acceptodds
Withdrawn from ICLR 2027

Improving Multimodal MCMC with Learned Proposal Mixtures

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Local Markov chain Monte Carlo (MCMC) samplers can remain trapped in one mode. We train a shared normalizing flow to propose states from different regions across targets that share the state's conditional distribution given a known statistic. We study a reversible rule that controls how often the sampler switches proposal components, with a Metropolis–Hastings (MH) correction. More frequent switching can help cross modes, but can also increase rejection. For non-atomic targets, we identify when switching cannot increase covariance between successive values of any square-integrable observable. This holds exactly when the target-to-mixture density ratio has the same distribution under every component. Acceptance is then unchanged, and long-run estimation variance cannot increase. We also bound the additional variance of bounded observables caused by component and weight errors, without assuming a spectral gap. Across four asymmetric inverse designs at , proposal mixtures double the per-cycle efficiency of matched target-only flows for estimating mode probabilities; switching adds a further . On nonzero-field lattices, mixtures pass all 40 mixing checks, compared with 8/40 for target-only flows. Corrected flips and parallel tempering remain more efficient. Code is available at https://github.com/Alphachain123/Improving_multimodal_MCMC_with_learned_proposal_mixtures.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
conditional sampling, learned Markov kernels, Metropolis–Hastings, spectral acceleration, Bayesian computation, statistical physics

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_a4CUOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper propose a general MCMC framework to collect samples from multimodal distributions. Ideally, these samples are used for downstream inference tasks.(e.g. estimating E(f(x))E(f(x)) for some function ff)

Motivated by linear inverse problem, this paper assume that the posterior distribution can be determined by some statistics of state xx, e.g. π(x∣y)∝wy(U(x))π⋆(x)\pi(x \mid y) \propto w_y(U(x)) \pi^\star(x), where xx is the state, yy is the observation, π⋆\pi^\star is the prior, UU is some known function. Given this structure, the authors aim to resolve one hardness of the task: If the sampler is trapped within one mode, even though the estimation has low empirical variance, the inference could still be significantly biased; while if one propose to switch mode, two aspects may show up and be harmful: (a) variance increase; (b)high rejection rate.

To address this hardness, the authors propose the following procedure:

  1. Partition the state space in to KK components and introduce an auxiliary label variable.
  2. Train a black-box model (Normalizing flow) to approximate the conditional distribution given observation yy and component label aa.
  3. Use mixture of conditional distributions per component as proposal to run a MH sampler.
  4. Within the sampling process, they propose a whitebox routing rule to switch the component.

On the theory side, they proved that if a calibration condition holds, the proposed routing rule would not increase the lag-one covariance. They then give some case studies for this calibration assumption and discuss the robustness in the Appendix.

On the experiment side, they test their method's performance on Gaussian mixtures and nonzero-field lattices.

Critical strengths

The main novelty of this paper is sustained by their Theorem 1, where they show the lag-one covariance is not increasing if and only if the calibration condition holds.

Critical weaknesses
  1. Writing is unacceptable. The abstract and intro are either ai-generated or heavily ai-assisted. The ICLR 2027 policy did not forbid this but in this specific case, the readability of those parts is very restricted. I suggest the authors put more human effort to polish the writing and try to make the article readable. I would like to make two concrete examples to justify my statement:

(a). The core contribution of this paper is the analysis of the routing rule in Theorem 1. In the intro, the authors define and discuss the condition required for inverse problem or posterior in the very begining and followed the notation afterwards, which is unnecessary and redundant. The whole statement still holds for general multimodal distribution π(x)\pi(x) and all the discussion about posterior could be first deleted and push back to later sections if they would like to incorporate case studies.

(b). Their conclusions on hard routing and indepent mixing is straightforward and should serve as intuitive arguments instead of a notable contribution. Intuitively, the blocks discussed in this paper are disjoint so if the learned component is exact, a hard routing would never touch other components. The exact version could be relaxed, where small KL/chisq error implies small probability mass on support mismatch. The proof to quantify this and give downstream arguments are straightforward and standard. The authors should simplify their argument on this part and convey the essential idea. Ideally, the saved space could used to show more content about the calibration assumption.

  1. The contribution of this paper remains unclear and the applicability is somehow restricted, at least in this version. I would like to justify my arguments in the following aspects:

(a). Restricted to posterior sampling, the authors did not discuss and compare with the diffusion and other flow based sampler, which is also an important line of research e.g. [1-3]. In particular, [1,2] studied the linear inverse problem; [3] targets on more general distributions. The settings of these papers may be different with this one, but a proper dicussion would make the scope of this paper clearer. In particular, because of this, the applicability of this work to posterior sampling problem remains unclear: Given the empirical success of diffusion based samplers and the theoretical results/understandings of them, the author should justify the advantage or contribution of their framwork. I think the comparsion is fair and necessary as the author also trained a blackbox model to approximate the conditional distribution per component.

(b). The proposed framework in this paper requires learning components. However, the advantage of this approach remains unclear to me. Are there any principled way for us to partition the state space? Could we get more refined sample complexity bound by using this framework? How could we establish and test the calibration condition with different partition designs? Some of my questions are reflected in the experiments, but I did not see a well-orqganized discussion.

(c) The experiment is not convincing. The advantage of the proposed method is not clear compared with modified HMC on the lattice experiment. And the scale of the Gaussian experiment is rather restricted. Given the fact that linear inverse problem is extensively discussed in image generation field, experiments on real-world data would be much more convincing.

[1] Bruna J, Han J. Posterior sampling with denoising oracles via tilted transport[J]. arXiv preprint arXiv:2407.00745, 2024.

