{"id":"7f5015e4-5127-4f0a-907f-a6b4eb5aa108","arxiv_id":"2507.02619","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"L-VAE applies Kendall et al.'s uncertainty weighting to learn the beta weight of beta-VAE from data, achieving competitive but not consistently top-two results.","lead":"This paper proposes L-VAE, a variational autoencoder that learns the weighting between reconstruction error and KL divergence during training instead of tuning it by hand. It compares against five VAE variants on four disentanglement benchmarks, claiming consistent top-two performance, but the evidence is weakened by missing error bars and overclaims.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The abstract's 'consistently best or second best' claim is contradicted by Table 2: on dSprites MIG, L-VAE (0.30) ranks behind ControlVAE (0.34) and DynamicVAE (0.31), making it third. Without seeds or error bars, the central performance claim is not established and needs softening.","rationale":"I focused on the central empirical claim rather than the loss derivation. The Kendall-based derivation has a real gap (log sigma vs sigma^2 changes the optimum), but even a heuristic weighting scheme could be a useful contribution if the performance claim held. The paper's own Table 2 makes the strongest form of the claim untenable, and the ablation raises further doubt about the dynamic-learning benefit. The reader's weakest_assumption points to Eq. 4-6, but their rationale already flags the Table 2 inconsistency; I treat the table contradiction as the primary load-bearing concern, with the derivation gap as secondary. The CONDITIONAL verdict remains appropriate: the paper should soften the abstract, report seeds/error bars, and reconcile Table 3 with the text.","tokens_in":15816,"tokens_out":6712,"duration_ms":67503,"concrete_test":"Re-rank every metric-dataset cell in Table 2 and verify whether L-VAE is first or second in all 24 cells; dSprites MIG already violates this. Then re-run the dSprites MIG comparison with 10 random seeds and report mean and standard deviation for L-VAE and ControlVAE; if ControlVAE's mean exceeds L-VAE's mean by more than the pooled standard error, the 'consistently best or second best' claim must be removed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: L-VAE 'consistently provides the best or the second best performances.' Table 2 is the paper's only quantitative support, and it contradicts the claim. In dSprites MIG, L-VAE scores 0.30, while ControlVAE scores 0.34 and DynamicVAE scores 0.31, so L-VAE is third (tied with beta-VAE). In Falcor3D MIG and SAP, L-VAE (0.05, 0.05) is not top-two, with beta-VAE and ControlVAE outperforming it. The paper's own Section 7.2 uses the softer statement 'generally produces better or on par performance,' so the abstract overstates the table. The ablation in Table 3 also fails to support the claimed benefit of dynamic learning: in Falcor3D, beta-VAE with the learned beta-hat = 2.34 achieves higher beta-VAE (0.93 vs 0.88), FactorVAE (0.65 vs 0.47), MIG (0.12 vs 0.05), and SAP (0.09 vs 0.05) than L-VAE; in MPI3D, beta-VAE with beta-hat matches or exceeds L-VAE on several metrics. Because no error bars, standard deviations, or multiple seeds are reported, the small differences in Table 2 cannot be distinguished from noise. The load-bearing problem is not the derivation gap (Eq. 4 vs 5), which is secondary to an empirical method; it is that the headline performance claim is internally contradicted by the presented evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes L-VAE, a variational auto-encoder that learns the relative weight between the reconstruction term and the KL term by introducing two trainable parameters, σ0 and σ1, together with a quadratic regularizer. The effective β is defined as β̂ = σ0^2/σ1^2. The method is compared with VAE, β-VAE, ControlVAE, DynamicVAE, and σ-VAE on dSprites, MPI3D, Falcor3D, and Isaac3D using six disentanglement metrics and reconstruction loss, plus qualitative latent traversals on CelebA. The paper claims that L-VAE removes the need for empirical β search and consistently achieves the best or second-best performance among the compared methods, and that the learned β̂ aligns with the empirically tuned β.","tokens_in":16223,"tokens_out":7431,"duration_ms":71522,"significance":"If the claims were fully supported, the contribution would be practically useful: a simple, auto-differentiable