{"id":"e975ae39-50e1-479b-a833-ec80d75f37fb","arxiv_id":"2505.22243","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"UDuo replaces the static arrival assumption in online matching with forecasted dual values and reports modest A/B improvements, but its optimality and feasibility guarantees are stated without proof.","lead":"This paper proposes UDuo, a framework for online coupon or ad allocation that forecasts future user demand instead of assuming a fixed statistical arrival model. It reports real-world A/B gains on a food delivery platform, but the promised theoretical guarantees of optimality and feasibility are not actually proven in the text.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central feasibility/optimality claim is asserted without proof, and the grid-search construction in Eq. 6 can provably miss the true dual optimum under distribution drift.","rationale":"The reader's weakest_assumption already identifies the forecast-accuracy and grid-coverage issue in Eq. 6. My stress-test agrees and sharpens it: the grid Λ is centered at the stochastic-model solution λ*' with no stated bound on distribution drift, so the true optimum can fall outside the grid even with perfect forecasting. This is a concrete, checkable failure mode, not merely a stylistic concern about missing proofs. The empirical results cannot compensate because the central theoretical claim is the paper's main advertised contribution, and the experiments are thin (no error bars, no ablations, broken table formatting). The verdict of REJECT stands, whether framed as UNCHANGED or as an independent rejection based on the unsupported feasibility/optimality guarantee. I would keep the reader's verdict unchanged.","tokens_in":11008,"tokens_out":1939,"duration_ms":25472,"concrete_test":"Construct a two-period deterministic example in the paper's notation. Let T=2, |C|=2, |T|=2, with reward/cost matrices such that the stochastic-model optimum is λ*'=1, and the true dynamic optimum in period 2 is λ*_2=1+Kε (i.e., a shift of Kε). With ε and K as in Eq. 6 and Λ centered at λ*', λ*_2 lies outside Λ, so binary search returns a boundary point and is suboptimal. Compute the resulting objective gap L(λ_boundary) − L(λ*_2). If the gap is positive, the claim that UDuo 'achieves asymptotically optimal solutions' in dynamic environments fails for this case; the paper would need an adaptive grid or a drift bound to restore the claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—'asymptotically optimal solutions with constraint feasibility guarantees in dynamic environments'—is never supported by any theorem, lemma, or proof. Section 3.1 says UDuo maintains 'rigorous theoretical guarantees on budget feasibility and regret bounds,' and Section 5 says 'By theoretically proving its constraints on feasibility guarantees,' but no such proof appears. This is not merely a missing formalization; there is an identifiable technical gap in the construction itself. In Eq. 6, the search grid is Λ = {λ_low + kε}, with λ_low = λ*' − (K/2)ε, where λ*' is solved from the stochastic arrival model. The true optimal λ* under a shifted distribution can lie outside this interval whenever the shift exceeds Kε/2 relative to the stochastic-model optimum. The paper provides no mechanism to expand Λ, no bound on how far λ* can drift, and no error bound connecting forecast error in v_t(λ) to suboptimality of the selected λ. Furthermore, the budget-feasibility claim is unsupported: the pacing module allocates per-period budgets B_t with Σ B_t ≤ B, but the online decisions are made using forecast-dependent λ_t; there is no argument that actual cumulative spend Σ c_ij x_ij ≤ B holds when forecasts are wrong. Thus both optimality and feasibility rest on unstated assumptions about forecast accuracy and drift magnitude, making the headline theoretical claim load-bearing and currently unsubstantiated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UDuo, a dual-optimization framework for online matching under budget constraints. It replaces the stationary stochastic arrival model with time-series forecasts of a user-arrival representation vector v(λ) defined in Eq. (4), selects the dual variable by a grid search in Eq. (6), and introduces budget pacing strategies plus a multi-scene transformer (MiFormer) for forecasting. The authors report online A/B tests on the Eleme food-delivery platform and forecasting benchmarks, and the abstract claims that UDuo 'achieves asymptotically optimal solutions with constraint feasibility guarantees in dynamic environments.'","tokens_in":11340,"tokens_out":3170,"duration_ms":34939,"significance":"The v(λ) reformulation in Eq. (4) is a clean and useful way to separate user arrival dynamics from resource scarcity, and the paper points to a practically relevant problem: adapting online resource allocation to nonstationary arrival patterns. The deployment experience and the MiFormer forecasting architecture are also potentially valuable engineering contributions. However, the paper's headline claims of asymptotic optimality, feasibility guarantees, and 'rigorous