{"id":"dfac9cd6-f7d2-4dfc-8f85-9c0678098838","arxiv_id":"2608.12470","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Tailed-Uniform proposal distributions, which pad the uniform training box with decaying tails, improve neural posterior estimates near and beyond the prior boundaries in simulation-based inference.","lead":"This paper proposes a new way to sample training simulations for neural posterior estimation: instead of a hard-edged uniform box, it adds smooth tails beyond the box edges. Training with these 'tailed uniform' proposals makes the estimated posteriors more accurate near and beyond the training boundaries, which matters for large astrophysical simulation-based inference campaigns.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Under Eq. (1), tail samples carry zero importance weight for Uniform assumed priors, so the reported Uniform-prior improvements in Sec. III are not mechanistically possible; the implementation or seed noise must be checked.","rationale":"The reader's declared weakest assumption is the Section V MCMC log-normal prior mismatch, which is real but only affects the cosmological demonstration. The more load-bearing concern, also present but secondary in the reader's rationale, is the zero-importance-weight inconsistency: for a hard-truncated Uniform assumed prior, tail samples contribute exactly zero to the loss in Eq. (1), so they cannot improve the trained posterior through the mechanism the paper invokes. The toy experiment in Section III and Table II is the primary support for the abstract's 'even for Uniform assumed priors' claim, so if this inconsistency is not resolved, that central claim is unsupported. The Gaussian-prior results and the empirical scaling studies may still be valid, but they do not justify the strong Uniform-prior statement. The reader's CONDITIONAL verdict already requires clarification of the Uniform-prior effect and repeated-seed error bars; our analysis strengthens that requirement without changing the overall verdict. Hence UNCHANGED, with the caveat that if the code check shows exact zero weights and the benefit disappears, the verdict should move toward REJECT for the current claim.","tokens_in":12616,"tokens_out":8566,"duration_ms":91850,"concrete_test":"Inspect the loss implementation used by the LtU-ILI pipeline for the Gaussian Linear Uniform Task, then rerun the experiment for Uniform and Gaussian-tailed proposals with 10-20 independent seeds, reporting seed-marginalized C2ST error bars at each grid point. If the weighted loss in Eq. (1) is used, a minibatch containing only tail samples has identically zero gradient, and any Tailed-Uniform advantage must vanish; if an advantage persists, confirm whether the code drops the P/Ptilde weight. Analytically, compute the gradient of L_NPE with respect to flow parameters for a tail-only minibatch under P=0; it must be zero, so a nonzero empirical benefit would prove the implementation deviates from Eq. (1).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (1) defines the NPE loss with an explicit importance weight P(theta)/Ptilde(theta). In the Gaussian Linear Uniform Task (Sec. III), the assumed prior is P = U([-1,1]^d), while every Tailed-Uniform proposal places nonzero sampling mass outside that box. For every tail sample, P(theta)=0, so that sample's contribution to L_NPE and to its gradient is exactly zero. The mechanism offered in Sec. III.B, that 'extra samples outside the original boundaries ... constrain network behavior as it approaches the boundary,' therefore cannot operate: a zero-weight sample carries no learning signal. Under the stated loss, a Uniform-trained and a Tailed-Uniform-trained network with identical in-box proposal density and identical random seed should be statistically indistinguishable in the Uniform-prior task; Table II instead reports a consistent benefit. Either the implementation does not actually apply Eq. (1) (for example, it trains unweighted on the proposal, in which case q_w targets the proposal posterior rather than the assumed Uniform posterior and the comparison to the analytic reference is invalid), or the differences are within seed noise, which is not controlled for anywhere in the paper. Both possibilities undermine the abstract's 'even for Uniform assumed priors' claim. The cosmological Section V prior mismatch is a related but secondary confound; the zero-weight issue affects the central toy demonstration directly.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces Tailed-Uniform, a family of proposal distributions for simulation-based inference that replace the hard edge of a Uniform simulation box with decaying tails, and studies whether training neural posterior estimators on such proposals improves posterior accuracy near and beyond the original boundaries. The paper reports consistent C2ST improvements over a Uniform-trained baseline in two analytic Gaussian toy problems, in sensitivity sweeps over dimension, simulation budget, and network architecture, and in a cosmological matter-power-spectrum example. The central claim is that tail samples help even when the assumed prior is uniform, and that the benefits grow with dimensionality.","tokens_in":12777,"tokens_out":8966,"duration_ms":83384,"significance":"The practical question is timely: boundary artifacts in NPE training are a known nuisance in astrophysical simulation suites, and a data-efficient proposal