way to replace the β search in β-VAE. The paper has genuine strengths: a systematic sensitivity analysis of β in Section 4.1 (Observations 1–4, Fig. 2), a broad evaluation across four datasets and six disentanglement metrics, and inclusion of several adaptive-β baselines. However, the central empirical claim is not established by the reported single-run tables, and the ablation in Table 3 does not support the importance of dynamic weighting. The significance of the paper is therefore conditional on substantial revision and additional evidence.","major_comments":[{"comment":"The abstract and the contributions list state that L-VAE 'consistently provides the best or the second best performances', but Table 2 contradicts this. On dSprites MIG, L-VAE scores 0.30, behind ControlVAE (0.34) and DynamicVAE (0.31), placing it third together with β-VAE. On Falcor3D MIG, L-VAE (0.05) trails β-VAE (0.07), ControlVAE (0.06), and DynamicVAE (0.06); on Falcor3D SAP it is tied for third with β-VAE (0.05) behind ControlVAE (0.07) and DynamicVAE (0.06). Section 6.3 reports hyperparameter searches but no multiple seeds, and Table 2 gives no error bars or standard deviations, so the small differences cannot be distinguished from noise. Section 7.2 uses the softer statement that L-VAE 'generally produces better or on par performance', which is consistent with the table but not with the abstract. The headline claim needs to be corrected or supported by repeated runs with statistical measures.","section":"Abstract and Section 1; Table 2"},{"comment":"The derivation from Kendall et al. is not carried over correctly. Equation (4) is presented as the Kendall et al. multi-task loss with log(σ0) + log(σ1) regularizers, but Equation (5) replaces these with σ0^2 + σ1^2 without any derivation or justification. This is not a minor notational change: the two regularizers lead to different optimization landscapes and different interpretations of the learned σi. Moreover, β̂ = σ0^2/σ1^2 is by definition the effective weight on the KL term in Eq. (5), so the reported 'alignment' between β̂ and a searched β is a consistency check on fitted values, not an independent prediction. The fitted ratios (1.39, 1.05, 2.34, 0.95) differ from the searched β values (2 or 4) by up to about a factor of two, which further weakens the alignment claim in Section 7.2. The authors should either derive Eq. (5) from a specific noise model or explicitly state that the σi^2 penalty is a heuristic regularizer.","section":"Section 5, Eqs. (4)-(5)"},{"comment":"The ablation study does not support the claim that dynamic weighting is important. For Falcor3D, a standalone β-VAE trained with the learned β̂ = 2.34 outperforms L-VAE on β-VAE score (0.93 vs 0.88), FactorVAE (0.65 vs 0.47), IRS (0.35 vs 0.30), MIG (0.12 vs 0.05), and SAP (0.09 vs 0.05). For MPI3D, β-VAE with β̂ = 1.05 achieves equal or better scores on all six disentanglement metrics (e.g., β-VAE 0.74 vs 0.71, FactorVAE 0.46 tie, MIG 0.17 vs 0.16, SAP 0.21 vs 0.20) and a lower reconstruction loss (10.63 vs 10.79). The conclusion in Section 9 states that β-VAE with β̂ achieves better disentanglement 'with the cost of an increased reconstruction loss', but this is contradicted by the MPI3D row and by the dSprites row (reconstruction 18.13 vs 21.14). These results undercut the paper's central novelty, which is the benefit of learning the weights dynamically instead of fixing them.","section":"Section 7.3, Table 3"}],"minor_comments":[{"comment":"The text 'Kp to 0.0.01' appears to be a typo; presumably the intended value is 0.001 as in the ControlVAE baseline.","section":"Section 6.3"},{"comment":"The affiliation contains the typo 'Universty' and the Figure 4 caption contains 'experimens'; these should be corrected.","section":"Title page and Figure 4"},{"comment":"The text refers to 'Equation 6' for the overall loss, but the loss is given in Eq. (5) and the regularization term is not separately numbered. Cross-references to equations should be fixed.","section":"Section 6.5"},{"comment":"The statement that 'we select the σi values that maximize the β-VAE score' should clarify that this selection is performed on the validation set (as described in Section 6.5) and that the reported results are on a held-out test set, to avoid any appearance of selecting on the test metric.","section":"Section 7.