theoretical validity' are not supported by any theorem, lemma, or proof, and the reported experimental evidence lacks the statistical detail needed to verify them. As presented, the significance of the framework is therefore not established.","major_comments":[{"comment":"The search grid Λ = {λ_low + kϵ} is centered at λ*′, the solution of the stochastic arrival model, with half-width Kϵ/2. The paper claims asymptotic optimality in dynamic environments, but it provides no bound on how far the true optimal dual variable λ* can drift from λ*′, and no mechanism to expand Λ when drift is large. Under any distribution shift for which λ* lies outside [λ*′ − Kϵ/2, λ*′ + Kϵ/2], the argmin in Eq. (6) cannot equal λ*, so the claimed optimality fails. A regret bound or drift bound that accounts for this truncation and discretization is required.","section":"§3.3, Eq. (6)"},{"comment":"The manuscript repeatedly asserts 'rigorous theoretical guarantees on budget feasibility and regret bounds' (§3.1) and states that UDuo's feasibility guarantees are 'theoretically proving' (§5), yet no theorem, lemma, or proof appears anywhere in the paper. The only formal statement is the O(log|C| 1/ϵ) complexity of binary search, which concerns computational cost, not optimality or regret. Moreover, the pacing module only guarantees Σ_t B_t ≤ B; it does not imply that actual cumulative spend Σ_t Σ_ij c_ij x_ij ≤ B when decisions are based on forecast-dependent λ_t. The conditions under which forecast errors preserve budget feasibility are never stated.","section":"§3.1, §3.4, §5"},{"comment":"The text states that Table 1 shows 'statistically significant improvements in all key metrics,' but the table reports only percentage lifts with no sample sizes, confidence intervals, or significance tests. Table 2 reports MSE and MAE averaged over horizons without error bars or significance tests, despite the text claiming a robust 8% improvement. Since these tables are the sole empirical support for the headline claims of efficiency and generalization, the experimental claims are unsubstantiated as reported.","section":"Table 1 and Table 2"},{"comment":"The paper's own checklist leaves the answers to 'Claims', 'Limitations', 'Theory assumptions and proofs', 'Experiment statistical significance', and related items as [TODO]. This is not a stylistic issue: it confirms the absence of the required theoretical and statistical support for the central claims, and no limitations section is provided elsewhere in the manuscript.","section":"NeurIPS Paper Checklist"}],"minor_comments":[{"comment":"There is a capitalization inconsistency ('We' after 'In this work,') and the paper's name is used informally; please copyedit throughout.","section":"Abstract"},{"comment":"The caption contains a typo: 'Elme's' should be 'Eleme's'.","section":"Table 1 caption"},{"comment":"The notation |C| is used both for the cardinality of the user set and as a set index, which is confusing; please use a separate index set notation such as i = 1, ..., m.","section":"Eq. (4)"},{"comment":"The phrase 'replacing the neighboring optimal solution approximation' is undefined and should be clarified.","section":"§3.1"},{"comment":"The gating scaling factor γ is set to 1.5 without justification or ablation; please provide evidence for this choice or state that it is a tuning hyperparameter.","section":"Eq. (7)"},{"comment":"The caption ends with the incomplete sentence 'Our full results are in.'; this appears to be a placeholder and should be completed or removed.","section":"Table 2 caption"}],"recommendation":"reject","confidential_remarks":"This manuscript reads as an early preprint rather than a polished submission. The core theoretical claims are asserted without proof, and the only algorithm-specific statement that could pass as theoretical is a computation-complexity remark. The grid-search construction in Eq. (6) has a concrete failure mode under distribution drift that cannot be repaired by minor editing. Given that the abstract and conclusion stake the paper's contribution on these guarantees, I do not see how a revision within the normal scope of a journal submission could address the gap. The practical system and the MiFormer results might be suitable for an applied venue if the statistical evidence is added and the claims are scaled back."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"UDuo is a paper with a genuine kernel buried under unsupported claims and an unfinished draft. The v(λ) reformulation—writing the dual objective's user-dependent term as a function of λ and forecasting that function—is a legitimate and potentially useful way to handle nonstationary arrivals. The multi-scene gating mechanism for iTransformer is a plausible engineering addition, and the A/B setup on Eleme is a real deployment. If the reported lifts hold up, that is a meaningful result.