{"id":"919d57bc-7e62-43b2-acf4-2f33e446fb8d","arxiv_id":"2606.08360","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Generative Frontier Planning replaces Monte-Carlo sampling with a deterministic backup over a latent covariate-coverage surrogate to achieve tractable (1-1/e)-approximate per-round allocation in covariate-dependent peer-referral systems.","lead":"The paper introduces Generative Frontier Planning (GFP), a model-based method that uses a latent surrogate to plan adaptive referral allocations when future recruits depend on the covariates of current referrers. A smart generalist might read it to see how generative models combined with structured planning can improve resource allocation in real-world recruitment for hidden populations.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Surrogate may not guarantee diminishing returns when generative model conditions on referrer covariates","rationale":"The reader's weakest_assumption directly identifies the surrogate properties as load-bearing. The concrete test above would settle whether those properties actually survive the conditional generative model used in the paper.","tokens_in":1815,"tokens_out":289,"duration_ms":10703,"concrete_test":"Extract the exact surrogate definition and finite-dimensional summaries from the GFP design section; then, on the calibrated simulation, evaluate the per-round objective for all subsets of size k and k+1 (k=1..10) over 100 random referrer states and check whether marginal gains are nonincreasing for every state. If any state violates diminishing returns, the approximation guarantee does not hold.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim rests on the surrogate making the per-round objective monotone submodular, enabling the (1-1/e) greedy guarantee. The abstract states this is achieved by making expected next-frontier value depend on the offspring model only through finite-dimensional summaries. When the generative model is conditional (covariates of new recruits depend on referrer), the coverage-value function can lose submodularity unless the summaries are chosen to preserve it exactly; nothing in the given description shows that the chosen summaries (whatever they are) provably retain the property for arbitrary conditional distributions learned from censored data.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper claims that Generative Frontier Planning (GFP) solves adaptive referral-resource allocation under covariate-dependent arrivals by replacing Monte-Carlo rollouts with a deterministic backup over a latent covariate-coverage value surrogate; the surrogate is constructed so that the per-round objective is monotone submodular (yielding a (1-1/e) greedy guarantee) and depends on the conditional generative model only through finite-dimensional summaries amortized offline. On a simulation environment calibrated to a real respondent-driven sampling dataset, GFP outperforms random, RL, and i.i.d. dynamic-programming baselines across four discount factors.","tokens_in":1975,"tokens_out":379,"duration_ms":14441,"significance":"If the surrogate construction rigorously preserves monotonicity and diminishing returns for arbitrary conditional distributions learned from censored count data, the approach would supply a tractable, approximately optimal planner for non-i.i.d. recruitment dynamics that are common in hidden-population studies; the simulation results provide initial evidence of practical gains over simpler baselines.","major_comments":[{"comment":"Abstract (paragraph on GFP design): the central (1-1/e) guarantee rests on the surrogate inducing monotonicity and diminishing returns for the per-round objective when the offspring generative model is conditional on referrer covariates. The manuscript states that this is achieved by making expected next-frontier value depend only on finite-dimensional summaries, but supplies neither the explicit form of those summaries nor a proof that submodularity is retained for arbitrary conditional distributions fitted to censored data; without this, the approximation claim is not yet substantiated.","section":"Abstract (GFP design paragraph)"}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":"The evaluation is simulation-only with no public code or calibrated dataset; this limits independent verification of the reported outperformance and of whether the surrogate construction actually works on the real RDS data used for calibration."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their careful reading and for identifying the need to strengthen the substantiation of the (1-1/e) guarantee. We address the major comment below and will incorporate the requested details in the revised manuscript.","responses":[{"response":"We agree that the current manuscript provides only a high-level description of the surrogate and does not include the explicit form of the finite-dimensional summaries or a self-contained proof of submodularity retention under arbitrary conditional generative models fitted to censored data. In the revision we will (i) explicitly define the summaries in Section 3.2 as the vector of expected coverage statistics E[φ(X) | referrer covariates] obtained from the amortized conditional generative model, and (ii) add a dedicated appendix containing the formal proof that the surrogate value function remains monotone and submodular because it is linear in these coverage statistics; the linearity argument holds for any valid conditional distribution and therefore applies to models learned from censored count data. These additions will directly substantiate the (1-1/e) per-round guarantee.","revision_made":"yes","referee_comment":"[Abstract (GFP design paragraph)] Abstract (paragraph on GFP design): the central (1-1/e) guarantee rests on the surrogate inducing monotonicity and diminishing returns for the per-round objective when the offspring generative model is conditional on referrer covariates. The manuscript states that this is achieved by making expected next-frontier value depend only on finite-dimensional summaries, but supplies neither the explicit form of those summaries nor a proof that submodularity is retained for arbitrary conditional distributions fitted to censored data; without this, the approximation claim is not yet substantiated."}],"tokens_in":1427,"tokens_out":358,"duration_ms":12386,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The punchline is that this paper gives a way to plan referral allocations when new recruits' covariates depend on who refers them, using a surrogate to avoid sampling and get a greedy guarantee. That's the core advance over i.i.d. dynamic programming baselines.\n\nThey model the arrival process with a censored count model and a conditional generative model, then build a latent covariate-coverage value surrogate so the expected next frontier depends only on finite-dimensional summaries amortized offline. This lets them do deterministic backup and claim the per-round objective is monotone submodular, so marginal greedy gets (1-1/e). On simulations from a real respondent-driven sampling dataset, it beats random, RL, and i.i.d. DP across four discount factors.