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is not ready for publication in its current form. The abstract's ranking claim is directly contradicted by Table 2, and the ablation in Table 3 undermines the claimed benefit of dynamic weighting. The authors should be asked to rerun experiments with multiple seeds, report standard deviations or statistical tests, correct the abstract, and either derive the modified regularizer or clearly label it as a heuristic. The core idea is simple and potentially useful if these issues are resolved, but the current evidence does not support the central claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWhat you should know: L-VAE is a simple auto-weighted beta-VAE. It applies Kendall et al.'s multi-task uncertainty weighting to the VAE loss, learning both the reconstruction scale and the KL scale. The paper is honest that sigma-VAE is an instantiation with one learned scale, so the real delta is a second learnable weight and a sigma-squared regularizer instead of log-sigma. That's a routine extension, but it is a convenient one.\n\nThe paper's value is mainly empirical. They run four datasets, six metrics, five baselines, and include an ablation where a static beta-VAE is trained with the learned beta-hat. The observations about beta sensitivity (including beta < 1 working better on MPI3D and Isaac3D) are worth having. The writing is clear, and the method is easy to adopt.\n\nNow the soft spots. The abstract says L-VAE \"consistently provides the best or the second best performances.\" Table 2 does not support that: on dSprites MIG, L-VAE (0.30) is behind ControlVAE (0.34) and DynamicVAE (0.31); on Falcor3D MIG and SAP, L-VAE is not top-two. The paper's own Section 7.2 uses the softer \"better or on par.\" The abstract overstates the data. Second, there are no seeds or error bars. Many differences are a few hundredths of a metric, so you can't distinguish signal from noise. Third, the ablation cuts the other way in some places: on Falcor3D, static beta-VAE with the learned beta-hat (2.34) beats L-VAE on beta-VAE (0.93 vs 0.88), FactorVAE (0.65 vs 0.47), MIG (0.12 vs 0.05), and SAP (0.09 vs 0.05). That undermines the claim that dynamic weighting is itself the reason for the gains. On MPI3D the static version matches or beats L-VAE on several metrics.\n\nThe derivation gap (Kendall uses log-sigma regularizers; the paper uses sigma^2 without derivation) is secondary, but it does mean the method is a heuristic weight-annealing scheme rather than a principled uncertainty estimate.\n\nOverall: a plausible convenience, well-written, with a broad but statistically weak evaluation. The central performance claim as stated is not credible, but a softened version might be. Fix the abstract, add seeds and error bars, and reconcile the ablation, and this becomes a reasonable short paper or workshop contribution. I would not cite it in my own work as is, but I would bring it to a reading group to discuss the gap between claim and table. A serious editor should send it to review — it is not a desk reject — but the referee should require the claims to match the evidence.","headline":"L-VAE is a simple Kendall-style weight learner for beta-VAE, but the 'consistently best or second best' claim contradicts the paper's own tables.","tokens_in":16800,"tokens_out":2995,"would_cite":false,"duration_ms":30738,"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 variational autoencoder that learns the weight of its own loss terms matches or beats hand-tuned beta-VAE.","keywords":["variational autoencoder","disentangled representation learning","learnable beta","uncertainty weighting","beta-VAE","multi-task learning","hyperparameter optimization"],"falsifier":"Train L-VAE and $\\beta$-VAE across a fine grid of $\\beta$ on a fixed benchmark: if the learned ratio $sigma_0^{2}$/$sigma_1^{2}$ does not track the $\\beta$ value that maximizes the chosen disentanglement metric, or a static $\\beta$-VAE trained at the learned ratio matches L-VAE everywhere, the dynamic-benefit claim would fail. A direct check of the derivation is to train with the original log-$\\sigma$ regularizer of Eq. 4 instead of the $\\sigma$-squared term of Eq. 5 and compare the learned weights and scores.","tokens_in":15576,"feed_emoji":"🧩","tokens_out":4580,"duration_ms":47417,"temperature":0.7,"pith_summary":"This paper tries to show that the $\\beta$ hyperparameter of $\\beta$-VAE, the weight balancing reconstruction quality against disentanglement in the