\n\nThe problem is the gap between what the abstract promises and what the paper delivers. The central claims—asymptotic optimality, constraint feasibility guarantees, rigorous theoretical validity—are asserted in the abstract, Section 3.1, and the conclusion, but no theorem, lemma, or proof appears anywhere. That is not a minor stylistic issue; those claims are load-bearing. The grid search in Eq. 6 is centered on the stochastic-model optimum λ*'. If the actual distribution drifts far enough, the true optimal λ can fall outside the grid, and the paper gives no bound on the drift or the resulting suboptimality. The feasibility guarantee is equally unsupported: pacing ensures Σ B_t ≤ B, but actual spend depends on forecast-dependent λ_t, and there is no argument that cumulative cost stays under budget when forecasts err.\n\nThe empirical sections are too thin to rescue this. The A/B table reports percentage lifts with no sample sizes, confidence intervals, or significance tests—despite the text calling the results 'statistically significant.' The forecasting benchmark table is broken (missing separators, truncated header) and has no error bars. The paper is unfinished: the NeurIPS checklist is entirely [TODO], and there are incomplete sentences.\n\nTo be fair, the v(λ) idea is a reasonable step beyond the stationary-arrival assumption, and the sliding-window recalibration is a sensible closed-loop design. The paper's combination of forecasting the dual representation with a gated transformer variant does not appear in the cited literature. But the draft as submitted is not ready for peer review. It needs either real proofs or explicitly weakened claims, full experimental details, and a completed manuscript.\n\nI would not cite this in its current form, but I would keep an eye out for a revised version. If the authors deliver the missing theory and statistics, the core idea deserves a serious referee. As is, this should be desk-rejected.","headline":"A promising reformulation of the online matching dual, but the paper asserts its main theoretical claims without proof and is too unfinished to review.","tokens_in":11833,"tokens_out":2994,"would_cite":false,"duration_ms":30010,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"UDuo replaces the stationary stochastic arrival assumption with forecasted user-arrival representation vectors and claims asymptotically optimal allocation with budget-feasibility guarantees in dynamic environments.","keywords":["online resource allocation","budget constraints","dual optimization","user arrival distribution drift","time-series forecasting","budget pacing","multi-scenario transformer","dynamic pricing"],"falsifier":"Run UDuo on a synthetic user-arrival process with a known offline optimum, apply gradual and abrupt distribution drift, and compare the online allocation and budget spend against the offline LP optimum; if forecast error on $v_{t+H}(\\lambda)$ is not tightly coupled to suboptimality or budget violation, the claimed guarantees fail.","tokens_in":10790,"feed_emoji":"🎯","tokens_out":4731,"duration_ms":49724,"temperature":0.7,"pith_summary":"The paper proposes UDuo, a framework for online resource allocation under budget constraints that abandons the classical stationary stochastic arrival assumption. Instead of estimating a fixed user distribution, UDuo builds a time-indexed representation of user arrivals, forecasts its future values with a time-series model, and binary-searches a dual variable to make allocation decisions. The paper argues this achieves asymptotically optimal solutions with budget-feasibility guarantees in dynamic environments, and reports online experiments in food-delivery pricing where it improves orders, GMV, and profit over a stochastic-model baseline. The practical point is that online matching can be driven by short-horizon forecasts of a compressed arrival summary rather than by a long-run stationary distribution.","feed_headline":"Forecast-driven duals beat stationarity in online matching","feed_subtitle":"UDuo predicts user-arrival summaries and paces budgets to keep allocations near-optimal as demand drifts.","key_machinery":"The central object is the temporal user arrival representation vector $v_t(\\lambda) = \\sum_i \\max_j (r_{ij} - \\lambda c_{ij})$, which collapses the population's best adjusted rewards into a one-dimensional function of the dual variable $\\lambda$. UDuo forecasts future values of this vector rather than the raw arrival distribution, then binary-searches the discretized dual objective $\\lambda B_t + v_t(\\lambda)$ over the grid $\\Lambda$ to obtain $\\lambda_t^*$, which feeds the decision rule $x_{ij}=1$ iff $j=\\arg\\max_j (r_{ij}-\\lambda_t^* c_{ij})$.","core_discovery":"The central proposal is that the dual objective for the fractional matching LP can be re-expressed through the user arrival representation vector $v(\\lambda)=\\sum_i\\max_j(r_{ij}-\\lambda c_{ij})$, so the online problem becomes: forecast $v_{t+H}(\\lambda)$, pace the budget, and binary-search the grid $\\Lambda$ for $\\lambda_t^*$ using $\\lambda_t^*=\\arg\\min_{\\lambda_i\\in\\Lambda}[\\lambda_i B + v_t(\\lambda_i)]$. UDuo claims this forecast-then-search procedure achieves equivalent optimality to gradient descent with $O(\\log|C|\\,1/\\epsilon)$ complexity, keeps the total budget feasible