{"id":"78cea556-54a5-4506-9b06-fd9a10a21895","arxiv_id":"2512.04954","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Training a normalizing flow on prior samples weighted by likelihood can approximate a posterior, but matching the base distribution's number of modes to the target is needed to avoid spurious bridges.","lead":"The paper trains a normalizing flow to approximate a Bayesian posterior by drawing samples from the prior and weighting them by their likelihood, avoiding posterior samples. It shows on 2D/3D toy tasks that a Gaussian-mixture starting distribution reproduces multi-modal posteriors better than a unimodal base, though a heavy-flavour application promised in the abstract is missing from the text.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Empirical support for the topology-alignment claim is single-run and internally non-monotone; the paper's own optimization caveats make run-to-run variation a live confound.","rationale":"I read the paper as claiming two things: (i) the likelihood-weighted objective (Eq. 2.11) is a valid surrogate for D_KL(posterior||q), and (ii) the method works in practice, especially when base and target mode counts match. The first is mathematically sound; no concern there. The second is the load-bearing empirical claim and is where the evidence is weakest. The paper's own caveats about optimization instability and combinatorial mode assignment (Sec. 5, footnote 1) are not backed by any diagnostic that lets the reader distinguish a genuine topological effect from a run that got stuck. The non-monotone entries in Tables 2 and 3 make this worry concrete: if a two-mode base is meant to be topologically intermediate between one and three modes, its metrics should roughly fall between them, but they are the worst in both tables. Without repeated seeds or loss curves, the favorable three-mode result could simply be the one run that found a good assignment. Additionally, the abstract promises a high-energy physics application and MCMC comparison, but the body contains no such section; this removes the one independent, realistic validation that would have made the practical claim credible. I agree with the reader's weakest assumption; my concern is the same confound, sharpened by the non-monotone numbers. The proposed re-run would settle it. Since the central contribution is empirical and that evidence is missing, I support the reader's REJECT verdict.","tokens_in":7767,"tokens_out":7732,"duration_ms":70739,"concrete_test":"Run every configuration in Tables 2 and 3 (2D/3D/non-Gaussian, base modes 1, 2, 3) with 20 random seeds under the same architecture, recording final loss curves. Compute mean and 95% CI for D_KL and W1, and regard the topology-alignment claim as unsupported if the matched-mode model is not significantly best or if its loss curves do not consistently reach the same final value as unmatched models. Also track mode assignment: for each base component, compute the mean latent z mapped into each target mode across seeds; if assignments vary by seed, the GMM initialization is not controlling the outcome.","verdict_should_be":"REJECT","load_bearing_attack":"The central empirical claim—Section 4.1/4.2's conclusion that when the number of base modes matches the target, 'we achieve the lowest distance measures and the highest fidelity reconstruction'—rests on Tables 2 and 3. These are single runs with no error bars, no convergence curves, and no evidence that the conditional 'provided the required loss is reached for all models' (Sec. 4.1) is satisfied. The tables are not even monotone in base-mode count: for a three-mode 2D target, Model-2D2 (two base modes) has W1=0.2372, worse than Model-2D1 (one base mode) with W1=0.1352; in the non-Gaussian case Model-nonGauss2 has W1=4.4237, worse than Model-nonGauss1's 1.6173. Given the paper's own statement that 'the network lacks explicit guidance on which base mode should map to which target mode, leading to combinatorial ambiguity and optimization instability' (Sec. 5) and footnote 1 ('reaching an optimum loss while training is not very trivial in higher dimensions'), bad local optima are a plausible explanation for the reported differences. The abstract also promises a heavy-flavour application and MCMC comparison that do not appear in the body, removing the independent realistic validation. The mathematical identity in Eqs. 2.9-2.11 is not the issue; the issue is that the paper's central practical claim is not established by the reported evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a likelihood-weighted normalizing flow objective: by minimizing -1/N sum_i L(theta_i) log q_phi(theta_i) for prior samples, the flow is claimed to converge to the posterior p(theta|D) without posterior samples. Experiments on 2D and 3D