{"id":"2a429ae7-d34f-4616-9917-5b2e28ae5c34","arxiv_id":"2608.00123","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"LLM-OSDA couples bid-dependent optimal stopping with envelope pricing to auction a single native ad insertion in multi-turn LLM conversations; the ideal mechanism is DSIC in expectation, and a learned version gains 11% net revenue over fixed-timing baselines in simulation.","lead":"An auction mechanism that decides not just which ad wins but which turn of a multi-turn LLM conversation the ad appears in, using optimal stopping and standard truthful-pricing machinery. On a simulated conversational shopping corpus, it reports an 11% net-revenue gain over the best fixed-timing baseline.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Bid leakage through bid-aware synthesis violates A1/A4 and makes the frozen-replay envelope payment invalid under a bid-dependent dialogue distribution.","rationale":"Good-faith reading: the paper is an auction-design contribution with a clean single-parameter model. Under A1-A4, Lemma 1 and Theorem 2 are a standard envelope argument and are not the weak spot. The exact-Bellman monotonicity proof is correct: fixing any feasible plan, the objective is affine in the score, so optimal allocations are monotone by Topkis. The special case T=1 recovering Myerson is a nice sanity check. The novel empirical claim is that the layered implementation preserves these properties in a simulated LLM environment.\n\nThe load-bearing soft spot is the gap between the model's bid-independence assumption and the data-generation protocol. The appendix says the assistant may see candidate bids during synthesis. This is not an artifact; it is a deliberate platform tilt. It creates an indirect channel: dialogue histories used to train/evaluate G_eta and to condition the renderer/judge may be bid-correlated even after bids are stripped at data-loading time. The frozen-replay payment calculation compounds this because it reuses dialogues generated at the true bid for all counterfactual bids, which is only valid under bid-independent D. The paper's own 'bid-in-pCTR' and 'bid-in-renderer' ablations test direct leakage, not this indirect textual leakage, so they do not rule it out.\n\nCould this be fixed by a different test? Yes. A counterfactual regeneration with bid-blind synthesis would reveal whether the 11% gain and the monotonicity diagnostics survive. If they do, the concern is resolved; if not, the headline empirical claims need to be downgraded to conditional on a bid-independent data-generation protocol.\n\nVerdict: the theoretical core is sound; the empirical evaluation is conditional on eliminating bid leakage. The reader's conditional verdict is appropriate and should be retained.","tokens_in":18145,"tokens_out":13739,"duration_ms":160186,"concrete_test":"Regenerate the evaluation corpus in two versions: (a) the current synthesis protocol where the assistant sees true candidate bids, and (b) a bid-independent control where the assistant sees randomly permuted bids (or no bid information) for the same sampled dialogues, keeping all other generation prompts and seeds identical. Then rerun the full pipeline (G_eta, StopNet, renderer, envelope payment with the same frozen-replay procedure) on both versions. If the LLM-OSDA vs. Always-Round-1 net-revenue gap in the control shifts by more than the reported +/-0.02 standard error, or if the monotonicity rate drops by more than 5 percentage points, the bid leakage is consequential and the 11% claim is confounded; if both remain within noise, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central DSIC and revenue claims rest on A1-A4: the dialogue distribution D and click law q_i must be independent of bids. The Data Generation appendix states that during dialogue synthesis 'the assistant may see candidate bids, reflecting a platform's tilt toward higher-paying ads.' This directly violates A1/A4: bids can enter the dialogue text, so the pre-insertion history h_t (and hence q_i(h_t) and the judge's click labels) can correlate with bids even though G_eta and pi_R never receive bid features. The paper's layering ablation (Table 2) only tests direct bid features (bid-in-pCTR, bid-in-renderer), not this indirect textual channel. Moreover, the payment computation replays a frozen evaluation schedule across the 64-point bid grid: the same dialogue outcomes are reused for every counterfactual bid z. If D is bid-dependent, the counterfactual dialogue at bid z differs from the actual dialogue generated at the true bid, so the estimated allocation curve x_hat_i^(R)(z) is not a valid estimate of the expected-click allocation under report z, and the envelope integral is computed under