through pacing $\\sum_t B_t\\le B$, and adapts to distribution drift through a sliding-window update. On Eleme's online traffic the paper reports that UDuo lifts orders by 0.72%, GMV by 3.63%, and profit by 3.37% relative to a stochastic online-gradient-descent baseline, and that its MiFormer forecaster reduces MSE by 8% across user-arrival-vector and benchmark datasets.","pith_inferences":["If one could bound forecast error in $v_t(\\lambda)$, UDuo's framework would extend naturally to multiple resource constraints by adding one dual variable per constraint, a direction the paper only sketches as future work.","The representation $v_t(\\lambda)$ collapses an entire user population into a scalar curve; a testable extension is to compare this compressed forecast against forecasting the full arrival distribution and then aggregating, to see when the compression loses information.","The feasibility guarantee is only as strong as the pacing scheme; a testable extension is an online controller that adjusts $B_t$ from realized spend, protecting feasibility when historical pacing patterns fail."],"forward_implications":["Replacing stationary stochastic arrivals with forecasted representation vectors gives a principled way to keep dual solutions fresh in nonstationary environments.","The binary-search procedure over a discretized grid achieves the same solution as gradient descent at $O(\\log|C|\\,1/\\epsilon)$ complexity, so the forecasting step, not the optimizer, is the practical bottleneck.","Budget pacing plus the total-budget constraint $\\sum_t B_t \\le B$ turns a global budget constraint into per-period pacing, enabling proactive allocation across time slots.","On the reported A/B test, UDuo improves orders by 0.72%, GMV by 3.63%, and profit by 3.37% over the stochastic-model baseline, while the MiFormer forecaster cuts MSE by 8% on held-out forecasting benchmarks."],"supporting_citations":[{"why":"Introduces the drifting user arrival distribution model that motivates the non-stationary setting UDuo targets.","marker":"[1]"},{"why":"Supplies the L-BFGS gradient descent method that UDuo compares against for solving the dual problem.","marker":"[2]"},{"why":"Provides the binary-search approach for the dual variable that UDuo adapts in Eq. (6).","marker":"[3]"},{"why":"PepNet's parameter-and-embedding personalization paradigm inspires MiFormer's gated scenario-aware parameterization.","marker":"[4]"},{"why":"iTransformer is the base time-series architecture that MiFormer extends with scenario embeddings.","marker":"[5]"}],"fun_headline_variants":["UDuo: forecast arrivals, pace budgets, match online","Forecast-then-search duals for drift-adaptive online matching","UDuo predicts arrival vectors to keep allocations near-optimal","Dual forecasting beats stochastic models in online matching","UDuo: forecast the dual, pace the budget, match online"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the forecasted user-arrival summaries $v_{t+H}(\\lambda)$ are accurate enough for the grid search in Eq. (6) to return a near-optimal dual solution, and that the fixed grid $\\Lambda$, centered at the stochastic model's solution, still contains the true optimal $\\lambda$ after the distribution drifts.","fun_headline_variants_meta":{"raw":{"variants":["UDuo: forecast arrivals, pace budgets, match online","Forecast-then-search duals for drift-adaptive online matching","UDuo predicts arrival vectors to keep allocations near-optimal","Dual forecasting beats stochastic models in online matching","UDuo: forecast the dual, pace the budget, match online"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000355,"raw_usage":{"total_tokens":1922,"prompt_tokens":934,"completion_tokens":988,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":919}},"tokens_in":550,"tokens_out":988,"duration_ms":9873,"temperature":1.0,"reasoning_tokens":919,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:11:32.556936+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run UDuo on a synthetic user-arrival process with a known offline optimum, apply gradual and abrupt distribution drift, and compare the online allocation and budget spend against the offline LP optimum; if forecast error on $v_{t+H}(\\lambda)$ is not tightly coupled to suboptimality or budget violation, the claimed guarantees fail.","supporting_citations":[{"cited_title":"Robust online matching with user arrival distribution drift","cited_arxiv_id":null,"evidence_quote":"Introduces the drifting user arrival distribution model that motivates the non-stationary setting UDuo targets."},{"cited_title":"Direct heterogeneous causal learning for resource allocation problems in marketing","cited_arxiv_id":null,"evidence_quote":"Provides the binary-search approach for the dual variable that UDuo adapts in Eq. (6)."},{"cited_title":"Pepnet: Parameter and embedding personalized network for infusing with personalized prior information","cited_arxiv_id":null,"evidence_quote":"PepNet's parameter-and-embedding personalization paradigm inspires MiFormer's gated scenario-aware parameterization."}],"review_version":1}