[2] Xun Z, Gupta S, Price E. Posterior sampling by combining diffusion models with annealed langevin dynamics[J]. Advances in Neural Information Processing Systems, 2026, 38: 76108-76165.

[3] Parulekar A, Rout L, Shanmugam K, et al. Efficient approximate posterior sampling with annealed langevin monte carlo[C]//International Conference on Learning Representations. 2026, 2026: 40513-40543.

Questions for authors

See weakness.

Reviewer_RjNUOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

suppose users want to sample from a multimodal distribution, and they already have a mixture proposal (through e.g. normalizing flow) which estimates of each mode's local distribution.

The paper studies how to select these components to construct a better proposal, especially how frequent we should switch from modes. Theorem gives an exact condition under which more frequent switching cannot help. The main information seems to be, suppose the estimation error (between proposal and target) are uniform over components, then adding more switching will not hurt mixing.

The authors also provide toy and flow-based experiments to support their result.

Critical strengths

I think Theorem 1 is moderately interesting. It provides a reasonable metric on quantifying the effect of switching and is intuitive.

Critical weaknesses
  1. The paper is extremely difficult to read. I've spent quite a lot of time but still cannot totally get the main contribution.

  2. I felt the paper's main contribution is Theorem 1, i.e., it feels more of a theory paper. Meanwhile I believe Theorem 1 does not rely on neural net/normalizing flow.

On the other hand, the writing (especially title and abstract) feels closer to a method paper. For example, title contains "learned proposal", and the intro mentions the components are learned from a normalizing flow. However, the experiments seem too limited for a methods paper, while the theory do not seem sufficiently substantial or novel for a theory paper.

  1. I also have reservations about the problem formulation itself. I think the main challenge in multimodal sampling is to estimate the number of modes, locate them, and then fit the distribution around these modes. However, the paper treats all these as given, and focuses on calibrating an already-learned proposal. In my opinion, if the proposal is learned poorly, the subsequent adjustment will not make up for its inefficiency. If the proposal is already learned well, then adjusting the switching frequency will at best have very modest gains.
Questions for authors

Can you clarify your main contribution, is it theoretical or methodological?

Reviewer_PYGWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper looks to extend work on global-local MCMC schemes, popularised as flow based global moves with local “classic” MCMC exploration, this line of work typically aims to address problems that MCMC chains struggle with such as multimodality. This paper proposes to train flows on exact posterior draws, and introduces a scalar switching to encourage moves between isolated modes. The primary result is a theoretical claim about the efficacy of these switching results.

Critical strengths

This is a challenging problem and one that is seeing active work in venues such as ICLR. The construction around Theorem 1 appear to be novel and sound. The result for which switching cannot increase covariance appears sound and a useful point for methods aiming to improve exploration of multimodal posterior landscapes.

Critical weaknesses

My main concern is that this work exists in the regime where it is assumed that one has an exact oracle of the target, or one fitted to asymptotically correct draws from the target. This can be useful for studying theoretical properties but I think that any practical insight is fundamentally limited as a result of this. In this setting I think the theoretical results would have to be stronger to justify acceptance, or have a clearer route to application to practical problems.

In the setting of multimodal targets, the challenging part is estimation of the normalizing constant of each mode effectively, or even discovering modes in the first place without apriori knowledge. This level of oracle knowledge seems required to perform the proposed switching, and for cases where the symmetry is apriori known, as in for example lattice models with a Z2 symmetry, the advantage over simple MH proposals that correspond to the known symmetry looks to be minimal.

The theoretical characterization of switching is interesting, but its connection to practical performance remains uncertain, particularly as the authors do not verify that the required conditions hold for their learned proposals. The empirical results do not fully resolve this concern. Much of the reported improvement comes from using conditional mixture proposals rather than the switching mechanism itself, and the lattice experiments remain substantially less efficient than existing methods based on corrected symmetry flips or parallel tempering. I am concerned that for perfectly fitted conditional proposals over disjoint modes, the calibration condition reduces to requiring the correct mixture weights, in which case the resulting mixture already generates exact independent target samples without any switching needed.

Questions for authors

When I was initially reading this I thought the flow would be used in a blocked gibbs like sampler, i.e. allowing conditional moves of subsets of parameters. As this work is posed in the context of lattices, could the authors expand on how the block moves correspond to cluster moves that are established for Ising models (1)? The sharper question on these kinds of systems is correlation length close to criticality, and I would wonder if the proposed switching has any utility there?

Could the authors demonstrate their approach on a more realistic inverse problem, ideally with real data, where exact posterior samples and mode probabilities are not already available? In particular, how would the training data and partitions be obtained in such a setting, and what advantage would the resulting sampler provide over the method used to generate those training samples? Additionally could the authors show that their method scales well when he multimodality is less confined to a subspace.

There is quite a lot of recent work on learned flow based samplers in multimodal settings, and the limitations of this. If the authors think their approach could improve these settings then these might be useful in terms of experimental benchmarks to compare to (2) and (3) for example. I also think this line of literature should appear more prominently in the introduction and motivation, and I think there are other references on this line that would be more useful for an exposition of why multimodal sampling is difficult/interesting.

(1) Nonuniversal critical dynamics in Monte Carlo simulations (Swendsen and Wang)

(2) Improving the evaluation of samplers on multi-modal targets. Grenioux, Louis, Maxence Noble, and Marylou Gabrié.

(3) Learned Reference-based Diffusion Sampler for multi-modal distributions. Maxence Noble, Louis Grenioux, Marylou Gabrié, Alain Oliviero Durmus

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.