DSIC\" claim is the empirical regret computed with the same procedure that is minimized during training: for each test profile, 100 misreports are initialized and 200 gradient-ascent steps are run, and the maximum over those restarts is reported. For the nonconcave misreport-utility functions induced by a neural allocation and payment rule, a finite number of gradient-ascent restarts yields a lower bound on the true worst-case regret, not an upper bound. A reported regret below 0.001 therefore does not certify near-DSIC, and the revenue numbers in Tables 1–3 are computed under truthful reports that may not be incentive-compatible. I request an independent verification at least in the smallest setting (Setting A), such as an exhaustive grid or exact optimization over the single-dimensional misreport space, or a provable upper bound derived from the network architecture; the same metric used as the training penalty is not an independent certificate.","section":"Section 3.2 / Section 4.1, Eqs. (2)–(4), Algorithm 1, Tables 1–3"},{"comment":"The statement that the matrix S3 = min(column-softmax(S1), row-softmax(S2)) is \"doubly stochastic\" is not correct for the rectangular matrices used here. The column-wise softmax gives column sums of 1, and the row-wise softmax gives row sums of 1 including the non-allocation column; after taking the elementwise minimum and deleting that column, the row and column sums are at most 1 but are not equal to 1 in general. For example, when the number of candidate rows R+m differs from K, equality of both total sums is impossible. The feasibility constraints (a) and (b) only require inequalities, so the allocation construction can still be repaired, but the lemma as stated and its citation to [8] do not establish the claimed property. This needs a corrected statement and proof for a doubly sub-stochastic matrix.","section":"Section 3.1, Lemma 1 and the construction of S3"},{"comment":"The sentence \"Since we focus on DSIC mechanisms, and \\tilde{p}_{i\\cdot} \\in [0,1] and \\tilde{p}_{\\cdot j} \\in [0,1], the payments cannot exceed the values (i.e., bids), which implies that IR condition is satisfied\" is false as written. Because the independent-store allocation term includes the quality factor \\alpha_i, and because a store can win both an independent slot and bundle slots (Footnote 3), \\sum_k a_{i\\cdot k} q_{i\\cdot k} can exceed b_i; payments can then exceed the bid. The correct IR argument is that, under truthful reporting, payment equals \\tilde{p}_{i\\cdot} v_i g_i, so truthful utility is (1-\\tilde{p}_{i\\cdot}) v_i g_i \\ge 0. The proof should be corrected, and the model should clarify whether a store may appear in multiple winning bundles simultaneously, since this affects both the IR argument and the interpretation of the allocation constraints.","section":"Section 3.1, payment network and IR discussion"}],"minor_comments":[{"comment":"The displayed definition of rgti·(v) includes an expectation over v ∼ F while also depending on a fixed value profile v; this appears to conflate ex-post regret with expected regret. The empirical version in Eq. (2) is clearer, and the definition should be rewritten to match it.","section":"Section 2.2, definition of ex-post regret"},{"comment":"The notation for bids and values is inconsistent: the mechanism is defined with bids b, but payment rules are written as p : V → R≥0, and the utility formulas mix b and v. Aligning the notation would remove avoidable confusion about which arguments are truthful reports.","section":"Section 2.1 and Section 3.1, notation"},{"comment":"The DSIC metric ru is reported only as a uniform bound \"<0.065\" for every method and every C value, which makes the column uninformative for comparing methods. The actual ratios, or at least their standard deviations, should be reported.","section":"Table 3"},{"comment":"The test-time regret evaluation uses 100 restarts and 200 gradient steps, but no details are given about the step size, the initialization distribution, or the stopping rule for the gradient-ascent misreport search; these details are necessary for other researchers to reproduce the reported regret values.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for an applied mechanism-design venue, and the empirical setup is more realistic than many AMD papers. The main risk is the DSIC certificate: if the test-time regret is only a lower bound, the headline revenue gain may not be achievable. The authors should be asked to supply an exact or exhaustive verification in at least one small setting and to correct the false lemma and IR proof. If the paper is intended for a more theory-oriented journal, the incorrect Lemma 1 is a serious issue; if the venue is applied, it is still a load-bearing proof error that must be fixed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, useful extension of neural auction design to a hybrid store/bundle format. HRegNet reports consistent revenue gains over strong baselines across synthetic settings, value distributions, and three real-data splits. The main soft spot is the DSIC validation, which is circular and not a certified bound; the IR proof is misstated but the result actually holds.