{"id":"a54d8c01-1d92-4a42-a531-4dc5f16634b7","arxiv_id":"2411.13162","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"DFP adjusts per-click payments using real-time conversion data to keep advertisers' realized CPA equal to their target, which the authors claim preserves truthful bidding even with sparse clicks.","lead":"The paper introduces Decoupled First-Price Auction (DFP), an autobidding mechanism that adjusts per-click payments using real-time conversion outcomes so that advertisers' realized cost per acquisition stays at their target. The authors claim this preserves truthful bidding during runtime, especially for advertisers with sparse clicks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2 proves TIC only for a hindsight payment rule: p* is the optimum of bPPl_t, which uses the stage's realized conversions and clicks, whereas online per-click payments must be set before this information arrives.","rationale":"The reader's weakest_assumption is exactly the load-bearing gap I find. Section 4.2 defines bPPl_t with stage-level realized conversions and clicks; Theorem 2 concludes TIC by plugging in the optimum of that offline program. The paper itself flags the missing information in the paragraph following Theorem 2, but then pivots to an RL heuristic without stating any guarantee for the estimated case. Thus the headline claim 'DFP is TIC' is not established for the online OCPC setting. The TIC definition and the decoupling construction are interesting, and Lemma 6 correctly characterizes TIC in terms of the stage-level equality; the flaw is the step from that characterization to an implementable mechanism. The experiments compare the trained DFP against baselines on the realized tCPA/CPA ratio, but the training reward is a function of the same realized bZmNt, so the evaluation does not demonstrate causal TIC. Because the central theoretical result is conditional on unavailable information, I concur with the reader's REJECT; the paper could be resubmitted with a qualified theorem (e.g., ϵ-TIC with an explicit error bound in terms of prediction error) and a causal evaluation.","tokens_in":18373,"tokens_out":5580,"duration_ms":56772,"concrete_test":"Recompute the Section 6.1 experiment in a strictly causal setting: for each click at round n, allow the payment to depend only on information available at that click (e.g., CTR/CVR estimates, clicks and conversions observed before n, and a forecast of stage conversions trained only on past stages), then compute tCPA/CPA per stage for the 5000 bidders. If the DFP quartile range widens beyond the reported ϵ ≈ 0.1, or if any stage attains tCPA/CPA ≠ 1, the theorem's hindsight assumption is empirically falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Definition 7 defines DFP's payment as bpmnk = p*_mnk, where p* is the optimal solution to bPPl_t (Section 4.2). The objective of bPPl_t is sum_m | bZmNt · tCPA_m / (Σ_{n∈[Nt],k∈[K]} bymnk · bpmnk) − 1 |. Both the numerator bZmNt and the click indicators bymnk in the denominator are realized values for the whole stage t. In OCPC, per-click payments are committed at click time, before the stage's conversions (and often before later clicks) are observed. The paper acknowledges this: 'the platform lacks knowledge of bZmNt when determining bpmnk.' Therefore Theorem 2's proof, which substitutes p* and concludes the objective is zero, establishes TIC only for a clairvoyant mechanism that knows the stage's future outcomes. No theorem or bound covers the causal/estimated variant that is actually deployed; the RL section optimizes a reward that also uses the realized bZmNt, so the empirical tCPA/CPA measurement does not break the circularity. This is a load-bearing gap because the paper's central claim—that DFP is TIC in a runtime sparse-click system—rests entirely on this oracle payment rule.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies autobidding for tCPA bidders under the OCPC pricing model. It first proposes Coupled First-Price Auction (CFP), a first-price mechanism claimed to be Autobidding Incentive Compatible (AIC). It then introduces Time-Invariant Incentive Compatibility (TIC), a stricter notion intended to capture the behavior of risk-averse bidders who evaluate realized outcomes at the end of each stage, and proposes Decoupled First-Price Auction (DFP), in which the payment rule is defined as the solution of the stage-level optimization problem bPPl_t. A PPO-based RL algorithm with a two-part reward is proposed to implement the DFP payment rule. Experiments on 31 days of data from 5000 bidders compare tCPA/CPA ratios and payment fluctuation across CPA, Pacing, CFP, and DFP.","tokens_in":18784,"tokens_out":8984,"duration_ms":101372,"significance":"If DFP were truly incentive compatible in a causal online sense, the paper would address a real practical problem: sparse-click advertisers deviating from truthful bidding because realized CPAs deviate from the tCPA. The first-price AIC observation for CFP is clean