{"id":"fefa30c9-323b-4c1e-a17a-76583224e460","arxiv_id":"2607.25376","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Student's t likelihood (ν=5) is a robust default for VI-trained BNNs, improving CRPS in most tested settings while occasionally losing on MSE to Gaussian under lognormal noise.","lead":"The paper tested three ways to model prediction uncertainty inside Bayesian neural networks and found that a heavy-tailed Student's t assumption usually beats the usual Gaussian assumption. That gives practitioners a nearly free change to the training objective that can improve predictive accuracy and uncertainty quality on regression tasks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Moment-matched likelihood is not the ELBO likelihood of Eq. (3.4); the reported Student-t advantage may be an artifact of this parameterization.","rationale":"The reader's weakest assumption—that the moment-matched estimator may be an artifact rather than a faithful likelihood—is the same concern I consider most load-bearing. I sharpen it: the objective used in training is not the ELBO of Eq. (3.4), because the likelihood depends on multiple weight samples, not on a single W. This threatens both internal consistency and external applicability. The abstract's overclaim ('independent of the data distribution') is real but is a scope/wording issue that the paper's own conclusion already softens; the moment-matching issue is more fundamental because it determines whether the reported ranking is about likelihood families at all. The reader's conditional verdict remains appropriate: the claim should be accepted, if at all, only after the proposed per-sample ELBO comparison is run. No adjustment to the reader's verdict is needed.","tokens_in":29656,"tokens_out":7673,"duration_ms":85374,"concrete_test":"Re-run Table 3 (multivariate, Gaussian and lognormal noise) with the conventional per-sample ELBO estimator: for each of T=10 weight draws W_t, compute log p(y | f_{W_t}(x)) using (i) Gaussian with fixed/learned scale and (ii) Student's t with ν=5, then average the log-likelihoods to estimate Eq. (3.4). Keep architectures, seeds, and early stopping identical. If Student's t no longer dominates CRPS/MSE, the moment-matched parameterization is responsible for the reported ranking; if it still dominates, the central claim survives this check.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 defines the VI objective as ELBO = E_{W~q}[log p(Y|X,W)] (Eq. 3.4). Section 4 then parameterizes p(Y|X,W) using sample moments \\hat{\\mu} and \\hat{\\sigma} computed over multiple weight draws (§5.5: 10 samples per forward pass). The density actually evaluated is therefore p(y | \\hat{\\mu}(W_1..W_T), \\hat{\\sigma}(W_1..W_T)), not p(y|X,W) for a single weight draw. This is a moment-matched approximation to the posterior predictive, not the likelihood appearing in the ELBO. The Gaussian/skew-normal/Student-t comparison thus tests three predictive-density approximations under a non-standard objective, and the ranking may not transfer to conventional VI-BNNs where likelihood parameters are direct network outputs or fixed per-sample noise. The paper never bounds the gap between log p(y; \\hat{\\mu},\\hat{\\sigma}) and (1/T)\\sum_t log p(y|f_{W_t}(x)), nor compares against the standard parameterization. Since the practical recommendation is to replace the Gaussian likelihood in VI-BNNs, this unvalidated design choice is the load-bearing assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the choice of likelihood distribution inside variational-inference Bayesian neural networks (VI-BNNs). For regression, it compares Gaussian, skew-normal, and Student's t likelihoods, where the likelihood parameters are computed from empirical moments of Monte Carlo weight samples. Experiments cover univariate (y=x^3+noise), multivariate (20-dimensional input, 5-dimensional output) synthetic data with five noise families, and two real-world regression tasks (POWER, ENTSO-E), across various MLP depths/widths and dataset sizes, with 30 seeds. The paper reports that Student's t (with fixed ν=5) is usually the best or competitive likelihood, especially on CRPS, and concludes that Student's t is a robust default; the abstract makes a stronger universal claim of superiority independent of data distribution and architecture.","tokens_in":30011,"tokens_out":6238,"duration_ms":65380,"significance":"If the finding is valid, the practical recommendation to replace the default Gaussian likelihood in VI-BNNs with a location-scale Student's t (ν=5) is useful and easily actionable. The paper's strengths are the breadth of experiments (30 seeds, several architectures, five noise families, two real