the wrong distribution. The empirical monotonicity (88.5%) and revenue (11%) diagnostics are therefore potentially confounded. The exact-Bellman theorem is unaffected, but the paper's claim that 'bids enter only the committed auction mechanism' is not supported by the evaluation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LLM-OSDA, a dynamic cost-per-click auction for native advertising in multi-turn LLM conversations. A bid-independent LLM layer estimates contextual click quality and renders the winning ad, while a committed auction mechanism decides when to stop, which advertiser wins, and the per-click price. The main theoretical result is that, under assumptions A1–A4 (bid-independent dialogue paths and click law), the ideal Bellman expected-discounted-click allocation is monotone in each advertiser's bid, and envelope pricing yields DSIC in expectation and click-contingent individual rationality. A learned StopNet approximates the Bellman action values; the paper provides a conditional approximate-IC bound. Experiments on a simulated conversational corpus report an 11% net-revenue gain over the strongest fixed-timing baseline, with diagnostics on monotonicity and regret.","tokens_in":18465,"tokens_out":9026,"duration_ms":99246,"significance":"The mechanism-design question is timely and the theoretical construction is valuable. Lemma 1 and Theorem 2 are standard single-parameter envelope arguments correctly adapted to endogenous stopping, and the T=1 reduction to a quality-weighted Myerson auction is a useful sanity check. The appendix gives full proofs, and the paper is unusually transparent that the deployed approximation is not covered by a uniform theoretical guarantee. The main weakness is that the empirical evaluation does not actually instantiate the bid-independent-dialogue assumption: the data-generation process lets the synthesizing assistant see candidate bids, and the payment computation replays the same frozen dialogue outcomes across the bid grid. This undermines the empirical monotonicity, envelope-pricing, and revenue claims as a test of the theoretical model. The exact-Bellman theory is unaffected, but the bridge from theory to the reported 11% result needs substantial rework.","major_comments":[{"comment":"The empirical evaluation violates the bid-independent dialogue assumption that Lemma 1 and Theorem 2 require. The Data Generation appendix states: 'During synthesis the assistant may see candidate bids, reflecting a platform's tilt toward higher-paying ads.' This makes the dialogue distribution D bid-dependent, contradicting the main-text assumption that D is 'the distribution over bid-independent potential dialogue paths, assumed independent of bids.' Because the payment computation replays a frozen evaluation schedule across the 64-point bid grid, the same dialogue outcomes are reused for every counterfactual bid z; under a bid-dependent D, the counterfactual dialogue at report z differs from the actual dialogue at the true bid, so the estimator \\hat{x}_i^{(R)}(z) is not a valid estimate of the expected-click allocation under report z. Consequently the envelope integral in Eq. (2) is c","section":"Theoretical Guarantees (A1/A4) and Additional Experiments: Data Generation; Envelope pricing, Eq. (2)"},{"comment":"The abstract states that for practical deployment 'we bound the resulting incentive loss in terms of its approximation error.' Proposition 4 is a conditional bound, but it is stated in terms of unknown quantities δ_x and δ_m, and the paper explicitly says: 'The reported finite-grid regret does not estimate this uniform bound.' The diagnostics (88.5% monotone sessions; 89% sessions with zero observed regret) are therefore not a bound on the incentive loss. The main text is appropriately cautious, but the abstract should be qualified so readers do not confuse empirical diagnostics with a proved approximation guarantee.","section":"Mechanism Diagnostics and Proofs (Proposition 4)"},{"comment":"The click and retention labels used for evaluation are produced by an LLM judge from the same model family as the labels used to train the quality model and StopNet. The paper claims 'the reported revenue is not a self-evaluation' because the judge is not the pCTR model, but the judge is still an internal simulator, not external ground truth. The 11% net-revenue figure is therefore a simulator-internal result and should be presented as such, rather than as evidence about real user behavior or real advertiser payments.","section":"Metric Definitions and Experimental Setup"}],"minor_comments":[{"comment":"The statement that 'the assistant may see candidate bids during synthesis' should be reconciled with the main-text claim