loss, does not need to be tuned by hand. The proposed L-VAE learns two scale parameters sigma_0 and sigma_1 alongside the network weights, so that the ratio $sigma_0^{2}$/$sigma_1^{2}$ plays the role of $\\beta$ automatically. Across dSprites, MPI3D-complex, Falcor3D, and Isaac3D, the paper reports that L-VAE gives the best or second-best scores on six disentanglement metrics, while also finding that the optimal effective $\\beta$ is below 1 on some datasets. If correct, this removes the expensive $\\beta$ search that practice requires, and it makes the trade-off adapt to the data rather than being fixed in advance.","feed_headline":"Learnable weight replaces beta search in disentangled VAEs","feed_subtitle":"L-VAE sets reconstruction and KL weights by gradient descent and matches or beats tuned beta-VAE on four benchmarks.","key_machinery":"The mechanism is Kendall et al.'s uncertainty weighting for multi-task losses, transplanted from Gaussian log-likelihood tasks to the two terms of the VAE objective. Two learnable scalars sigma_0 and sigma_1 divide the reconstruction and KL terms, with a regularization term, here $sigma_0^{2}$ + $sigma_1^{2}$, that stops the weights from growing without bound; the ratio $sigma_0^{2}$/$sigma_1^{2}$ is the effective $\\beta$. This turns hyperparameter search into an ordinary optimization problem, since the sigma_i sit in the optimizer's parameter set and are updated with the network weights. The paper's argument relies on the learned ratio being aligned with the empirically tuned $\\beta$ of $\\beta$-VAE.","core_discovery":"The paper's central claim is that the relative weight between the reconstruction term and the KL-divergence term of the ELBO can be learned by gradient descent instead of being set by search, and that doing so preserves or improves disentanglement while lowering reconstruction loss. L-VAE is $\\beta$-VAE with the loss rewritten as (1/$sigma_0^{2}$) L_R + (1/$sigma_1^{2}$) D_KL + $sigma_0^{2}$ + $sigma_1^{2}$, where sigma_0 and sigma_1 are optimized together with the encoder and decoder. The learned ratio $\\beta$-hat = $sigma_0^{2}$/$sigma_1^{2}$ lands where the optimal tuned $\\beta$ lands in their experiments: 1.39 on dSprites, 1.05 on MPI3D, 2.34 on Falcor3D, and 0.95 on Isaac3D. An ablation that freezes this ratio into a plain $\\beta$-VAE gives similar or worse disentanglement than the dynamic version, which the paper reads as evidence that continuously learning the weights, not just the final ratio, is doing the work.","pith_inferences":["The observation that beta below 1 can beat beta above 1 on real-image datasets is a testable challenge to the common assumption that stronger KL weighting always helps; applying L-VAE to new datasets would cheaply map where that assumption fails.","The paper swaps the log-sigma regularization of the source derivation for a sigma-squared term without justification, so an ablation of that regularizer choice would show whether the learned ratios are sensitive to it.","Model selection in the experiments uses the beta-VAE measure on the validation set, so rankings might shift under other selection criteria; reporting L-VAE's performance across selection metrics would test the strength of the comparison.","All sigma_i are initialized to 1 and learned with a OneCycle schedule, so L-VAE's dynamic behavior may interact with optimization schedules such as KL annealing; combining L-VAE with such schedules is a natural extension."],"forward_implications":["Users of beta-VAE can drop the beta grid search and initialize sigma_0 = sigma_1 = 1, letting the optimizer find the trade-off.","Datasets where low beta is better, such as MPI3D and Isaac3D, are handled automatically because the learned ratio can go below 1 without special-casing.","Because freezing the learned ratio into a static beta-VAE gives similar or slightly worse quality, continually learning the weights appears to be part of the gain, not just the final value.","The same parameterization transfers to other two-term losses, since the paper's derivation is not specific to VAEs.","Across the six reported metrics no single method wins everywhere, but L-VAE is best or second best on the aggregate, which is the paper's stated comparative claim."],"supporting_citations":[{"why":"Supplies the uncertainty-weighting method for learning relative loss weights that L-VAE is built on.","marker":"[19]"},{"why":"Defines beta-VAE, the main