Gaussian mixture targets and one non-Gaussian target compare flows with 1, 2, or 3 base modes. The central claim is that when the base mode count matches the target, the model achieves the lowest distance/divergence metrics and avoids spurious bridges. The abstract also announces a heavy-flavour physics application with MCMC comparison, but no such application appears in the body.","tokens_in":8069,"tokens_out":6407,"duration_ms":57668,"significance":"The mathematical derivation in Eqs. (2.9)-(2.11) is correct, and the bridge artifact is a well-motivated and clearly visualized consequence of the diffeomorphic constraint. If the cardinality-matching claim could be established robustly, the method would be a simple, architecture-agnostic contribution to amortized posterior estimation. However, the empirical evidence is currently based on single runs with no error bars or convergence verification, and the experiment design preselects the base mode count from the known ground truth. The absence of the advertised heavy-flavour application further weakens the real-world validation.","major_comments":[{"comment":"The central claim that matching cardinality gives the best reconstruction rests entirely on Tables 2 and 3. These are single runs with no error bars, no repeated-seed statistics, and no training curves. The text conditions the claim on 'provided the required loss is reached for all models' (Sec. 4.1), but no loss values or convergence evidence are reported. The tables are also non-monotone: for the three-mode 2D target, Model-2D2 has W1=0.2372, worse than Model-2D1's 0.1352; for the non-Gaussian target, Model-nonGauss2 has W1=4.4237, worse than Model-nonGauss1's 1.6173. Given the paper's own statement about combinatorial ambiguity and optimization instability (Sec. 5), the reported differences may reflect local optima rather than the topology-alignment effect. Multiple seeds, error bars, and convergence checks are needed to support the claim.","section":"§4.1, Tables 2–3"},{"comment":"The evaluation protocol is circular with respect to the central conclusion: the base mode count is set equal to the known ground-truth mode count, and then the matching model is reported to perform best. This demonstrates that if one already knew the number of target modes, a GMM base with that cardinality yields good results; it does not provide a method or prediction for inferring the mode count. The conclusion itself calls for future 'adaptive methods to formally characterize and match the number of modes', which is a direct admission that the current approach does not address the unknown-mode-count setting. Please state this limitation explicitly in the abstract and Section 4, or add an experiment in which the mode count is estimated without ground-truth knowledge.","section":"§4.1, §5"},{"comment":"The abstract and introduction promise a realistic validation: extraction of the Wolfenstein parameters from the CP asymmetry in B^0 -> J/psi K^0 and comparison against a well-converged MCMC reference. The body contains no such application, no mention of this decay, and no MCMC comparison; the conclusion is the last section and only summarizes the toy benchmarks. This is a substantial mismatch between the claimed contribution and the actual content. Either add the missing experiment or revise the abstract/introduction to state that validation is limited to synthetic benchmarks.","section":"Abstract vs. body"}],"minor_comments":[{"comment":"Implementation details are incomplete: report architecture hyperparameters, training steps, learning rate, batch size, optimizer, and random seeds. Without these, the experiments are not reproducible.","section":"§3.2"},{"comment":"The column header 'W-Sliced KL Divergence' is unclear; it seems to refer to a Sliced-Wasserstein metric, not a KL divergence. Please define each metric and specify the Monte Carlo approximations used.","section":"Tables 2 and 3"},{"comment":"There are numerous typos and formatting issues: 'M odel' spacing, 'probabilitybridges', inconsistent 'modelled/modeled', 'T able'. A copy-editing pass is advised.","section":"Throughout"},{"comment":"Footnote 1 states that reaching the optimum loss is 'not very trivial in higher dimensions' — this is vague. Please report final loss values and specify the dimensions in which this statement is meant to apply.","section":"Footnote 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is relevant to cs.LG and the proposed loss is simple and well-founded. However, the missing advertised heavy-flavour application and the thin empirical evaluation are serious issues. If the authors can add repeated-seed experiments with error bars and convergence evidence, and either include the heavy-flavour study or revise the abstract, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: don't let the title fool you. The core objective in Eq. (2.11) is the same likelihood-weighted negative log-likelihood used by Neural Importance Sampling (Müller et al., ref [23]). The paper even says it is 'akin to Neural Importance Sampling,' and the KL derivation is textbook. What is new here is not the method but the empirical observation about base-mode cardinality and flow topology.