that 'bids are stripped at data-loading time.' They are not literally contradictory, since the synthesizing assistant is not the deployed renderer, but the current wording makes the bid-independence assumption look violated in the evaluation corpus.","section":"Additional Experiments: Data Generation"},{"comment":"Table 4 reports LLM-OSDA net revenue 0.777 and Always-Round-1 0.689, which do not match any row of Table 1 (0.769/0.818 for LLM-OSDA and 0.693/0.733 for Always-Round-1). Please specify whether Table 4 uses base or refined components and under which evaluation round.","section":"Adaptive Timing Baseline (Table 4)"},{"comment":"The phrase 'we bound the resulting incentive loss in terms of its approximation error' should say 'conditional on the approximation errors δ_x, δ_m' to avoid overstating the strength of Proposition 4.","section":"Abstract"},{"comment":"The caption has a typo/formatting issue: 'PENV Pcrit' should be typeset with subscripts; also clarify the units and whether the scatter is per-session or per-advertiser.","section":"Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The theoretical contribution is sound and the paper is honest about many limitations, but the data-generation issue is a serious gap between the model assumptions and the empirical test. I do not think this requires rejection: the authors could regenerate their corpus with truly bid-independent dialogue synthesis, or explicitly model the bid-dependent dialogue distribution and re-derive the mechanism. However, as written, the headline empirical claims (monotonicity diagnostics, envelope CPC, 11% revenue gain) are not valid tests of the paper's DSIC/spricing theory. The paper would also benefit from stating in the abstract that the empirical evaluation is wholly within an LLM-based simulation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper builds a clean optimal-stopping auction for multi-turn native ads, with a standard envelope-pricing core that works if the dialogue distribution is bid-independent. The theory is fine. The empirical evaluation, though, is compromised by the data-generation step: during dialogue synthesis the assistant may see candidate bids, so the conversation history itself can carry bid information. That breaks the A1/A4 assumption the whole pricing argument rests on, and it makes the frozen-replay envelope payment invalid. The authors should fix this before the 11% number is taken seriously.\n\nWhat's actually new: combining Bellman stopping with single-parameter envelope pricing for a single-insertion CPC auction, and keeping the quality estimator bid-independent so that the allocation monotonicity lemma goes through. The exact-Bellman DSIC/IR proof is standard (Myerson/Milgrom-Segal), but the application to bid-dependent timing is the contribution. The paper is honest about the learned StopNet: it doesn't claim a uniform bound, reports monotonicity and regret as diagnostics, and provides stopping-stability and error-to-IC transfer lemmas. That's good practice.\n\nThe soft spots, in order of severity:\n\n1. Bid leakage. The appendix says the assistant may see candidate bids during dialogue synthesis. So the bid can influence the dialogue path, which means q_i(h_t) is not bid-independent, and the single-parameter structure is gone. The payment computation replays the same dialogue across the bid grid; that's only valid if the dialogue distribution is bid-independent. If it's not, the allocation curve x_hat^(R)(z) is estimated under the wrong distribution, and the envelope integral is meaningless. The layering ablation only tests direct bid features in G_eta or renderer; it does not test the indirect textual channel. This is a real gap.\n\n2. Self-referential evaluation. The envelope CPC is computed from the system's own pCTR estimates, and even though the judge is independent, the payment is internal. Combined with the first issue, the 11% net-revenue improvement over the fixed-timing baseline is not externally grounded. The myopic baseline in the appendix also uses the same potentially biased G_eta, so it doesn't cleanly separate look-ahead from leakage.\n\n3. Minor: the monotonicity diagnostic (88.5%) is an empirical grid check, which the authors acknowledge. That's fine, but it doesn't rescue the payment computation.