\n\nWhat's new: the hybrid format itself—each slot allocated to either an independent store or a store-brand bundle, with a cap C on the number of bundles—is not covered by RegretNet, JRegNet, or the theoretical joint-ad work. The two-subnetwork allocation with a C-softmax constraint is a genuine architectural change, and the use of quality factors α_i for independent display is reasonable. The experiments are broad and the comparisons are honest, including a revised JRegNet baseline. The statistical significance testing is standard.\n\nThe soft spots, in proportion. First, the IR proof in Section 3.1 is wrong as stated: when α_i > 1 and a store wins both an independent slot and a bundle slot, the allocated expected CTR can exceed 1, so a payment p_i = ptilde_i * v_i * g_i can exceed the bid. But the truthful utility is (1 - ptilde_i) * v_i * g_i, which is nonnegative because ptilde_i is in [0,1]. So IR holds by construction; only the proof needs fixing. That's minor.\n\nThe more serious issue is DSIC. Test-time regret is computed by 100 random restarts and 200 gradient-ascent steps, using the same objective minimized during training. For a nonconcave neural mechanism, this yields a lower bound on true regret, not an upper bound. The reported \"<0.001\" is the regret found, not a certified guarantee. Since revenue is computed under truthful bidding, a better misreport would make that revenue unachievable. This is a standard limitation in the RegretNet literature, but the paper should state it as a limitation rather than presenting near-DSIC as established. The real-data ratio ru < 0.065 is better than nothing but still the same gradient-ascent search.\n\nNo code or data is released, which hurts reproducibility but is common in this area. The citation pattern is fine: RegretNet, JRegNet, and the joint-ad theory are all present and used correctly.\n\nWho benefits: researchers and practitioners in automated mechanism design, especially those working on sponsored search and ad auction formats. It deserves a serious referee. My recommendation: engage with it, but request strengthened DSIC validation for small cases (e.g., exhaustive grid or an upper-bound certificate) and a corrected IR proof. With those changes, it's a citable contribution.","headline":"A practical, well-executed extension of neural auction design to hybrid store/bundle formats, but the near-DSIC claim rests on a circular and non-exhaustive regret check, so treat revenue as conditional on near-truthful bidding.","tokens_in":16107,"tokens_out":2594,"would_cite":true,"duration_ms":29835,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91B26","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Letting an ad slot show either an independent store or a store-brand bundle, with the bundle count capped at C, raises platform revenue beyond either pure format.","keywords":["hybrid advertising","sponsored search","auction mechanism design","automated mechanism design","neural network architecture","regret minimization","incentive compatibility","individual rationality"],"falsifier":"Run a one-slot hybrid auction with one store having $\\alpha_i = 1.5$, one bundle containing it, and network probabilities of 0.7 for the independent store and 0.4 for the bundle; the store's expected allocated value is $1.45$, so a sigmoid payment scalar near 1 produces a payment above the store's value, directly contradicting the paper's IR argument if such a trained run exists.","tokens_in":15037,"feed_emoji":"🛒","tokens_out":7132,"duration_ms":72386,"temperature":0.7,"pith_summary":"This paper tries to establish that online ad platforms can raise revenue by mixing two ad formats in one auction: each slot can go to an independent store or to a store-brand bundle, with the number of bundles capped at C. The claim is that a learned mechanism, HRegNet, can allocate slots and set payments for both formats while keeping each advertiser's incentive to misreport negligible and keeping truthful participation non-negative. On synthetic distributions and on three days of real e-commerce auction logs, the paper reports that this hybrid mechanism beats store-only, bundle-only, and welfare-maximizing baselines in revenue. If right, the contribution is a practical template for ad systems that want the reach of store ads and the two-sided revenue of bundle ads in a single ranked list.","feed_headline":"Hybrid ads beat store-only and bundle-only auctions","feed_subtitle":"A learned mechanism mixes both formats in one slot list, keeping regret under 0.001 and lifting platform revenue.","key_machinery":"The load-bearing object is HRegNet's allocation-and-payment construction. Two subnetworks score independent stores and bundles separately; row-wise and column-wise softmax operations on the raw scores, followed by an elementwise minimum, produce a doubly stochastic allocation matrix that respects the one-slot-per-advertiser and one-advertiser-per-slot constraints. A C-softmax normalization caps the total probability mass assigned to bundles at C. Payments are computed by scaling each advertiser's expected allocated value by a sigmoid parameter in [0,1], which the paper argues keeps payments below bids and hence enforces individual rationality. Training minimizes negative expected revenue plus an augmented Lagrangian penalty on empirical ex-post regret, with optimal misreports found by gradient ascent.","core_discovery":"The central discovery claimed is that hybrid allocation is not a compromise but a revenue improvement: allowing each slot to be either an independent store or a bundle lets the platform collect from brands as well as stores while still