under the paper's own definitions, and the empirical comparison over 155,000 bidder-day instances is a useful descriptive exercise. However, the central theoretical claim is not supported as stated: Theorem 2 holds only for a hindsight payment rule that uses realized stage-level conversions and clicks, and the TIC concept is built on a behavioral rule that presupposes the desired equivalence. The paper provides no formal model of risk-averse utility maximization, no approximation guarantee for the RL-based causal payment rule, and no code or machine-checked proofs. The empirical results are suggestive but do not repair the gap between the deployed mechanism and the theorem.","major_comments":[{"comment":"The proof of Theorem 2 assumes the platform knows the stage-level realized conversions bZmNt and clicks bymnk when choosing the per-click payments bpmnk. The objective of bPPl_t is sum_m | bZmNt * tCPA_m / (sum_{n,k} bymnk * bpmnk) - 1 |, and the optimal p* is defined after these whole-stage realized values are available. In OCPC, per-click payments must be committed at click time, before delayed conversions (and often before later clicks) are observed; Section 4.2 itself acknowledges that \"the platform lacks knowledge of bZmNt when determining bpmnk.\" Theorem 2 therefore proves TIC only for a clairvoyant hindsight mechanism, not for the causal payment rule that is actually deployed.","section":"§4.2, Definition 7 and bPPl_t"},{"comment":"Lemma 6 is essentially a restatement of Lemma 4: the condition bZmNt * tCPA_m = sum bymnk * bpmnk is exactly the requirement tCPA_m / CPA_mNt = 1. It is not specific to CFP, and the bPPl_t objective is deliberately designed to make this equality hold. Substituting p* into the objective and concluding that the objective value is zero is true by construction, so Theorem 2 has no independent game-theoretic content as a statement about bidder incentives.","section":"§4.2, Lemma 6 and Theorem 2"},{"comment":"The RL implementation does not close the oracle gap. The reward r^(1)_n is defined using the whole-stage realized quantities bZmNt and bymnk, and the state description states that bZmn is updated only once at the end of each stage. The empirical tCPA/CPA ratios in Table 1 and Figures 3-5 therefore mainly show how well the reward function is optimized, not that truthful bidding is an optimal strategy in a mechanism whose payments are set causally at click time. No theorem or bound is given for the error introduced by estimating bZmNt, so the main claim is not established for the deployed variant.","section":"§5.1, reward r^(1)_n and state definition"},{"comment":"The paper models risk aversion as a behavioral rule rather than as a preference. Definition 5 says a bidder is risk-averse if she adjusts her bids or withdraws when she observes deviations, and Lemma 3 then asserts that such a bidder bids truthfully iff tCPA_m / CPA_mNt = 1 at every stage. This is not derived from any utility maximization problem; it is an assumption that the desired equivalence holds. Consequently, TIC (Definition 6) is a post-hoc consistency condition on realized CPA rather than an incentive property established for risk-averse utility maximizers.","section":"Definition 5 and Lemma 3"},{"comment":"The derivation of the click-volume threshold in Lemma 5 controls the two Chernoff tails separately and then takes the larger threshold, but for the conjunction of both inequalities to hold with probability at least 1 - epsilon, a union bound would require each tail to be controlled at about epsilon/2 rather than epsilon. Moreover, Chernoff gives only an upper bound on tail probabilities, so it cannot establish the stated \"necessary\" condition on bYmNt; at best the derivation gives a sufficient condition under the specific concentration bound used. The lemma as stated is therefore not rigorous.","section":"Appendix C.6, proof of Lemma 5"}],"minor_comments":[{"comment":"There are multiple typos and rendering problems: \"desgin\" in Section 1, \"vloume\" in Section 3, \"MES\" for MSE in Section 5.2, \"clcik\" in Figure 2, and \"an bidder\" in Section 6.2.2 and Appendix A. The accents used for expected and realized values, e.g. [CPAmNt and bPmn, are also not rendered consistently, which makes the distinction between ex ante and realized objects hard to follow.","section":"Throughout"},{"comment":"The text says that if bZmNt were known, bPPl_t would reduce to a standard online decision-making problem, but the objective is evaluated on whole-stage realized clicks and conversions; this is a hindsight problem, not an online one. The wording should be corrected to avoid implying causal information is available.","section":"§4.2, bPPl_t"},{"comment":"The payment fluctuation comparison mixes different information regimes: CPA and Pacing are offline baselines with complete knowledge of all outcomes, while DFP is the deployed system. The descriptive quartile comparison is useful, but the text should acknowledge that this is not a controlled