\n\nThe genuinely useful part is the demonstration that a unimodal base forces a diffeomorphic flow to draw bridges between disconnected posterior modes, and that a GMM base with matching mode count removes those artifacts. The figures make that point convincingly. The paper also flags the combinatorial ambiguity in mode assignment and the optimization difficulty in higher dimensions (Sec. 5, footnote 1). That is honest and shows good scientific instinct.\n\nThe soft spots are proportionate but load-bearing. Tables 2 and 3 are single runs with no error bars, no convergence curves, and no baseline against NIS, MCMC, or even a second seed. The results are not monotone in mode count: Model-2D2 is worse than Model-2D1 on Wasserstein. That non-monotonicity is exactly what you would expect from bad local optima, which the paper itself lists as a known issue. The 'alignment works' conclusion also uses the known ground-truth mode count to choose the base; that is a diagnostic, not a prediction. On top of that, the abstract promises a heavy-flavour application and MCMC comparison that never appear in the body, and the 'amortized' framing is misleading—the flow is trained for one dataset and is not conditioned on observations.\n\nThe non-Gaussian benchmark in Sec. 4.2 is also underspecified: we see pictures and numbers, but no explicit density. So the whole empirical contribution currently rests on under-reported experiments.\n\nWho is this for? Someone interested in topology constraints in normalizing flows will find the bridge demonstration worth a look. But as a methods paper, it is not ready. I would want repeated-seed results, proper baselines, the missing physics application, and an honest framing that this is an application of neural importance sampling, not a new algorithm. With that revision it could become a solid short paper. As is, a serious referee would reject on evidence and novelty grounds, but I would not desk-reject outright—the derivation is clean and the topology observation is worth referee attention.","headline":"The math is correct and the bridge artifact is real, but the method is neural importance sampling with a new name, and the key empirical claim rests on single runs and a missing application section.","tokens_in":8598,"tokens_out":2541,"would_cite":false,"duration_ms":24970,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A normalizing flow trained with likelihood-weighted prior samples can reconstruct multi-modal posteriors without posterior samples, but only if its base distribution is initialized with as many modes as the target has.","keywords":["normalizing flows","posterior estimation","likelihood-weighted importance sampling","multi-modal distributions","topological mismatch","Gaussian mixture base","amortized inference","inverse problems"],"falsifier":"Re-run the matched-cardinality likelihood-weighted training on a four- or five-mode posterior in 3-5 dimensions with many random seeds and a single fixed architecture; if most seeds produce collapsed modes, bridged modes, or stalled losses, the alignment result is specific to the paper's low-mode runs rather than a general property of the objective.","tokens_in":7584,"feed_emoji":"🌊","tokens_out":6314,"duration_ms":51492,"temperature":0.7,"pith_summary":"This paper tries to show that a normalizing flow can learn a Bayesian posterior directly from prior samples and likelihood evaluations, with no posterior samples or MCMC chains. The key twist is that the flow's base distribution must have the same number of modes as the posterior; otherwise the flow is forced by topology to draw spurious bridges of probability between modes. The paper derives a likelihood-weighted loss that approximates KL divergence from the posterior to the model, then demonstrates on 2D and 3D benchmarks that matching the base mode count to the target yields the best Wasserstein and KL metrics. It also applies the method to a multimodal, non-Gaussian particle-physics posterior and finds agreement with a converged MCMC reference. If it holds, the result turns flow topology from a detail into the main design choice for amortized inference.","feed_headline":"Match base-mode count to target modes and the flow learns them all","feed_subtitle":"Likelihood-weighted flows learn posteriors from prior samples alone, if their base has every target mode","key_machinery":"The