datasets) and the availability of source code. The statistical significance testing against the Gaussian baseline adds rigor. However, the universal formulation in the abstract is inconsistent with the paper's own tables, and the moment-matched likelihood parameterization is not the standard ELBO likelihood, so the central claim is currently not established for conventional VI-BNN likelihoods. A reframed or extended version could be a solid empirical contribution.","major_comments":[{"comment":"The abstract's claim that Student's t is better 'independent of the data distribution and MLP architecture' is contradicted by the paper's own results. In Table 3, under lognormal noise with σ=3 and n=10,000, Gaussian is the significant best for MSE in all four architectures (G***); at n=50,000 it remains best in the 2×128 and 2×64 cells. In Table 4, Gaussian attains the best MSE with G*** for POWER 8×4 and for ENTSO-E 4×4 and 8×4. Table 2 with n=300 and Gaussian noise also favors Gaussian on MSE. The later text already softens to 'generally more robust'; the abstract and conclusion should be aligned with the actual scope, or the claim needs to be conditioned.","section":"Abstract"},{"comment":"The density actually evaluated is not the ELBO likelihood p(Y|X,W). Eq. (3.4) defines the ELBO with an expectation of log p(Y|X,W), but §4 parameterizes all candidate likelihoods with sample moments \\hat{\\mu} and \\hat{\\sigma} computed from T=10 stochastic forward passes (§5.5). The evaluated density is log p(y|\\hat{\\mu},\\hat{\\sigma}), not (1/T)\\sum_t \\log p(y|f_{W_t}(x)). This is a moment-matched approximation to the posterior predictive, not the likelihood appearing in the VI objective. The paper never compares against the standard parameterization in which likelihood parameters are direct network outputs or a learned/fixed noise variance, nor bounds the approximation gap. The observed Student-t advantage could therefore be an artifact of the moment-matched estimator rather than of the likelihood family. This issue must be addressed, either by ablating T, adding standard-parameterizatio","section":"§4 vs. §3 Eq. (3.4)"},{"comment":"The degrees of freedom ν=5 is selected via a sensitivity analysis performed only on the univariate Gaussian-noise, n=3,000 case (Table 1) and then fixed across all noise families, architectures, and real datasets. The robustness and 'data-independent' wording therefore depend on a hyperparameter tuned on one specific condition. The authors should either report sensitivity of the main conclusions to ν across a broader set of conditions or explicitly state that ν=5 is a selected, fixed hyperparameter and discuss the risk of transfer.","section":"§5.7, Table 1"}],"minor_comments":[{"comment":"The skew-normal parameterization is ambiguous: δ is introduced only as |δ|, but α is then defined as δ/sqrt(1-δ^2), which requires a signed δ. Specify δ = sign(\\hat{\\gamma}) |δ| or the equivalent, otherwise the skew-normal implementation is under-specified.","section":"§4, Eqs. (4.1)–(4.2)"},{"comment":"The caption says rows are grouped by noise type '(Gaussian, Gamma, Student's t)', but the table also contains Lognormal and Laplace rows. Update the caption to list all five noise families.","section":"Table 3 caption"},{"comment":"The figure captions label the model as '32x8' and '32x1', while the surrounding text describes depth 1 width 16 and depth 1 width 32. Please check the labels and keep them consistent.","section":"Figures 6 and 7"},{"comment":"Typo: 'POWER ans ENTSO-E' should be 'and'. Please proofread the final text.","section":"§5.9"},{"comment":"The paper reports significance tests on training epochs and wall-clock time in Table 4. These are computational-cost measures, not predictive quality; it would help to state this explicitly and clarify whether the t-tests on time are adjusted for multiple comparisons or only used descriptively.","section":"§5.6"}],"recommendation":"major_revision","confidential_remarks":"The experimental campaign is impressive, but the mismatch between the advertised 'ELBO likelihood' and the moment-matched estimator is the main substantive risk. If the authors cannot provide a standard-parameterization comparison, the manuscript should be reframed as an empirical study of moment-matched predictive densities. The abstract's universal claim also needs to be corrected to match the tables. This is fixable within the scope of the paper, hence major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a careful, extensive empirical study that answers a real question—whether the Gaussian likelihood default in VI-trained BNNs should be replaced—and its main empirical claim (Student's t with fixed ν=5 is a robust choice, especially for CRPS) is supported by the experiments. But the abstract oversells it: the paper's own tables show Gaussian winning MSE under lognormal noise and in several real-world configurations, so \"independent of data distribution and MLP architecture\" is not accurate.