\n\nThe theory is correct and worth having. The empirical claims need stronger support. I'd send this to peer review, asking for a revised data-generation protocol that eliminates bid influence on dialogue, plus a test for indirect leakage (e.g., correlation between log bid and dialogue features or judge clicks). The paper is not ready as-is, but it's a serious piece of work.","headline":"Solid theory for bid-dependent stopping with envelope pricing, but the simulated evaluation has a bid-leakage problem that taints the 11% revenue claim.","tokens_in":18970,"tokens_out":4300,"would_cite":true,"duration_ms":47589,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91B26","60G40"],"pacs":[],"model":"deepseek-v4-flash","headline":"A dynamic auction that decides when a chatbot shows an ad keeps truthful bidding in expectation and raises net revenue 11% over fixed-timing insertion.","keywords":["dynamic auction","optimal stopping","envelope pricing","LLM-native advertising","multi-turn conversation","incentive compatibility","cost-per-click","Bellman recursion"],"falsifier":"Replay the deployed mechanism with one advertiser's bid swept over a fine grid while holding rival bids and dialogue rollouts fixed, scoring clicks with an independent judge rather than the learned pCTR; any downward step in the expected discounted-click allocation curve—or any session where a higher bid moves insertion to a lower-click turn without changing the winner—would falsify the monotonicity on which DSIC relies. A sharper test: generate two versions of the same dialogues, one where the assistant saw candidate bids during synthesis and one where it did not, and compare the non-monotone","tokens_in":18044,"feed_emoji":"📣","tokens_out":12557,"duration_ms":125584,"temperature":0.7,"pith_summary":"This paper tackles a new unit of sale in advertising: instead of selling a fixed slot, the platform sells one moment inside a multi-turn LLM conversation. The central claim is that the platform can decide when to insert the ad, which advertiser wins, and what price per click to charge—all in a way that keeps truthful bidding dominant in expectation, even though the insertion time itself responds to bids. The paper achieves this by keeping the LLM's click-quality estimates and ad rendering blind to bids, so that only the auction mechanism sees the bids, and by pricing through a single-parameter envelope formula built on the expected discounted number of clicks. Under an exact Bellman stopping rule the allocation is monotone in each bid, which makes envelope pricing incentive-compatible; a learned approximation is shown to deviate only near the stopping boundary. On a simulated shopping-dialogue corpus, the mechanism raises net revenue by 11% over the strongest fixed-timing baseline without lowering user retention.","feed_headline":"Auction that times chatbot ads lifts revenue 11 percent","feed_subtitle":"It decides when to insert a sponsored reply, keeps bidding truthful, and holds user retention.","key_machinery":"The load-bearing object is the expected discounted-click allocation x_i(b_i,b_{-i}) = E[γ^{τ(b)-1} q_i(h_{τ(b)}) 1{I(b)=i}], where τ is the stopping time, γ the per-turn continuation probability, and q_i the click probability at the insertion turn. The monotonicity lemma shows this single-parameter curve is nondecreasing in the bid, and the envelope payment m_i(b_i)=b_i x_i(b_i)-∫_{θ_i}^{b_i}x_i(z)dz converts that monotonicity into DSIC. Supporting machinery: the Bellman recursion V_t(h_t;b)=max{max_i ψ_i(b_i)q_i(h_t), γ E[V_{t+1}|h_t]}, a bid-independent LLM quality estimator G_η and renderer π_R that keep the click law bid-independent, and a learned StopNet whose two heads regress the stop","core_discovery":"The paper establishes that classical single-parameter auction machinery survives the move from fixed slots to endogenous stopping times, provided the allocation object is redefined as the expected discounted click probability and the LLM's quality and rendering layers are kept bid-independent. Formally, with a bid-independent dialogue distribution and click law, the ideal Bellman mechanism—stop at the first turn where max_i ψ_i(b_i)q_i(h_t) exceeds the discounted continuation value, allocate to the argmax, and charge the envelope CPC m_i(b_i)/x_i(b_i) with m_i(b_i)=b_i x_i(b_i)-∫_{θ_i}^{b_i}x_i(z)dz—makes truthful bidding weakly dominant in expectation and satisfies click-contingent individu","pith_inferences":["Editorial inference: the observed 5.1% gap between the envelope CPC and the critical-bid CPC could be used as a direct diagnostic for how much of a mechanism's revenue comes from timing option value rather than winner selection; a fixed-timing mechanism should show zero gap.","Editorial inference: the assumption that bids never leak into dialogue content is testable in production by measuring whether the winning bid correlates with the LLM's wording or with judged click probability; the paper's own synthesis step, where the assistant may see candidate bids, is a place this leakage could appear.","Editorial inference: the single-insertion restriction suggests a natural