fix would be widely useful. The paper has real strengths: the toy problems are analytically tractable and evaluated against external references, the code is public, and the sensitivity analysis covers tail shapes, widths, budgets, architectures, and dimensions in a systematic way. If the mechanism were established, this would be a solid methods contribution. However, the abstract's 'even for Uniform assumed priors' claim is in direct tension with the importance-weighting loss stated in Eq. (1), and the cosmological validation uses different priors for the MCMC reference and the NPEs; these issues are load-bearing and need to be resolved.","major_comments":[{"comment":"The central claim that Tailed-Uniform improves NPEs 'even for Uniform assumed priors' is not consistent with the loss defined in Eq. (1). When the assumed prior is P(theta)=U([-1,1]^d), every tail sample with theta outside the box has P(theta)=0, so its importance weight P(theta)/Ptilde(theta) and hence its contribution to L_NPE and its gradient are exactly zero. The mechanism described in Section III.B, that 'extra samples outside the original boundaries ... constrain network behavior as it approaches the boundary,' therefore cannot operate under the stated objective: the effective training objective is the standard uniform-prior NPE loss evaluated only on in-box samples. The reported advantage in Table II must then be either a finite-sample/seed effect or evidence that the implementation does not actually apply Eq. (1) (e.g., it trains unweighted on the proposal, in which case q_w targets the posterior under the proposal prior, not the assumed Uniform prior, and the comparison to the analytic truncated reference is invalid). Please clarify exactly which loss is implemented, and add a seed-controlled experiment in which the same in-box samples are used for Uniform and Tailed-Uniform training, with weights applied and then removed, to isolate any real mechanism.","section":"Section II, Eq. (1); Section III.B, Table II"},{"comment":"The cosmological validation is confounded by a prior mismatch. The MCMC reference posteriors are generated under a log-normal prior on (Omega_m,h) truncated at ±3 sigma, while all NPE networks are trained with a Uniform assumed prior on the boxes Omega_m in [0.27,0.37] and h in [0.63,0.71]. The C2ST comparisons in Figure 10 and Table III therefore measure not only boundary handling but also differences between the priors, and a Uniform-trained NPE may be penalized at the test point simply because the MCMC prior permits support outside the box. Please either generate the MCMC reference under the same Uniform box prior used by the NPEs, train the NPEs under the log-normal prior, or otherwise demonstrate that the comparison is insensitive to the prior choice.","section":"Section V.A, Fig. 9, Table III"},{"comment":"The abstract states that 'these benefits increase in high dimensions,' but the paper's own Figure 5 shows that at d=8 the Tailed-Uniform variants perform better near the boundary and extrapolation bin but degrade in the interior, and at d=12 the advantage disappears entirely (93% of the budget is in the tails). The claim should be qualified to reflect the crossover dimension d*(sigma) and the bias-variance trade-off described by Eq. (5); as written, the abstract overstates the high-dimensional benefit.","section":"Section IV.A, Fig. 5; Abstract"}],"minor_comments":[{"comment":"The continuity condition is stated verbally, but the derivation of the normalization constants A and B from normalization plus continuity is not shown; a one-line derivation would make the construction easier to verify.","section":"Section II.B, Table I"},{"comment":"The x-axis label '2 Extrap' is ambiguous; it should read '2σ Extrap' or be explicitly defined in the caption.","section":"Figures 5-8"},{"comment":"The mean and standard deviation of the log-normal prior on (Omega_m,h) are not specified; please provide the full prior definition so the reader can assess the overlap with the Uniform training box.","section":"Section V.A"},{"comment":"The sentence 'Uniform posteriors leak probability mass beyond the prior support' is misleading for the uniform-assumed-prior task: the true posterior has no support outside the box, so what leaks is the NPE estimate. Please state this explicitly.","section":"Section VI, Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The zero-weight issue in the uniform-prior experiments is serious enough that the central abstract claim cannot stand as written. I would ask the authors to verify the implemented loss in the public code before revision; if the code trains unweighted on the proposal, the toy 'Uniform assumed prior' section is not testing the stated method and the conclusion needs substantial reframing. The cosmological prior mismatch is straightforward to fix. I do not see a circularity problem: the method is evaluated against external references, and the tail-width hyperparameters are fixed by heuristic rather than tuned to the test points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nRead this one for the Gaussian-prior experiments, not for the headline. The Uniform-prior claim in the abstract does not hold up against the paper's own loss function.