{"id":"b78fe9e9-4812-43ef-9926-97fdd858934a","arxiv_id":"1908.09744","paper_version":4,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper proposes VIS, a variational guide that embeds several iterations of SGLD or SGD and auto-tunes the step size, claiming tighter ELBOs and faster mixing, with experiments on VAEs and state-space models.","lead":"This paper introduces a variational inference method that runs a short MCMC sampler inside the approximating posterior and backpropagates through it to tune the sampler's step size. The authors report strong empirical gains, but the theoretical claim of a tighter ELBO is not proven and relies on dropping an entropy term.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 4.2's 'tighter refined ELBO' omits the log-Jacobian/change-of-variable term for the deterministic SGD flow; without it the objective is not a lower bound and can exceed log p(x).","rationale":"The reader's weakest assumption identifies the missing entropy/change-of-variable term, and I agree that this is the load-bearing issue. Section 4.2's claim is the paper's only formal justification that the VIS objective is a tighter ELBO, and it is exactly what makes optimizing the refined guide a principled variational-inference procedure. The Gaussian example shows the omission cannot be waved away: neglecting the log-Jacobian turns a lower bound into an upper bound of log p(x), and the true refined ELBO can be lower than the original ELBO. This is an internal inconsistency rather than a disagreement with the field, because VIS-D explicitly acknowledges that deterministic flows require the log-determinant. The empirical sections and the released code provide genuine evidence for practical utility, and the criticism is directed only at the theoretical framing. A revised version that either includes tractable Jacobians (or transition entropies as in VIS-MC) or explicitly reframes VIS-P/VIS-G as a heuristic objective would address the concern. Given that the empirical contribution is real and the theoretical flaw is precisely localizable, the reader's CONDITIONAL verdict remains appropriate; I would not move to reject.","tokens_in":14670,"tokens_out":8769,"duration_ms":90018,"concrete_test":"Analytical counterexample: set d=1, log p(x,z) = -z^2/2 + const, q0(z)=N(0,1), and T=1 SGD step F(z)=(1-eta)z. For eta=0.5 compute (a) L_approx=E_q0[log p(x,F(z))-log q0(z)], (b) L_true=L_approx+E_q0[log|1+eta*grad^2 log p(x,z)|], and (c) ELBO_0=E_q0[log p(x,z)-log q0(z)]. If L_approx > log p(x) while L_true < ELBO_0, the Section 4.2 claim fails; including the Jacobian term restores a lower bound, but it is not monotonically tighter in this example.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 (VIS-P and VIS-G) and Section 4.2 replace the transition in the refined guide by a Dirac delta for the deterministic step z=F(z0)=z0+eta*grad log p(x,z0), and set E_{q_phi,eta}[log q_phi,eta]=E_{q0}[log q0]. For a deterministic map the density transforms as q_phi,eta(z)=q0(F^{-1}(z))|det J_F^{-1}|, so the true ELBO is E_q0[log p(x,F(z0))-log q0(z0)+log|det J_F(z0)|]. The omitted Jacobian term is generically nonzero and is not an approximation error. Concretely, take d=1, log p(x,z) = -z^2/2 + const, and q0(z)=N(0,1). Then F(z)=(1-eta)z and log|det J_F|=log(1-eta)<0 for eta in (0,1). The paper's objective equals ELBO_0 + eta(2-eta)/2, which for eta=0.5 is +0.375 relative to ELBO_0 and above log p(x); the true refined ELBO is ELBO_0 + eta(2-eta)/2 + log(1-eta), which is -0.318 relative to ELBO_0. Thus the quantity optimized in VIS-P/VIS-G is not a valid lower bound, and the 'straightforward' tightness proof in Section 4.2 is false as stated. This is internally inconsistent with the paper's own VIS-D paragraph, which states that deterministic flows require the change-of-variable term.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a variational inference framework (VIS) that refines a tractable guide q0,φ(z|x) by applying T steps of a sampler Q_{η,T}(z|z0), such as SGLD or SGD, and then optimizes an ELBO-like objective with respect to both the guide parameters φ and the sampler parameters η. Several approximations to the entropy term are proposed: a particle/Dirac approximation (VIS-P), a pathwise Monte Carlo approximation (VIS-MC), a Gaussian approximation (VIS-G), an exact change-of-variable formulation for deterministic flows (VIS-D, not used in experiments), and a Fokker-Planck approximation (VIS-FP). The claimed theoretical contributions are an unbiasedness argument, a \"refined ELBO\" that is asserted to be tighter than the standard ELBO, and a Taylor-expansion interpretation of the refined gradient. Experiments cover a funnel density, hidden