\n\nWhat's new: prior work used Student's t in GPs and linear models; this is a systematic VI-specific comparison across five noise families and many architectures, with 30 seeds and released code. The finding that matching the likelihood family to the true noise does not buy much is genuinely useful. The table quality is good and the statistical testing is appropriate.\n\nSoft spots, in order of importance. First, the moment-matched likelihood parameterization. In Section 4, the likelihood is not the per-weight-sample likelihood in the ELBO of Eq. (3.4); it's a density evaluated at sample moments μ̂ and σ̂ from 10 weight draws. That's a plugin approximation to the posterior predictive, not the standard ELBO likelihood. The paper never compares against the usual parameterization where each weight draw produces its own likelihood, nor bounds the gap. Since the practical recommendation is to swap the likelihood in standard VI-BNNs, this unvalidated design choice is a real concern. It's not fatal—the comparison is fair across the three families within the scheme—but it limits the headline claim to \"under this moment-matched estimator.\"\n\nSecond, ν=5 is tuned on one Gaussian-noise dataset and then fixed everywhere. Minor, but it should be flagged as a hyperparameter, not a free lunch. Third, the abstract overclaim: Table 3 shows Gaussian wins MSE in several lognormal cells, and Table 4 shows Gaussian winning MSE in two real-world cells. The conclusion correctly says \"generally more robust,\" so the abstract should be aligned with that.\n\nThe citation pattern is clean, no self-citation issues, and the work is reproducible.\n\nRecommendation: send to peer review. A serious referee can push on the moment-matching issue and the overclaim, but the empirical contribution is substantial enough to be worth airing. The paper would be much stronger with either a comparison to the standard likelihood parameterization or a clearly narrowed claim.","headline":"Solid empirical study with an oversized abstract: Student's t is a robust default under their moment-matched VI scheme, but not unconditionally better as claimed.","tokens_in":30420,"tokens_out":2797,"would_cite":true,"duration_ms":31070,"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":"A location-scale Student's t likelihood with five degrees of freedom is a more robust default than Gaussian in Bayesian neural networks trained by variational inference, across tested noise families and MLP architectures.","keywords":["Bayesian neural networks","variational inference","likelihood distribution","Student's t","uncertainty quantification","model misspecification","continuous ranked probability score","regression"],"falsifier":"Re-run the comparison with likelihood parameters optimized directly (for example, maximizing validation log-likelihood for each family) instead of moment-matched, keeping ν=5; if Student's t no longer dominates Gaussian, the paper's conclusion is an artifact of moment matching. Alternatively, on a large dataset with clean Gaussian noise, if Gaussian does not at least tie Student's t when both receive the same estimated scale, the claim that heavy tails help regardless of noise is refuted.","tokens_in":29594,"feed_emoji":"📈","tokens_out":4861,"duration_ms":46723,"temperature":0.7,"pith_summary":"The paper tests whether the default Gaussian likelihood in Bayesian neural networks trained by variational inference can be replaced with something better. Across regression experiments on artificial data with known noise (Gaussian, Gamma, Student-t, Laplace, Lognormal) and two real-world datasets, a location-scale Student's t likelihood with five degrees of freedom consistently matches or beats the Gaussian on point error and probabilistic scores, regardless of the true noise family or MLP depth and width. The finding matters because it suggests practitioners do not need to know the noise distribution: a heavy-tailed likelihood is a robust default. The paper also finds that a skew-normal likelihood does not help, so tail behaviour, not skewness, drives the gain.","feed_headline":"Student's t likelihood outperforms Gaussian in Bayesian neural nets","feed_subtitle":"A simple drop-in replacement for Gaussian noise improves both point forecasts and uncertainty intervals.","key_machinery":"The likelihood parameters are set by moment matching from stochastic forward passes: the predictive mean