\n\nThe paper proposes Tailed-Uniform proposal distributions—uniform cores with Gaussian, exponential, linear, or flat tails—to fix NPE boundary pathologies. What's genuinely new is the systematic family of tails, the flat-tail control, and the dimensionality trade-off analysis (Eqs. 4–5). The sensitivity sweeps over tail width, simulation budget, architecture, and dimension are thorough, and the code is public. Under a Gaussian assumed prior, the tailed proposals clearly beat the hard-uniform baseline, and the mechanism is sound: tail samples carry nonzero importance weight and give the network signal beyond the box.\n\nThe soft spots are serious. First, in the Gaussian Linear Uniform task (Uniform assumed prior), Eq. (1) assigns zero importance weight to every sample outside the box. Those samples cannot contribute to the loss or gradient, so the reported advantage is either seed noise—no repeated seeds appear anywhere—or an implementation that silently omits the importance weight, in which case the network targets the proposal posterior rather than the Uniform-prior posterior and the comparison to the analytic reference is invalid. The paper acknowledges the objection but does not resolve it; I see no mechanism for tail samples to help under the stated loss.\n\nSecond, the cosmological validation compares NPEs trained with a Uniform prior to an MCMC reference drawn from a log-normal prior truncated at ±3σ. That is a prior mismatch, and the C2ST scores in Table III are not comparing like with like. Third, the paper lacks seed-averaged error bars; single runs carry the whole argument.\n\nCitation pattern is fine, relevant literature is covered. The paper is worth a serious referee—the Gaussian-prior results and the dimensionality analysis are well executed and useful to practitioners. A revision that drops or fixes the Uniform-prior claim, redoes the cosmology with matched priors, and adds repeated seeds would get my endorsement. As is, conditional at best. I'd bring it to reading group for the discussion, but I wouldn't cite the Uniform-prior result until it's fixed.\n\nBest,\n[Your name]","headline":"Useful proposal-distribution study with a clean Gaussian-prior result, but the Uniform-prior headline claim is undercut by an importance-weight inconsistency and a prior mismatch in the cosmology demo.","tokens_in":13364,"tokens_out":5874,"would_cite":false,"duration_ms":55798,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that sampling training simulations from a Tailed-Uniform proposal—a flat core with decaying tails—makes neural posterior estimators more accurate near and beyond the boundaries of the training box, even when the assumed…","keywords":["Tailed-Uniform","simulation-based inference","neural posterior estimation","proposal distribution","boundary effects","normalizing flows","cosmological parameter inference","classifier two-sample test"],"falsifier":"Rerun the cosmological benchmark with MCMC reference posteriors drawn under the same hard-truncated Uniform prior that the NPE assumes, rather than the truncated log-normal prior used in Section V; if the Tailed-Uniform advantage over the Uniform baseline disappears, the reported gain is confounded by prior mismatch.","tokens_in":12333,"feed_emoji":"🎯","tokens_out":5064,"duration_ms":47241,"temperature":0.7,"pith_summary":"This paper argues that the sharp edges of the Uniform box used to generate training simulations are the root cause of poor neural posterior estimates near parameter-space boundaries, and that a simple change to the sampling proposal fixes it. The proposed Tailed-Uniform proposal keeps a flat core over the original box but pads each side with a decaying tail, so the simulator spends part of its budget on points just outside the region of interest. The payoff, according to the paper, is posterior accuracy near and beyond the boundary that holds even when the assumed prior is itself Uniform, and the benefit grows with dimension because boundaries occupy a larger share of volume. The claim matters because large simulation campaigns are expensive: reallocating a fraction of an existing budget into tail samples would be far cheaper than expanding the box or generating more data.","feed_headline":"Tailed training data fixes neural posteriors at the box edge","feed_subtitle":"Padding the Uniform training box with decaying tails keeps posterior estimates accurate near and beyond the boundary, without extra…","key_machinery":"Tailed-Uniform is a family of hybrid proposal distributions that replaces the hard-truncated proposal $\\tilde{P} = U([a,b]^d)$ with a product of one-dimensional marginals, each flat with density $B/W$ on the core $[a,b]$ and decaying beyond it according to a unit-normalized tail profile $h$. Four profiles are compared: Gaussian, exponential, linear, and flat-tailed, with a shared half-normal scale $\\sigma$ that sets the tail probability mass; as $\\sigma \\to 0$, Tailed-Uniform recovers the Uniform baseline. The machinery works by concentrating samples in the shell where the density estimator previously had no support, at a cost quantified by the cube probability $P_{\\text{cube}} = B(\\sigma)^d$, which falls exponentially with dimension and drives the crossover dimension beyond which most of the budget lands in the tails.","core_discovery":"The central claim is that boundary degradation of NPE-trained posteriors is a structural property of the proposal's support, not a data-volume or capacity problem. Tailed-Uniform-trained networks match reference posteriors better near and beyond the training-box edge, even under a