\n\nThe construction is new in handling the conditional case without losing tractability. The simulation results look solid for what they are.\n\nThe soft spots are around the surrogate itself. The abstract says it's designed to preserve monotonicity and diminishing returns, but there's no derivation or proof sketch showing that the summaries work for general conditional distributions learned from censored data. The stress-test concern is fair: if the generative model conditions on referrer covariates, it's not obvious the value function stays submodular. Without the full construction or code, it's hard to verify. Also, everything is simulation; no real deployment or sensitivity checks on the model fit.\n\nThis paper is for researchers in adaptive sampling and public health modeling who want to move past i.i.d. assumptions. A reader working on submodular planning with generative models could find the surrogate idea useful if the details hold.\n\nIt deserves peer review because the problem is practical and the method is a genuine attempt to handle the conditional case, even if the guarantee needs more support.","headline":"GFP gives a surrogate trick for planning under conditional recruitment but the submodularity needs checking.","tokens_in":2476,"tokens_out":415,"would_cite":false,"duration_ms":13779,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Generative Frontier Planning adapts referral resources by replacing Monte Carlo with a deterministic surrogate that supports a (1-1/e) greedy approximation per round.","keywords":["generative frontier planning","peer-referral recruitment","respondent-driven sampling","covariate-dependent arrivals","adaptive resource allocation","surrogate planning","diminishing returns","approximation algorithms"],"falsifier":"If GFP does not outperform the random, reinforcement-learning, and i.i.d. dynamic-programming baselines on the simulation environment calibrated to the real respondent-driven sampling dataset across the four tested discount factors, the claimed practical advantage is refuted.","tokens_in":2727,"feed_emoji":"","tokens_out":745,"duration_ms":11936,"temperature":0.7,"pith_summary":"The paper models peer-referral recruitment where each referral's capacity and the new recruit's covariates depend on the referrer, learned from data via a censored count model and conditional generative model. It replaces standard Monte Carlo sampling in planning with a deterministic backup that uses a latent covariate-coverage surrogate whose value depends on the generative model only through finite-dimensional summaries computed offline. The surrogate is constructed so the per-round objective is monotone and exhibits diminishing returns, which lets a marginal-greedy rule achieve a (1-1/e) approximation guarantee. On simulations calibrated to a real respondent-driven sampling dataset, the resulting GFP planner outperforms random allocation, reinforcement learning, and i.i.d. dynamic programming baselines across four discount factors.","feed_headline":"Surrogate replaces Monte Carlo for adaptive referral allocation","feed_subtitle":"GFP uses a covariate-coverage surrogate to enable deterministic backups and a (1-1/e) greedy rule, beating baselines on real-data-calibrated","key_machinery":"The latent covariate-coverage value surrogate, which encodes expected frontier value through offline-amortized finite-dimensional summaries and induces a monotone diminishing-returns objective.","core_discovery":"GFP replaces per-step Monte-Carlo sampling with a deterministic backup over a latent covariate-coverage value surrogate. The surrogate is designed so that the expected value of the next frontier depends on the offspring generative model only through finite-dimensional summaries that are amortized offline, and so that the resulting per-round objective is monotone with diminishing returns. Together these properties make planning tractable and let marginal greedy allocation achieve a (1-1/e)-approximation for the per-round problem.","pith_inferences":["The same surrogate structure could be reused for other sequential allocation tasks whose state evolution is given by a conditional generative model.","If the finite-dimensional summaries preserve the main homophily effects, the method may transfer to other hidden-population interventions that rely on peer chains.","Live deployment would require checking whether the offline-amortized summaries remain accurate when the generative model is updated from new referrals.","The diminishing-returns property might allow hybrid planners that combine the greedy step with occasional lookahead without losing the approximation bound."],"forward_implications":["Each candidate allocation induces a different future-recruit distribution that the surrogate summarizes without repeated sampling.","Marginal greedy allocation on the per-round problem is guaranteed a (1-1/e) approximation.","Planning runs with deterministic backups instead of Monte Carlo rollouts.","The approach handles homophily and shared context that i.i.d. models ignore."],"fun_headline_variants":["Latent surrogate supports deterministic backup over covariate coverage in GFP","Amortized summaries make per round objective monotone with diminishing returns","GFP replaces Monte Carlo with deterministic planning for peer referral systems","Adaptive allocation under covariate dependent arrivals uses generative frontier planner"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The surrogate can be constructed so the expected next frontier depends on the generative model only through finite-dimensional summaries and the per-round objective is monotone with diminishing returns.","fun_headline_variants_meta":{"raw":{"variants":["Latent surrogate supports deterministic backup over covariate coverage in GFP","Amortized summaries make per round objective monotone with diminishing returns","GFP replaces Monte Carlo with deterministic planning for peer referral systems","Adaptive allocation under covariate dependent arrivals uses generative frontier planner"]},"model":"grok-4.3","cost_usd":0.004546,"raw_usage":{"total_tokens":2307,"prompt_tokens":762,"num_sources_used":0,"completion_tokens":65,"cost_in_usd_ticks":45462000,"prompt_tokens_details":{"text_tokens":762,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1480,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":762,"tokens_out":65,"duration_ms":8539,"temperature":1.0,"reasoning_tokens":1480,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T19:52:45.544967+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"If GFP does not outperform the random, reinforcement-learning, and i.i.d. dynamic-programming baselines on the simulation environment calibrated to the real respondent-driven sampling dataset across the four tested discount factors, the claimed practical advantage is refuted.","supporting_citations":[],"review_version":1}