and standard deviation are the sample moments of Monte Carlo weight samples, and the Student's t scale is derived to match that sample variance, with degrees of freedom fixed at five. This turns a heavy-tailed density into a one-line change to the ELBO loss, providing a robust training signal without extra learned parameters. The skew-normal version adds a third-moment shape correction but performs worse, isolating heavy tails as the operative feature.","core_discovery":"The central claim is that the likelihood distribution in mean-field variational-inference BNNs need not match the data noise to be optimal; a Student's t with five degrees of freedom is generally a more robust choice than the Gaussian. The evidence comes from controlled experiments: when the noise is Gaussian and the sample is large enough, Student's t outperforms or ties the Gaussian; when the noise is heavy-tailed or skewed, Student's t dominates. Only for extreme lognormal right-skew does the Gaussian give slightly lower MSE, while Student's t still gives better CRPS. The paper's own conclusion softens the abstract's universal wording to 'generally a more robust choice'.","pith_inferences":["The fixed ν=5 was chosen from a sensitivity analysis run on Gaussian noise; letting ν be learned per dataset might strengthen the result, but would weaken the 'no need to know the noise' message.","Because moment matching fixes the scale from sample variance, the comparison partly conflates distribution family with tail weight; an exact optimization of likelihood parameters could change the ranking, so a direct re-test with optimized scales would isolate the true source of the gain.","The result likely transfers to other ReLU-based Bayesian architectures, since the paper notes that node activations become heavy-tailed in deep layers; testing convolutional or transformer BNNs would be a natural extension.","Even in cases where MSE ties or slightly loses, CRPS improves, so practitioners optimizing for probabilistic forecasts — such as energy load forecasting — should prefer Student's t even if point accuracy looks similar."],"forward_implications":["Replacing the Gaussian likelihood with Student's t (ν=5) is a drop-in change in VI-trained MLP BNNs that improves or ties MSE and CRPS across tested noise families and architectures.","Matching the likelihood family to the data noise is not necessary for good performance; tail robustness matters more than exact distributional match.","Skew-normal likelihood, despite addressing asymmetry, is dominated; asymmetry alone does not drive the gains.","In some real-world settings, Student's t reduces training epochs or wall-clock time, while predictive intervals stay similarly calibrated.","The advantage grows with dataset size; with small samples and Gaussian noise, Gaussian remains competitive or better."],"fun_headline_variants":["Drop-in t-likelihood beats Gaussian in BNNs","Heavy-tailed likelihood wins for Bayesian nets","Student's t: new default for BNN likelihood?","Why BNNs should ditch Gaussian likelihood","t-likelihood robust for Bayesian regression"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The comparison hinges on the likelihood being parameterized by empirical moments of Monte Carlo weight samples; if that moment-matched parameterization is not a fair or consistent representation of the predictive distribution, the ranking between likelihood families could be an artifact of the estimator rather than the family.","fun_headline_variants_meta":{"raw":{"variants":["Drop-in t-likelihood beats Gaussian in BNNs","Heavy-tailed likelihood wins for Bayesian nets","Student's t: new default for BNN likelihood?","Why BNNs should ditch Gaussian likelihood","t-likelihood robust for Bayesian regression"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000862,"raw_usage":{"total_tokens":3576,"prompt_tokens":745,"completion_tokens":2831,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":2758}},"tokens_in":489,"tokens_out":2831,"duration_ms":19317,"temperature":1.0,"reasoning_tokens":2758,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T02:36:13.346507+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the comparison with likelihood parameters optimized directly (for example, maximizing validation log-likelihood for each family) instead of moment-matched, keeping ν=5; if Student's t no longer dominates Gaussian, the paper's conclusion is an artifact of moment matching. Alternatively, on a large dataset with clean Gaussian noise, if Gaussian does not at least tie Student's t when both receive the same estimated scale, the claim that heavy tails help regardless of noise is refuted.","supporting_citations":[],"review_version":1}