hard-truncated Uniform assumed prior; Uniform-trained networks leak probability mass and under-constrain parameters there. In the cosmological matter-power-spectrum test, all four tail variants outperform the Uniform baseline at boundary points, and the advantage persists across tail-width, budget, and architecture sweeps until very high dimensions. The paper also shows that increasing simulation count by two orders of magnitude or network width eightfold does not heal the Uniform boundary pathology, because no amount of interior data supplies information about the region beyond the edge.","pith_inferences":["If the mechanism is simply that the density estimator needs gradient information across the support edge, the same tail-padding should help neural likelihood and neural ratio estimators, not just NPE.","A testable extension is to allocate tail samples adaptively, starting with a uniform core and adding points only where a validation metric degrades near the boundary, preserving interior coverage at a fixed budget.","Per-dimension tail widths could be tuned along posterior degeneracy directions rather than independently along parameter axes, using the rotated-coordinate idea the paper lists as future work.","The boundary benefit may be strongest for compactly supported assumed priors whose probability mass sits near the truncation, where sharp density transitions are worst."],"forward_implications":["Public simulation suites that currently populate a hard-truncated Latin hypercube could reallocate part of their fixed budget into tail samples and expect better boundary posteriors without retraining.","Increasing simulation count or network width alone does not fix boundary degradation; changing the proposal's support is the effective lever.","In four dimensions, tail widths that put more than half the samples outside the box still improve boundary accuracy, and only near dimension twelve does the advantage disappear.","Smooth tails such as Gaussian, exponential, and linear are preferable to a widened uniform box at large tail width, because the smooth density transition itself matters.","The importance-weight correction in the NPE loss allows the proposal to differ from the assumed prior, so tail padding is directly compatible with a Uniform assumed prior."],"supporting_citations":[{"why":"Defines simulation-based inference and frames the inverse-problem setting the paper addresses.","marker":"[5]"},{"why":"Supplies the NPE training objective with the importance weight $P(\\theta)/\\tilde{P}(\\theta)$ that lets the proposal differ from the assumed prior.","marker":"[9]"},{"why":"Cited as the reason sharp proposal boundaries are difficult for gradient-based density estimators, motivating the tail padding.","marker":"[17]"},{"why":"Supplies the near-boundary volume growth argument used to explain why the benefit grows with dimension.","marker":"[18]"},{"why":"Provides the C2ST metric used to compare posterior samples at every evaluation point.","marker":"[19]"},{"why":"Provides the benchmark tasks and the comparison protocol that supports the boundary-focused evaluation.","marker":"[20]"},{"why":"Emulator used to generate the cosmological matter power spectra for the application.","marker":"[25]"}],"fun_headline_variants":["Tailed-Uniform training fixes posterior boundaries in SBI","Decaying tails on training box keep posteriors accurate at edges","Beyond the box: tailed proposals improve neural posteriors","Robust SBI with Tailed-Uniform: beating the boundary problem"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The cosmological comparison assumes the Markov-chain reference and the neural networks are answering the same question, but they are trained under different priors, so a mismatch in the reference prior could masquerade as a boundary effect.","fun_headline_variants_meta":{"raw":{"variants":["Tailed-Uniform training fixes posterior boundaries in SBI","Decaying tails on training box keep posteriors accurate at edges","Beyond the box: tailed proposals improve neural posteriors","Robust SBI with Tailed-Uniform: beating the boundary problem"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000175,"raw_usage":{"total_tokens":1237,"prompt_tokens":845,"completion_tokens":392,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":461,"completion_tokens_details":{"reasoning_tokens":321}},"tokens_in":461,"tokens_out":392,"duration_ms":3749,"temperature":1.0,"reasoning_tokens":321,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:07:38.915415+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the cosmological benchmark with MCMC reference posteriors drawn under the same hard-truncated Uniform prior that the NPE assumes, rather than the truncated log-normal prior used in Section V; if the Tailed-Uniform advantage over the Uniform baseline disappears, the reported gain is confounded by prior mismatch.","supporting_citations":[{"cited_title":"Cranmer, J","cited_arxiv_id":null,"evidence_quote":"Defines simulation-based inference and frames the inverse-problem setting the paper addresses."},{"cited_title":"Cornish, A","cited_arxiv_id":null,"evidence_quote":"Cited as the reason sharp proposal boundaries are difficult for gradient-based density estimators, motivating the tail padding."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Emulator used to generate the cosmological matter power spectra for the application."}],"review_version":1}