{"id":"c13f8dad-986e-482d-b90f-b49f82d176f7","arxiv_id":"2507.18106","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A new loss and uncertainty map built from a Beta posterior network aim to improve out-of-distribution pixel detection for semantic segmentation, but the reported gains are inconsistent and the novelty is limited.","lead":"This paper proposes a way to detect out-of-distribution objects in street scenes by combining a Beta-distribution uncertainty model with a free-energy-based segmentation framework. It claims better anomaly detection on several road-scene benchmarks, but key baselines are missing and results are mixed.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 12's OoD variance loss is identically zero under its own definitions, so the claimed supervision mechanism cannot operate and Eq. 13's gradient is wrong.","rationale":"The paper's central claim is that replacing RPL's fixed free-energy target with a Beta-posterior variance map improves OoD segmentation. The load-bearing component is the loss that injects this variance into RPL. Eq. 12 as written is identically zero: because m(ω)∈{0,1} and Var>0, −m·Var ≤0, and max(·,0)=0. Eq. 13's gradient is also incorrect; the gradient of max(−mVar,0) is zero in the relevant domain, not −∂Var/∂α − ∂Var/∂β. Consequently, the proposed mechanism cannot provide the 'dynamic adaptation' or 'gradient updates' claimed in Sec. 3.1. This is decisive: whether or not Beta variance is a meaningful uncertainty signal, the architecture's central supervision term has no effect. The reader's weakest_assumption concerned whether Beta variance reliably reflects OoD and whether Eq. 5 omits a log-determinant; those are legitimate but secondary. My check is analytic: evaluating the loss on one positive-variance OoD pixel suffices to show it is zero, and the gradient check confirms no signal flows. A reviewer can also ask for the exact training code to see whether a different sign convention was silently used; if so, the paper's stated objective is not what was implemented. This does not change the REJECT verdict, but it replaces a 'risk' with a concrete internal contradiction.","tokens_in":7534,"tokens_out":4187,"duration_ms":41525,"concrete_test":"Take any OoD pixel, e.g. α=β=2 (Var=1/20), and evaluate Eq. 12: max(−1·0.05,0)=0. Compute ∂Lout/∂α exactly; it is 0 for every α,β>0. A stronger empirical check: retrain with Eq. 12 removed from L_BUCE; if results are unchanged, Eq. 12 is inert. Even a code-level inspection would settle whether the implementation actually uses m=+1 with a sign flip that makes Lout nonzero; if it does, the paper's written objective is not the implemented one.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 defines the core mechanism as Lout = Σ_{ω∈Ω} max(−m(ω)·Var(p(ω)), 0), with m(ω)∈{0,1} a binary OoD mask (Eqs. 11–12). Since Var>0 for α,β>0, every term is max(0, negative)=0. Concretely, for an OoD pixel (m=1), Lout=0; for an inlier (m=0), Lout=0. The gradient displayed in Eq. 13 (∇ Lout = −∂Var/∂α − ∂Var/∂β) is the derivative of −Var, not of max(−Var,0), which is 0 on the whole domain. Thus the proposed variance-based supervision injects no gradient into the flow parameters, so the paper's central mechanism — replacing RPL's fixed free-energy target with the Beta variance map — cannot be what drives the reported gains. Any improvement would have to come from other loss terms (UCE/variance BCE), not from Eq. 12. This is an internal inconsistency, independent of whether Beta variance is a good OoD signal.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Free-Energy Posterior Network that combines a Beta-posterior flow module with the RPL framework for out-of-distribution detection in semantic segmentation. It claims that using the variance of the Beta posterior as a supervisory signal, instead of RPL's fixed free-energy targets, improves uncertainty estimation and OoD detection. Experiments are reported on Fishyscapes, RoadAnomaly, and Segment-Me-If-You-Can.","tokens_in":1724,"tokens_out":2890,"duration_ms":82640,"significance":"The topic is relevant and the paper attempts a novel combination of posterior-network uncertainty and energy-based flow models. However, the central mechanism as written is non-functional: the proposed OoD variance loss in Eq. (12) is identically zero, and the displayed gradient in