extension—multi-insertion sessions would require tracking a second stopping-allocation dimension, and the monotonicity and envelope arguments would need to be re-derived over the joint timing sequence."],"forward_implications":["Platforms can sell a single native insertion per conversational session as a committed mechanism: timing, winner, and click-contingent price are all set by the auction before the ad is rendered.","The envelope CPC prices the option value of timing: because raising a bid can change both the winner and the insertion turn, the payment integrates over the whole allocation curve, which is why it can exceed the single critical-report threshold.","The single-turn limit recovers classical auctions: with T=1, identity scoring gives a quality-weighted welfare auction, and a virtual-value score gives a quality-weighted Myerson auction.","A learned approximation preserves the guarantees approximately: StopNet decisions differ from optimal only near the stopping boundary, and the value loss and incentive loss are bounded by its approximation error.","Timing can raise revenue without hurting retention in the tested setting: the learned trigger turn sits between round-1 and last-round, retention is unchanged, and a myopic threshold baseline does not match the Bellman look-ahead."],"supporting_citations":[{"why":"Supplies the single-parameter envelope construction and virtual-value framework that the payment rule and T=1 reduction build on.","marker":"Myerson 1981"},{"why":"Supplies the envelope theorem used to prove that a monotone expected-click allocation plus envelope payment yields DSIC in expectation.","marker":"Milgrom and Segal 2002"},{"why":"Provides the optimal-stopping theory and Bellman recursion that define the insertion-time rule.","marker":"Peskir and Shiryaev 2006"},{"why":"Gives the selling-options interpretation of timing, which the paper adapts to price the option value of waiting.","marker":"Board 2007"},{"why":"Represents the token-level LLM auction approach that the paper contrasts with its session-level stopping design.","marker":"Dütting et al. 2024b"},{"why":"Represents the RAG-based segment auction baseline that allocates ads within a response but does not choose timing.","marker":"Hajiaghayi et al. 2024"},{"why":"Closest independent work on the same optimal-stopping LLM ad problem, using a distributional user-belief model instead of LLM-estimated signals.","marker":"Alaei, Makhdoumi, and Malekian 2026"},{"why":"Provides the LLM-as-a-judge paradigm that generates click and retention labels for the simulated corpus.","marker":"Zheng et al. 2023"},{"why":"Supplies the Amazon Reviews 2023 corpus from which the simulated shopping dialogues and ad pool are drawn.","marker":"Hou et al. 2024"}],"fun_headline_variants":["Auction that times chatbot ads lifts revenue 11%","Timing ad bids in chat boosts revenue 11%","Optimal-stopping auction for LLM ads gains 11%","Dynamic ad auction in chats: truthful, +11% revenue","When to insert a chatbot ad? Auction decides, up 11%"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The truthful-bidding guarantee assumes bids never influence the dialogue paths or the user's click behavior—only the auction's stop/allocate/price decision—and the paper's own data-generation step, where the assistant may see candidate bids, is where that assumption could silently break.","fun_headline_variants_meta":{"raw":{"variants":["Auction that times chatbot ads lifts revenue 11%","Timing ad bids in chat boosts revenue 11%","Optimal-stopping auction for LLM ads gains 11%","Dynamic ad auction in chats: truthful, +11% revenue","When to insert a chatbot ad? Auction decides, up 11%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000114,"raw_usage":{"total_tokens":933,"prompt_tokens":800,"completion_tokens":133,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":60}},"tokens_in":544,"tokens_out":133,"duration_ms":3230,"temperature":1.0,"reasoning_tokens":60,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T04:21:34.370030+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replay the deployed mechanism with one advertiser's bid swept over a fine grid while holding rival bids and dialogue rollouts fixed, scoring clicks with an independent judge rather than the learned pCTR; any downward step in the expected discounted-click allocation curve—or any session where a higher bid moves insertion to a lower-click turn without changing the winner—would falsify the monotonicity on which DSIC relies. A sharper test: generate two versions of the same dialogues, one where the assistant saw candidate bids during synthesis and one where it did not, and compare the non-monotone","supporting_citations":[],"review_version":2}