Markov and dynamic linear models, VAEs on MNIST and fashion-MNIST, and a conditional VAE used as a deep Bayes classifier, with code released online.","tokens_in":14985,"tokens_out":8217,"duration_ms":82874,"significance":"If the theoretical claims were sound, this would be a practically valuable framework: it is simple to implement, compatible with probabilistic programming, and it demonstrates how to tune sampler hyperparameters by automatic differentiation. The paper deserves credit for releasing runnable code and for evaluating the method on standard benchmarks; the reported gains, e.g., MNIST test log-likelihood of -82.74 for VIS-5-10 versus -100.91 for a plain VAE, are substantial and worth explaining. However, the central mathematical assertion—that the proposed \"refined ELBO\" is a valid, tighter lower bound—is not substantiated and is false as stated for the Dirac/SGD and Gaussian approximations. Because the entire title and Section 4 rest on this claim, the theoretical contribution needs major reworking, even though the empirical approach may remain useful as a heuristic.","major_comments":[{"comment":"The claim that the Dirac-delta/SGD refined objective is a tighter ELBO than the original is unproven and, as stated, false. For a deterministic flow F(z)=z+η∇log p(x,z), the density transform is q_{φ,η}(z)=q0(F^{-1}(z))|det J_{F^{-1}}(z)|, so the entropy term is not equal to E_{q0}[log q0]. The paper drops log|det J_F| when it sets E_{q_{φ,η}}[log q_{φ,η}] = E_{q0}[log q0] in VIS-P and VIS-G. Concretely, take d=1, log p(x,z) = -z^2/2 + const, and q0(z)=N(0,1). Then F(z)=(1-η)z, the original ELBO equals log p(x) because q0 is the prior, the paper's objective equals log p(x)+η(2-η)/2, which for η=0.5 is log p(x)+0.375, exceeding log p(x); the true refined ELBO including log(1-η) is log p(x)+η(2-η)/2+log(1-η), which is below log p(x). Thus Eq. (4) is not a valid lower bound for VIS-P/VIS-G, and the \"straightforward\" tightness proof must either include the Jacobian term or be restricted to a properly formulated pathwise bound.","section":"Section 4.2, Eq. (4)"},{"comment":"The approximations set the transition entropy to zero and identify E_{q_{φ,η}}[log q_{φ,η}] with E_{q0,φ}[log q0,φ], but no justification is given for why the omitted change-of-variable term is negligible. This is internally inconsistent with the paper's own VIS-D paragraph, which correctly states that deterministic flows require the change-of-variable formula. Since VIS-P and VIS-G are presented as general ELBO strategies, the authors need to specify under which conditions the zero-entropy replacement is valid, or, if it is not valid, explicitly label these objectives as heuristic rather than as ELBOs.","section":"Section 3.2, VIS-P and VIS-G"},{"comment":"The pathwise MC approximation is written as q_{φ,η}(z_T|x) = ∏_{i=1}^T q_η(z_i|z_{i-1}) q0,φ(z0|x), but the right-hand side is a joint density on z_{0:T}, not a density on z_T. If the authors intend to optimize an ELBO on the extended space q(z_{0:T}), then the objective is a valid lower bound—though a looser one than the marginal ELBO—and this should be stated explicitly. If they intend Eq. (4) with the marginal q(z_T), the displayed density is wrong. The VAE experiments rely on VIS-MC, so this ambiguity affects the interpretation of the main empirical results.","section":"Section 3.2, VIS-MC"},{"comment":"The statement that \"for the VIS-FP case ... KL(q_{φ,η}(z|x)||p(z|x)) ≤ KL(q0,φ(z|x)||p(z|x))\" is not generally true for the discretized deterministic flow in Appendix A, Eq. (6). The Fokker-Planck equivalence holds for the continuous-time SDE, not for the Euler step, and no proof is supplied for the discrete case. Since VIS-FP is not used in the experiments this is less central, but it is part of the theoretical analysis and should be corrected or removed.","section":"Section 4.2, VIS-FP"}],"minor_comments":[{"comment":"The HMM joint distribution is written as p(z_{1:τ}, x_{1:τ}, θ) = ∏_{t=1}^{τ} p(x_t|z_t,θ_em) p(x_t|x_{t-1},θ_tr) p(θ); the second factor should presumably be p(z_t|z_{t-1},θ_tr), not another p(x_t|·).","section":"Section 5.2, HMM model equation"},{"comment":"The caption says the evolution of the negative ELBO is shown at the \"bottom\", while the text says the trajectories are shown at the \"top\"; the placement is inconsistent and should be corrected.","section":"Figure 1 caption"},{"comment":"The expression for the refined gradient, ∇_z log p(x,z) + η ∇_z log p(x,z)^T ∇_z^2 log p(x,z), is notationally ambiguous with respect to row/column vectors; the authors should specify the convention and derive the second-order term consistently.","section":"Section 4.3"},{"comment":"The heading \"Unbiasedness\" is misleading: no estimator unbiasedness is proved. The paragraph only argues that consistency results of SG-MCMC apply as T→∞, which is an asymptotic statement, not finite-sample unbiasedness.","section":"Section 4.1"},{"comment":"The phrase \"By treating the gradient terms as points\" is vague; the authors should specify how the covariance σ_φ(x) is used after T deterministic steps and why the resulting distribution is Gaussian with mean z_T.","section":"Section 3.2, VIS-G"},{"comment":"The experiments report that sampler parameters are tuned via ∇_η rELBO, but Appendix B.3.2 gives only the initial value η=5e-5; the adaptation schedule and its effect on training should be described.","section":"Section 5.4 and Appendix B.3.2"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the omitted Jacobian is decisive and lands exactly on Section 4.2. The counterexample shows that the quantity optimized in VIS-P/VIS-G can exceed log p(x), so the paper's central 'refined bound' claim is not merely missing a proof; it is false as stated. I would ask the handling editor to ensure the revision squarely addresses this issue, either by adding the change-of-variable term, by reformulating the particle/Gaussian objectives as heuristics, or by restricting the theoretical claims to the pathwise VIS-MC objective."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: the empirical core is real, the theoretical core is not. The authors propose embedding an SGLD/SGD sampler inside a variational guide, forming a compound distribution, and backpropagating through the sampler to learn the step size. That is a genuinely useful trick, and they back it with released code and strong results on VAE benchmarks (test log-likelihood -82.74 on MNIST versus -96.16 for the same architecture with T=0) and a cVAE classifier (99.8% accuracy).\n\nThe main problem is Section 4.2. The paper claims the refined ELBO is tighter than the standard one, with the proof left as 'straightforward.' It isn't. For a deterministic map like SGD, the density transform carries a log-determinant Jacobian term. The particle and Gaussian approximations in Section 3.2 silently drop it. The stress-test example is clean: in 1D with a Gaussian target, the omitted term is -log(1-eta), which is not negligible. The quantity they optimize is not a lower bound on log p(x); it can exceed it. Worse, Section 3.2's VIS-D paragraph explicitly says deterministic flows require the change-of-variable term, so the paper is internally inconsistent. The same issue undermines Section 4.2's tightness claim and the KL inequality asserted for VIS-FP.\n\nWhere the paper does well: the experiments are extensive and honest. The comparison budgets are matched, the gains are large, and the code is released. The idea of tuning sampler parameters via automatic differentiation is a real contribution, even if the bound story collapses.\n\nThe honest framing is: this is a heuristic that works well empirically, not a variational method with a valid lower-bound objective. If the authors either prove the bound under explicit assumptions or reposition the paper as a sampler-tuning heuristic with empirical support, it could become a solid contribution. As it stands, the theory does not support the headline claim.\n\nFor peer review, I would send it out with the clear expectation that the theoretical claims be fixed or removed. The empirical work deserves an audience; the current theory would mislead them.","headline":"Empirical winner, theoretical miscalibration: the sampler-in-the-guide idea works, but the claimed tighter ELBO is not a valid bound because the log-Jacobian term is dropped.","tokens_in":15575,"tokens_out":2727,"would_cite":false,"duration_ms":28070,"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":"Refining a variational posterior with MCMC sampler steps yields a tighter lower bound and learns the sampler's step size by backpropagation.","keywords":["variational inference","probabilistic programming","Markov chain Monte Carlo","stochastic gradient Langevin dynamics","evidence lower bound","implicit variational distributions","automatic differentiation","Bayesian classification"],"falsifier":"Compute the exact pushed-forward density for a one-dimensional Gaussian target after one SGD step, including the log-Jacobian term, and compare its ELBO with the particle-approximation ELBO over a range of step sizes; the tighter-bound claim collapses at the first step size where the two bounds disagree by a non-negligible amount.","tokens_in":14398,"feed_emoji":"🎲","tokens_out":5916,"duration_ms":58760,"temperature":0.7,"pith_summary":"The paper proposes building variational posteriors by starting from a standard