Eq. (13) is inconsistent with that loss. The empirical section also misreports a table and omits the most relevant baseline. As a result, the claimed contribution is not established by the evidence presented.","major_comments":[{"comment":"Eq. (12) defines Lout = sum_omega max(-m(omega)*Var(p(omega)), 0), with m(omega) in {0,1} and Var(p(omega)) > 0 for alpha,beta > 0. Every summand is therefore zero: for an OoD pixel the argument is negative and max(.,0) returns 0, while for an inlier pixel the argument is 0. Thus Lout is identically zero and injects no gradient into the flow parameters. The gradient in Eq. (13) is the derivative of -Var, not of max(-m*Var,0), whose gradient is zero on the whole domain. This invalidates the paper's central claim that the Beta variance map replaces RPL's fixed free-energy target as a supervisory signal; any observed difference between RPL and RPL+DE would have to be attributed to other terms, such as Eq. (18), not to Eq. (12).","section":"§3.1, Eqs. (11-13)"},{"comment":"The only operational supervision of the Beta variance is the binary cross-entropy loss Lvar = BCE(Var(i), y(i)), which directly fits the variance to binary OoD labels. The paper claims that the method estimates distributional uncertainty without relying on handcrafted thresholds, but Eq. (18) uses OoD labels as supervision. Consequently, reporting improved OoD detection does not establish that Beta variance captures intrinsic distributional uncertainty; it may simply reflect that the estimator is trained to predict the target labels. An ablation omitting Lvar or a held-out analysis would be needed to support the interpretation.","section":"§3.2, Eq. (18); §4.2"},{"comment":"The text in Section 4.2 misreports Table 3. For SMIYC-Anomaly, Table 3 shows RPL+DE with FPR 6.87 and AUPRC 89.48 versus RPL with FPR 7.18 and AUPRC 88.55, not AP 89.48 vs. 57.99. For SMIYC-Obstacle, Table 3 shows RPL+DE with FPR 0.11 and AUPRC 96.29 versus RPL with FPR 0.09 and AUPRC 96.91, not AUPRC 96.29 vs. 22.10. The numbers 57.99 and 22.10 do not appear in the table. Moreover, the Obstacle row shows that RPL+DE is worse than RPL on both FPR and AUPRC, contradicting the claim that RPL+DE consistently improves both AP and AUPRC.","section":"§4.2, Table 3"},{"comment":"Table 1 omits the RPL baseline even though the paper's contribution is an integration with RPL and Section 4.2 claims that the method outperforms all baselines. Without RPL numbers in Table 1, the comparison is incomplete, especially because Table 3 indicates that the gain over RPL is small or negative on some benchmarks. The claim should be restricted to the baselines actually listed in the table.","section":"§4.1, Table 1; §4.2"},{"comment":"Eq. (5) is presented as the log-likelihood in a flow-based framework, but it appears to be only the log-density of the latent Gaussian. A normalizing-flow density for the input requires the log-absolute-determinant of the Jacobian of the transformation. If Eq. (5) is intended only as the latent density, the text should say so explicitly. Since the free-energy interpretation of the flow output is central to the proposed framework, this needs clarification.","section":"§2.2, Eq. (5)"}],"minor_comments":[{"comment":"The definitions of alpha and beta are inconsistent: Eq. (7) uses log(1 + exp(z_in)) and log(1 + exp(z_out)), while Eq. (10) uses log P(z|c) and log Q(z|c). The relationship between these notations should be made explicit.","section":"§3.1, Eqs. (7) and (10)"},{"comment":"The threshold tau in Eq. (9) is not specified; if it is a hyperparameter, its value and sensitivity should be reported.","section":"§3.1, Eq. (9)"},{"comment":"The variable y_i in Eq. (15) is used as a class indicator, but its exact meaning relative to Eq. (9) is unclear for the multi-class segmentation setting. Please clarify whether it is a pixel-wise binary OoD indicator or a class label.","section":"§3.2, Eq. (15)"},{"comment":"The caption says that the differential entropy map replaces the free energy term in RPL's l_in loss, whereas Section 3.1 defines L_out for the OoD term. The notation should be reconciled.","section":"Figure 2 caption"},{"comment":"Section 4.3 refers to the Fishyscapes dataset, but the Figure 3 caption says Road Anomaly. Please correct the caption or the text.","section":"§4.3 and Figure 