serving users who prefer store-centric results. The paper formalizes this as a constrained mechanism-design problem and solves it with a neural network that takes store and brand value samples as input and outputs an allocation probability matrix and payments. The resulting mechanisms are reported to have average empirical ex-post regret below 0.001 on synthetic tests and a relative regret ratio below 0.065 on real data, with empirical revenue higher than every tested baseline across the reported settings and value distributions. The authors present this as a demonstrated property of the trained mechanism, not as a closed-form theorem.","pith_inferences":["The paper leaves implicit that the two-subnetwork pattern is a general recipe for mixed-format auctions: any two display classes with different click-through rates could be handled the same way, not just stores and bundles.","The claimed individual-rationality guarantee is only as solid as the premise that expected allocated value stays at or below 1; a store with quality factor above 1 winning both an independent slot and a bundle slot violates that premise, so high-alpha stress tests are the natural next check.","A concrete extension would freeze a trained payment network and sweep alpha values above 1 to see whether any sigmoid scalar saturates near 1 while the expected allocated value exceeds 1; that is the point where an IR violation would show up empirically."],"forward_implications":["Platforms can run one auction for both ad formats rather than choosing between store-centric and bundle-centric displays; the paper's experiments indicate the mixed format alone outperforms both pure formats in revenue.","The cap C acts as a platform dial: reported revenue rises as more bundles are allowed, so a platform can trade the ad-format mix against user experience by choosing C.","The learned mechanism is fast enough for online use: the paper reports roughly 1 to 3 milliseconds per auction sample at inference after 2 to 4 hours of offline training.","The revenue gain over a welfare-maximizing baseline comes with lower social welfare in the reported numbers, so a platform that weights advertiser welfare faces a real tradeoff.","The design pattern extends, per the paper, to slots displaying independent brands, exclusive displays, or bundles with more than two components with only minor modifications."],"supporting_citations":[{"why":"Supplies the deep-learning auction framework and the doubly stochastic construction, including Lemma 1, that HRegNet adapts to hybrid allocation.","marker":"[8]"},{"why":"Introduces the joint-advertising auction baseline that HRegNet extends and compares against in the experiments.","marker":"[28]"},{"why":"The classic VCG mechanism used as a welfare-maximizing DSIC and IR baseline.","marker":"[27]"},{"why":"Part of the VCG baseline definition and its incentive properties.","marker":"[4]"},{"why":"Completes the VCG baseline definition used in the revenue comparisons.","marker":"[14]"}],"fun_headline_variants":["Hybrid ad auctions beat single-format strategies","Mixing store and bundle ads lifts platform revenue","Learned hybrid auction outearns store-only and bundle-only","One slot serves store or bundle, boosting auction revenue","Hybrid advertising: neural auction combines formats for higher profit"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The individual-rationality guarantee rests on the assumption that each participant's expected allocated value, scaled by a sigmoid parameter in [0,1], can never exceed their bid; this can fail when a store with quality factor above 1 wins both an independent slot and a bundle slot in the same auction.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid ad auctions beat single-format strategies","Mixing store and bundle ads lifts platform revenue","Learned hybrid auction outearns store-only and bundle-only","One slot serves store or bundle, boosting auction revenue","Hybrid advertising: neural auction combines formats for higher profit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1349,"prompt_tokens":867,"completion_tokens":482,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":405}},"tokens_in":483,"tokens_out":482,"duration_ms":5194,"temperature":1.0,"reasoning_tokens":405,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:34:48.062305+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a one-slot hybrid auction with one store having $\\alpha_i = 1.5$, one bundle containing it, and network probabilities of 0.7 for the independent store and 0.4 for the bundle; the store's expected allocated value is $1.45$, so a sigmoid payment scalar near 1 produces a payment above the store's value, directly contradicting the paper's IR argument if such a trained run exists.","supporting_citations":[{"cited_title":"Optimal auctions through deep learning","cited_arxiv_id":null,"evidence_quote":"Supplies the deep-learning auction framework and the doubly stochastic construction, including Lemma 1, that HRegNet adapts to hybrid allocation."},{"cited_title":"Joint auction in the online advertising market","cited_arxiv_id":null,"evidence_quote":"Introduces the joint-advertising auction baseline that HRegNet extends and compares against in the experiments."},{"cited_title":"Multipart pricing of public goods","cited_arxiv_id":null,"evidence_quote":"Part of the VCG baseline definition and its incentive properties."}],"review_version":1}