tractable guide and pushing sampled latent values through several iterations of an MCMC-type sampler such as stochastic gradient Langevin dynamics. Its claim is that optimizing the resulting refined evidence lower bound gives a posterior approximation closer to the true posterior than the original guide alone, while automatic differentiation through the sampler steps automatically tunes the sampler's step size. This matters because it turns two separate tasks—designing a flexible variational family and hand-tuning an MCMC sampler—into one end-to-end optimization that can be implemented in any probabilistic programming system supporting algorithmic differentiation. The paper also supplies several approximations for the entropy of the implicit refined density and demonstrates the approach on funnel densities, state-space models, variational autoencoders, and a Bayesian classifier.","feed_headline":"Backprop-tuned MCMC tightens the variational lower bound","feed_subtitle":"A few sampler steps inside the variational guide improve posteriors in VAEs, HMMs, and Bayesian classifiers.","key_machinery":"The load-bearing object is the refined variational approximation, the compound distribution $q_{\\varphi,\\eta}(z|x)=\\int Q_{\\eta,T}(z|z_0)\\,q_{0,\\varphi}(z_0|x)\\,dz_0$, in which a tractable guide is evolved by a sampler transition. The load-bearing simplification is the particle approximation of that transition: approximating $Q$ by a sum of Dirac deltas sets the transition entropy to zero, reducing the refined ELBO to an expectation under the original guide of the log joint evaluated at the perturbed latent point. The paper also introduces MC, Gaussian, Fokker-Planck, and change-of-variable (VIS-D) approximations that track some portion of the lost entropy, and two automatic-differentiation modes, Full AD and Fast AD, that respectively do or do not differentiate through the sampler perturbation. This machinery converts an intractable implicit density into a computable objective that remains differentiable in both the guide and sampler parameters.","core_discovery":"The central discovery is that a compound distribution $q_{\\varphi,\\eta}(z|x)=\\int Q_{\\eta,T}(z|z_0)\\,q_{0,\\varphi}(z_0|x)\\,dz_0$, formed by evolving an initial variational guide through $T$ steps of an MCMC-like transition, can be used as the variational family even though the refined density is implicit. Using a Dirac-delta particle approximation for the transition, the entropy of the flow is dropped and the objective becomes an expectation under the original guide of the log joint evaluated at the perturbed latent point, with the perturbation being one steepest-ascent sampler step. The paper argues that for moderate step sizes this refined bound is tighter than the standard ELBO of the original guide, and that the framework permits tuning both the guide parameters and the sampler parameters by backpropagation. The underlying SG-MCMC sampler remains valid in the limit $T\\to\\infty$, so the learned initial distribution and step size can be reused at test time.","pith_inferences":["If the tighter-bound claim holds generally, the refined ELBO could be used as a convergence diagnostic for the sampler: one could measure how much the bound improves per sampler step and use that to choose an adaptive number of refinement steps.","The dropped Jacobian term suggests a cheap correction for deterministic flows: estimating $\\log\\left|\\det\\left(I+\\eta\\nabla^2\\log p\\right)\\right|$ along sampled trajectories would restore lower-bound validity without full Jacobian storage, bridging the particle and change-of-variable approximations.","The compound-distribution construction could likely be extended to discrete latent variables by replacing the diffusion with a differentiable relaxation, although the entropy approximation would then need a different treatment.","The large improvement from $T=0$ to $T=5$ in the VAE experiments hints that most of the modelling power comes from just a few refinement steps, so a training schedule that anneals $T$ during optimization could give both fast early progress and precise final approximations."],"forward_implications":["Optimizing the refined ELBO with $T>0$ gives tighter lower bounds and posterior approximations with better coverage than the original guide, as demonstrated on funnel, HMM, DLM, and VAE experiments.","The sampler's step size can be learned by gradient descent on the refined ELBO, removing the need to hand-tune that hyperparameter.","Differentiating through $T$ sampler steps costs $O(mT)$ in the number of sampler hyperparameters, keeping the approach practical when those hyperparameters are few.","Full AD and Fast AD offer a trade-off: differentiating through the sampler steps tunes the sampler, while stopping gradients is cheaper and still improves on plain variational inference.","At test time, a learned initial distribution combined with a few extra sampler steps can produce large accuracy gains, such as 99.8% accuracy on MNIST classification with a Bayesian cVAE classifier."],"supporting_citations":[{"why":"Supplies the stochastic gradient Langevin dynamics sampler used as the transition kernel that refines the initial variational guide.","marker":"[29]"},{"why":"Justifies using plain stochastic gradient descent as a deterministic refinement kernel, which underlies the particle and Gaussian approximations.","marker":"[34]"},{"why":"Provides the change-of-variable rule needed to track entropy in deterministic flows; the paper cites it for the VIS-D approximation and notes its computational cost.","marker":"[36]"},{"why":"Defines ADVI, the standard variational-inference approach that the refined approximation extends and compares against.","marker":"[9]"},{"why":"Gives the unbiased implicit variational inference baseline and experimental setup used for the VAE comparisons.","marker":"[23]"},{"why":"Presents the contrastive-divergence compound distribution most similar to VIS, used as a main comparison for the new formulation.","marker":"[28]"},{"why":"Shows the HMC-based deep latent Gaussian model baseline that VIS outperforms in the variational autoencoder experiments.","marker":"[25]"}],"fun_headline_variants":["Sampler steps tighten variational bounds","Backprop-tuned MCMC sharpens ELBO","Refined implicit guide betters VAEs and HMMs","MCMC-embedded variational posterior improves inference"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The particle and Gaussian approximations quietly ignore the randomness that the sampler itself contributes to the refined posterior, and the refined lower bound stands only if that discarded contribution is negligible.","fun_headline_variants_meta":{"raw":{"variants":["Sampler steps tighten variational bounds","Backprop-tuned MCMC sharpens ELBO","Refined implicit guide betters VAEs and HMMs","MCMC-embedded variational posterior improves inference"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00026,"raw_usage":{"total_tokens":1540,"prompt_tokens":847,"completion_tokens":693,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":634}},"tokens_in":463,"tokens_out":693,"duration_ms":7906,"temperature":1.0,"reasoning_tokens":634,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:03:31.556735+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the exact pushed-forward density for a one-dimensional Gaussian target after one SGD step, including the log-Jacobian term, and compare its ELBO with the particle-approximation ELBO over a range of step sizes; the tighter-bound claim collapses at the first step size where the two bounds disagree by a non-negligible amount.","supporting_citations":[{"cited_title":"Bayesian learning via stochastic gradient langevin dynamics","cited_arxiv_id":null,"evidence_quote":"Supplies the stochastic gradient Langevin dynamics sampler used as the transition kernel that refines the initial variational guide."},{"cited_title":"Stochastic gradient descent as approximate bayesian in- ference","cited_arxiv_id":null,"evidence_quote":"Justifies using plain stochastic gradient descent as a deterministic refinement kernel, which underlies the particle and Gaussian approximations."},{"cited_title":"Early stopping as nonparametric variational inference","cited_arxiv_id":null,"evidence_quote":"Provides the change-of-variable rule needed to track entropy in deterministic flows; the paper cites it for the VIS-D approximation and notes its computational cost."},{"cited_title":"Automatic differentiation variational inference","cited_arxiv_id":null,"evidence_quote":"Defines ADVI, the standard variational-inference approach that the refined approximation extends and compares against."},{"cited_title":"Unbiased im- plicit variational inference","cited_arxiv_id":null,"evidence_quote":"Gives the unbiased implicit variational inference baseline and experimental setup used for the VAE comparisons."},{"cited_title":"A contrastive di- vergence for combining variational inference and mcmc","cited_arxiv_id":null,"evidence_quote":"Presents the contrastive-divergence compound distribution most similar to VIS, used as a main comparison for the new formulation."},{"cited_title":"Learning deep latent gaussian mod- els with markov chain monte carlo","cited_arxiv_id":null,"evidence_quote":"Shows the HMC-based deep latent Gaussian model baseline that VIS outperforms in the variational autoencoder experiments."}],"review_version":1}