3"}],"recommendation":"reject","confidential_remarks":"The load-bearing flaw in Eq. (12) is straightforward and internal: the proposed OoD variance loss is identically zero, so the central supervision mechanism cannot operate. Combined with the misreported Table 3 results and the missing RPL baseline in Table 1, the manuscript does not support its claims. A corrected loss and fully rerun experiments would be needed before a new submission could be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: the headline contribution is a no-op. In Sec. 3.1, Eq. 12 defines Lout = Σ max(−m(ω)·Var(p(ω)), 0). Since m is 0/1 and Var is always positive, every term is max(negative, 0) = 0. The gradient in Eq. 13 (∇ = −∂Var/∂α − ∂Var/∂β) is the derivative of −Var, not of the max expression, which is 0 on the whole domain. So the variance map injects no gradient into the flow network. Whatever improvement the experiments show has to come from the other loss terms (LUCE or Lvar), not from the mechanism the paper is built around.\n\nWhat's actually new is narrow: a flow-based Beta posterior network, trained with a UCE-style loss plus a BCE variance term, plugged into RPL for semantic segmentation. The components are all prior work, but the assembly is new in a limited sense. The motivation — that RPL's fixed energy targets are too rigid — is reasonable, and the paper applies the idea to the right benchmarks.\n\nThe soft spots are severe. Eq. 5 is not the log-likelihood of a Gaussian under a normalizing flow; the log-determinant term is missing and the diagonal term should be log diag(U), with a normalization constant. So the free-energy interpretation is on shaky ground. Table 1 omits RPL, the method's own foundation, while the text claims to outperform all baselines. Table 3 shows RPL is actually better on Fishyscapes-Static FPR (0.85 vs 1.03) and on SMIYC-Obstacle (FPR 0.09→0.11, AUPRC 96.91→96.29). The text then misquotes numbers (e.g., 89.48 vs 57.99 and 96.29 vs 22.10) that don't match the table. There is no code and no hyperparameter values (λ1, λ2, τ, N), so the experiments can't be reproduced. Finally, the variance used in Eq. 18 is trained with BCE against the OoD labels, so calling it an unsupervised distributional uncertainty estimate is overstated.\n\nBottom line: this paper is not ready for a serious referee. The central loss is mathematically inert, and the experimental reporting is inconsistent. The authors could fix the loss, correct the flow equation, add the missing baseline, and release code — then there might be a modest, useful paper. As written, I would desk-reject and ask for a careful revision with those changes.","headline":"The paper's central mechanism — the variance-based loss that supposedly replaces RPL's free-energy target — is identically zero by Eq. 12, so the claimed contribution is not supported by the paper's own math.","tokens_in":8337,"tokens_out":4413,"would_cite":false,"duration_ms":44804,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that replacing RPL's fixed free-energy target with a variance map from a Beta posterior network improves out-of-distribution detection in semantic segmentation.","keywords":["out-of-distribution detection","semantic segmentation","uncertainty estimation","Beta posterior network","normalizing flows","free energy","residual pattern learning","evidential deep learning"],"falsifier":"Run the same RPL-plus-flow pipeline but replace the Beta variance in the loss with the raw free-energy score while keeping the identical binary OoD-mask supervision; if FPR and AuPRC do not worsen, the Beta posterior variance is not the operative component. A second check is to recompute the flow log-likelihood with the full change-of-variables Jacobian determinant and see whether the free-energy ranking of OoD pixels in Eq. (5) is preserved.","tokens_in":7332,"feed_emoji":"🎯","tokens_out":8251,"duration_ms":74229,"temperature":0.7,"pith_summary":"The paper proposes a Free-Energy Posterior Network for semantic segmentation: a frozen backbone is paired with a flow-based density estimator that outputs per-pixel parameters of a Beta distribution, and the variance of that Beta distribution is used as a training signal for out-of-distribution (OoD) regions. The claim is that this variance map is a better supervisory target than the fixed free-energy scores used by the residual pattern