baseline and the source of the beta-VAE disentanglement measure used throughout.","marker":"[14]"},{"why":"Provides the capacity-control view of beta-VAE that the paper's balance analysis extends.","marker":"[3]"},{"why":"Decomposes the ELBO and documents the reconstruction-disentanglement trade-off that motivates the paper's observations.","marker":"[5]"},{"why":"ControlVAE is an automatic-beta baseline that L-VAE compares against and seeks to simplify.","marker":"[37, 38]"},{"why":"DynamicVAE is another automatic-beta baseline that L-VAE compares against.","marker":"[39]"},{"why":"Sigma-VAE learns the reconstruction weight and is the closest prior instantiation, used as a baseline.","marker":"[35]"},{"why":"Provides evidence that hyperparameter selection drives disentanglement, motivating a learnable beta.","marker":"[27]"},{"why":"Supplies the six disentanglement measures used to evaluate all compared methods.","marker":"[4]"}],"fun_headline_variants":["L-VAE learns beta instead of tuning it by hand","Gradient descent sets VAE loss weights for disentanglement","Learnable beta for VAE matches tuned search on benchmarks","Learnable beta in VAE: matches or beats tuned beta-VAE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the uncertainty-weighting derivation, made for combining Gaussian log-likelihood task losses, still holds when one of the two terms is the KL divergence of the ELBO, and that replacing the derivation's log-sigma regularizer with sigma-squared leaves the learned trade-off intact.","fun_headline_variants_meta":{"raw":{"variants":["L-VAE learns beta instead of tuning it by hand","Gradient descent sets VAE loss weights for disentanglement","Learnable beta for VAE matches tuned search on benchmarks","Learnable beta in VAE: matches or beats tuned beta-VAE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000862,"raw_usage":{"total_tokens":3780,"prompt_tokens":1025,"completion_tokens":2755,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":2685}},"tokens_in":641,"tokens_out":2755,"duration_ms":24177,"temperature":1.0,"reasoning_tokens":2685,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:25:33.317198+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train L-VAE and $\\beta$-VAE across a fine grid of $\\beta$ on a fixed benchmark: if the learned ratio $sigma_0^{2}$/$sigma_1^{2}$ does not track the $\\beta$ value that maximizes the chosen disentanglement metric, or a static $\\beta$-VAE trained at the learned ratio matches L-VAE everywhere, the dynamic-benefit claim would fail. A direct check of the derivation is to train with the original log-$\\sigma$ regularizer of Eq. 4 instead of the $\\sigma$-squared term of Eq. 5 and compare the learned weights and scores.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp 7482–7491","cited_arxiv_id":null,"evidence_quote":"Supplies the uncertainty-weighting method for learning relative loss weights that L-VAE is built on."},{"cited_title":"In: International conference on learning repre- sentations","cited_arxiv_id":null,"evidence_quote":"Defines beta-VAE, the main baseline and the source of the beta-VAE disentanglement measure used throughout."},{"cited_title":"arXiv preprint arXiv:180403599","cited_arxiv_id":null,"evidence_quote":"Provides the capacity-control view of beta-VAE that the paper's balance analysis extends."},{"cited_title":"In: Advances in Neural Information Processing Systems, pp 2610– 2620","cited_arxiv_id":null,"evidence_quote":"Decomposes the ELBO and documents the reconstruction-disentanglement trade-off that motivates the paper's observations."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 19250–19259","cited_arxiv_id":null,"evidence_quote":"DynamicVAE is another automatic-beta baseline that L-VAE compares against."},{"cited_title":"In: International Conference on Machine Learning, PMLR, pp 9179–9189","cited_arxiv_id":null,"evidence_quote":"Sigma-VAE learns the reconstruction weight and is the closest prior instantiation, used as a baseline."},{"cited_title":"In: international conference on machine learning, pp 4114–4124","cited_arxiv_id":null,"evidence_quote":"Provides evidence that hyperparameter selection drives disentanglement, motivating a learnable beta."},{"cited_title":"IEEE Transactions on Neural Networks and Learning Systems","cited_arxiv_id":null,"evidence_quote":"Supplies the six disentanglement measures used to evaluate all compared methods."}],"review_version":1}