central object is the likelihood-weighted loss function, which weights each prior sample's log-density contribution by its data likelihood. The load-bearing mechanism is the diffeomorphic nature of normalizing flows: because the map is continuous and invertible, the base density's connectivity is preserved in the modeled density. The paper exploits this by choosing a Gaussian Mixture base with the same number of components as the target modes, so each base component can be assigned to one posterior mode.","core_discovery":"On the paper's own terms, the central claim is that the likelihood-weighted objective L(phi) = -(1/N) sum_i p(D|theta_i) log q_phi(theta_i) turns a normalizing flow into a posterior estimator that needs no posterior samples. The supporting empirical discovery is topological: a flow is a diffeomorphism, so it cannot change the connectivity of its base distribution; a unimodal base forced onto a multimodal posterior leaks probability through 'bridges' between modes. Initializing the base as a Gaussian mixture whose number of components equals the posterior's mode count removes those bridges and gives the lowest Wasserstein and KL values, including for a non-Gaussian 3D posterior and a physics-","pith_inferences":["Inference: The practical value likely depends on knowing or estimating the posterior's mode count ahead of time; an adaptive procedure that grows the base mixture until artifacts disappear is a natural extension the paper leaves implicit.","Inference: The combinatorial ambiguity in base-to-target mode assignment suggests that matched cardinality alone may not scale to high dimensions; adding a permutation-invariant or assignment-aware regularizer could stabilize training.","Inference: The paper's observation that KL stays low while Wasserstein rises under topology mismatch suggests that evaluation of multimodal posterior approximations should report spatial distances, not only KL.","Inference: A direct test would be comparing matched-mode flows against variational flows and annealed importance sampling on the same benchmarks to see when likelihood-weighted training is the cheaper route."],"forward_implications":["Posterior estimation becomes possible in likelihood-accessible problems using only prior sampling and likelihood calls, removing the need for precomputed posterior samples or long MCMC runs.","Base distribution topology is a design parameter, not a neutral choice; mismatched mode counts produce measurable artifacts (bridges) and inflated Wasserstein distances.","Matched mode counts should improve reconstruction fidelity in amortized settings across dimensions, as shown in 2D and 3D Gaussian-mixture and non-Gaussian benchmarks.","The same likelihood-weighted training can be applied to physics parameter extraction, such as Wolfenstein parameters from B0 -> J/psi K0 CP asymmetry, with results comparable to MCMC.","Loss values alone are not enough to judge multimodal fits; distance and divergence metrics capture topology errors that the loss itself barely reflects."],"fun_headline_variants":["Flow base mode count must equal target modes to avoid bridges","Likelihood-weighted flows need no posterior samples if base is multimodal","Fix spurious posterior bridges by matching base mixture to target modes","No posterior training data: set base topology to match target and go"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The empirical case rests on the trained flow actually reaching a good optimum and mapping each base mode to the correct posterior mode; the paper itself warns that this assignment is unguided and that reaching an optimal loss is not trivial in higher dimensions.","fun_headline_variants_meta":{"raw":{"variants":["Flow base mode count must equal target modes to avoid bridges","Likelihood-weighted flows need no posterior samples if base is multimodal","Fix spurious posterior bridges by matching base mixture to target modes","No posterior training data: set base topology to match target and go"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000273,"raw_usage":{"total_tokens":1469,"prompt_tokens":737,"completion_tokens":732,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":670}},"tokens_in":481,"tokens_out":732,"duration_ms":6687,"temperature":1.0,"reasoning_tokens":670,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T06:38:25.884114+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the matched-cardinality likelihood-weighted training on a four- or five-mode posterior in 3-5 dimensions with many random seeds and a single fixed architecture; if most seeds produce collapsed modes, bridged modes, or stalled losses, the alignment result is specific to the paper's low-mode runs rather than a general property of the objective.","supporting_citations":[],"review_version":1}