learning (RPL) framework, because it lets the network learn to increase uncertainty near ambiguous or unseen regions instead of relying on post-hoc energy thresholding. If the claim holds, OoD-aware segmentation can be trained end-to-end in a single forward pass, with no stochastic sampling and no hand-tuned thresholds. The paper reports the best FPR and AuPRC across most of the Fishyscapes, RoadAnomaly, and Segment-Me-If-You-Can benchmarks.","feed_headline":"Beta variance map lifts out-of-distribution detection in segmentation","feed_subtitle":"A flow-based Beta posterior trains pixel-level OoD awareness in one pass, no sampling or hand-tuned thresholds.","key_machinery":"The operative object is the Beta posterior variance map. A flow-based network transforms per-pixel latent features into log-probability scores $z_{\\mathrm{in}}, z_{\\mathrm{out}}$, which are converted to Beta parameters by $\\alpha = 1 + \\log(1 + e^{z_{\\mathrm{in}}})$ and $\\beta = 1 + \\log(1 + e^{z_{\\mathrm{out}}})$; the variance $\\mathrm{Var}(p) = \\alpha\\beta/((\\alpha+\\beta)^2(\\alpha+\\beta+1))$ then becomes both the uncertainty estimate and the training target. The key property is that variance is largest when $\\alpha \\approx \\beta$ with small values, so minimizing the outlier loss corresponds to reducing the sharpness of the predictive distribution on OoD pixels. The paper integrates this variance into RPL's energy loss and adds a Beta Uncertainty Cross Entropy (BUCE) loss that combines standard cross-entropy, a digamma-based uncertainty term, and binary cross-entropy between the variance map and the OoD mask.","core_discovery":"The paper's central claim is that distributional uncertainty, quantified as the variance of a per-pixel Beta posterior fitted by a normalizing-flow density estimator, can serve as a principled supervision signal for OoD detection in semantic segmentation. The method computes flow-based energy representations, converts them into Beta parameters $\\alpha(\\omega) = 1 + \\log(1 + e^{z_{\\mathrm{in}}(\\omega)})$ and $\\beta(\\omega) = 1 + \\log(1 + e^{z_{\\mathrm{out}}(\\omega)})$, and injects the variance $\\mathrm{Var}(p(\\omega)) = \\frac{\\alpha(\\omega)\\beta(\\omega)}{(\\alpha(\\omega)+\\beta(\\omega))^2(\\alpha(\\omega)+\\beta(\\omega)+1)}$ into RPL's energy-based loss. Because Beta variance is maximized when $\\alpha \\approx \\beta$ and both are small, the loss pushes the flow network to flatten the predictive distribution exactly where the OoD mask is active, while the proposed BUCE loss keeps inlier predictions confident. The claimed result is a semantically meaningful, sampling-free uncertainty map that outperforms MC Dropout, Shannon entropy, and plain free-energy baselines on real-world anomaly-segmentation benchmarks.","pith_inferences":["The same Beta-variance supervision could be plugged into other energy-based OoD detectors beyond RPL; the paper's closing remark gestures at this, but the mechanism itself is not specific to RPL.","Part of the reported gain may come from the binary cross-entropy term that directly supervises variance with OoD labels, so an ablation that trains the same BCE on raw energy scores would reveal how much of the improvement is due to the Beta parametrization rather than to outlier exposure.","If Beta variance is a genuine distributional-uncertainty signal, it should also improve misclassification detection and calibration on in-distribution inputs; the paper does not report those checks, but they are testable extensions.","The condition $\\alpha \\approx \\beta$ with small values is the evidential reading of 'conflicting and weak evidence,' which suggests the method could be unified with evidential deep-learning losses that penalize high uncertainty without explicit OoD labels."],"forward_implications":["OoD-aware segmentation can be trained with a single forward pass, eliminating the computational overhead of MC Dropout-style sampling and the need for post-hoc energy thresholding.","The Beta variance map provides dense pixel-level supervision, so the network can learn OoD regions during training rather than only at inference, which is what the reported FPR and AuPRC gains reflect.","The same uncertainty signal transfers across different anomaly types, since the paper reports consistent gains on Fishyscapes-Static, Fishyscapes-L&F, SMIYC-Anomaly, SMIYC-Obstacle, and