online experiment and should report the number of clicks or impressions underlying the variance statistics.","section":"§6.2.4 and Table 1"},{"comment":"The PPO, GAE, and entropy-regularization components are described without citations; standard references for proximal policy optimization and generalized advantage estimation should be added, since the paper claims to use these specific methods.","section":"§5.2"},{"comment":"The case of a stage with zero clicks or zero conversions is not handled: the ratio in bPPl_t is undefined when the denominator is zero, and Lemma 4 similarly assumes CPA_mNt is well defined. Given that the paper emphasizes sparse-click systems, this boundary case should be addressed explicitly.","section":"§4.2 and Definition 7"}],"recommendation":"reject","confidential_remarks":"The manuscript's central theoretical contribution is not supported in the form stated: Theorem 2 applies to a clairvoyant payment rule, and the RL section does not provide a bound or guarantee for the causal estimated variant. The TIC definition itself rests on a behavioral assumption rather than a derived optimal strategy, so the game-theoretic claim is largely definitional. The empirical work is substantial but cannot replace the missing causal guarantee. I would recommend rejection unless the authors substantially reframe the contribution as an empirical stabilization technique and prove approximation guarantees for a causal payment estimator under explicit assumptions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper introduces a genuinely new notion, Time-Invariant Incentive Compatibility (TIC), and a mechanism idea worth taking seriously: decouple allocation from payment and use real-time conversion feedback to adjust per-click prices. That is a real contribution to the autobidding mechanism-design literature. Second, the main theorem that DFP is TIC does not hold up. The payment p* is defined as the minimizer of an objective that, when set to zero, is exactly the TIC condition. So Theorem 2 is true by construction, but only for a payment rule that knows the stage's realized conversions and clicks. The paper itself acknowledges that the platform lacks knowledge of bZmNt when setting per-click payments. No theorem or bound covers the estimated/online variant that is actually deployed, and the RL section optimizes a reward using the same realized bZmNt, so the empirical tCPA/CPA ratio is essentially the training reward, not an independent test. The central claim—that DFP is TIC in a runtime sparse-click system—is therefore not established.\n\nWhat the paper does well: the CFP analysis is clean. Under a first-price payment rule, expected CPA equals the bid, so truthful bidding is optimal and CFP is AIC and IR. The Chernoff-style bound in Lemma 5 is a plausible statement about when an AIC mechanism becomes approximately TIC, even if the derivation is a bit informal about conditioning on realized clicks. The experiments use real data from a large ad platform, which gives credibility to the motivating observation that sparse-click bidders see realized conversions deviate from expectations and react by changing bids.\n\nThe soft spots are proportionate to the main flaw. The definition of risk-averse bidders is behavioral rather than utility-based, so TIC is really a design goal rather than a derived incentive property; that is fine, but it limits the theoretical claim. The proof of Lemma 5 also skips some details, though that is minor compared to the hindsight issue. A more serious editing problem is the abstract and conclusion overclaim: they state DFP retains IC during runtime without qualifying that the theorem only holds for a clairvoyant payment rule.\n\nThe paper is for mechanism design researchers and industrial ad-auction practitioners. The TIC concept and the failure mode it addresses are worth knowing, and the CFP result is a solid building block. To make the paper acceptable, the authors would need to either restrict the theorem to an offline benchmark or provide a guarantee for the estimated payment rule; without that, the main result is not as advertised.\n\nA serious editor should send this to peer review rather than desk-reject. The concept is novel, the motivating problem is real, and the CFP part is sound. The referee's attention should focus on the gap between the hindsight payment rule and the actual online implementation, and on whether the RL reward can be considered a legitimate estimate. If that gap cannot be closed, the paper should be rejected or reframed as a system design proposal without the TIC theorem.","headline":"TIC is a genuinely new concept and the CFP result is clean, but the main theorem that DFP is TIC is circular and only holds for a clairvoyant payment rule; the online variant is unsupported.","tokens_in":19247,"tokens_out":2880,"would_cite":true,"duration_ms":31340,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-12T16:46:34.637090+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}