RoadAnomaly.","Because the segmentation backbone is frozen and only the flow posterior and loss terms are trained, the added cost is localized to the uncertainty branch, making the approach more scalable than sampling-based alternatives.","The variance-maximizing condition ties OoD behavior to a concrete property of the Beta distribution, giving practitioners a diagnostic: anomalous pixels should be those where the posterior is flat and low-evidence."],"supporting_citations":[{"why":"RPL, the baseline framework whose fixed free-energy targets are replaced; it supplies the residual prediction branch and training protocol.","marker":"[13]"},{"why":"FlowEneDet, source of the flow-based density and free-energy estimation used to compute the Beta parameters.","marker":"[8]"},{"why":"Posterior Network, origin of the Dirichlet/Beta posterior formulation and the density-based pseudo-count idea.","marker":"[4]"},{"why":"UCE, the uncertainty-aware cross-entropy loss that the BUCE term extends to a Beta posterior.","marker":"[10]"},{"why":"Fishyscapes benchmark, providing the Static and LostAndFound evaluation sets.","marker":"[2]"},{"why":"Segment-Me-If-You-Can benchmark, providing the Anomaly and Obstacle tracks.","marker":"[3]"},{"why":"RoadAnomaly benchmark, providing the road-anomaly evaluation set.","marker":"[12]"}],"fun_headline_variants":["Free-energy posterior network uses Beta variance for OoD","Beta variance from free energy flags out-of-distribution pixels","One-pass OoD segmentation with Beta posterior variance","Distributional uncertainty as Beta variance in posterior nets","Sampling-free uncertainty: Beta variance from flow-based energy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the variance of the Beta posterior, trained with a binary cross-entropy loss against OoD labels, truly captures distributional uncertainty for unseen pixels and is a better training target than the original fixed free-energy score; if the variance is merely fitting the outlier labels, the claimed mechanism collapses.","fun_headline_variants_meta":{"raw":{"variants":["Free-energy posterior network uses Beta variance for OoD","Beta variance from free energy flags out-of-distribution pixels","One-pass OoD segmentation with Beta posterior variance","Distributional uncertainty as Beta variance in posterior nets","Sampling-free uncertainty: Beta variance from flow-based energy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000594,"raw_usage":{"total_tokens":2821,"prompt_tokens":1020,"completion_tokens":1801,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":636,"completion_tokens_details":{"reasoning_tokens":1723}},"tokens_in":636,"tokens_out":1801,"duration_ms":13892,"temperature":1.0,"reasoning_tokens":1723,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:37:50.912505+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same RPL-plus-flow pipeline but replace the Beta variance in the loss with the raw free-energy score while keeping the identical binary OoD-mask supervision; if FPR and AuPRC do not worsen, the Beta posterior variance is not the operative component. A second check is to recompute the flow log-likelihood with the full change-of-variables Jacobian determinant and see whether the free-energy ranking of OoD pixels in Eq. (5) is preserved.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RPL, the baseline framework whose fixed free-energy targets are replaced; it supplies the residual prediction branch and training protocol."},{"cited_title":"Gudovskiy, T","cited_arxiv_id":null,"evidence_quote":"FlowEneDet, source of the flow-based density and free-energy estimation used to compute the Beta parameters."},{"cited_title":"Charpentier, D","cited_arxiv_id":null,"evidence_quote":"Posterior Network, origin of the Dirichlet/Beta posterior formulation and the density-based pseudo-count idea."},{"cited_title":"The Fishyscapes Benchmark: Measuring Blind Spots in Semantic Segmentation","cited_arxiv_id":"1904.03215","evidence_quote":"Fishyscapes benchmark, providing the Static and LostAndFound evaluation sets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Segment-Me-If-You-Can benchmark, providing the Anomaly and Obstacle tracks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RoadAnomaly benchmark, providing the road-anomaly evaluation set."}],"review_version":1}