acceptodds
Under review as a conference paper at ICLR 2027

Iterative Exact Discrete Guidance for Energy-Based Sampling

Abstract

Sampling from unnormalized distributions over large discrete state spaces becomes difficult when a multimodal target is far from a tractable reference. We introduce *Iterative Exact Discrete Guidance* (IEDG), a population-exact, trajectory-wise guidance framework for unnormalized discrete targets. Rather than learn the full reference-to-target correction in one step, IEDG decomposes the global Boltzmann tilt along an annealing trajectory. Each stage learns a stage-local posterior correction for an incremental Boltzmann tilt of the current source, while the resulting corrections are accumulated relative to a fixed analytic posterior. At the population optimum, exact stage posteriors recover the correct reverse dynamics, whose exact simulation reproduces the target distribution. IEDG chooses stage increments by relative effective sample size (rESS), which controls Rényi- displacement and locally adapts the step size to the thermodynamic geometry of the annealing path. Our stagewise total-variation analysis shows that limited overlap amplifies Bregman fitting error by , while posterior, simulation, and truncation errors enter additively. IEDG improves all distribution-level errors over the neural baselines on ordered, exactly enumerated Ising , while substantially reducing one-shot errors on Ising/Potts across thermodynamic regimes and attaining the best neural-sampler result on several reported local-statistic and phase-coverage metrics. On Max-Cut, its best-of- and average-sample ratios exceed all the baselines.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Discrete neural sampler, Discrete flow matching, Annealed guidance

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_91dHOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work studies sampling from discrete Boltzmann distributions given access to an energy function. The authors propose Iterative Exact Discrete Guidance (IEDG), which uses annealing to gradually learn the correction from a reference distribution to the target. The corrections are accumulated in a single guide used for sampling at the target temperature. The method chooses annealing increments using relative effective sample size, and the authors analyze how overlap between successive distributions affects the propagation of learning errors. Experiments on Ising and Potts models and Max-Cut show improvements over one-shot discrete guidance.

Critical strengths
  1. I think IEDG is a practical extension of one-shot discrete guidance. Learning the correction gradually through annealing makes sense for a difficult source-to-target problem, and the final sampler still needs only one guide. I also find the error analysis useful because it connects overlap between stages to the effect of fitting errors on sampling (Proposition 3 and Theorem 1).

  2. The experiments show big improvements over one-shot DGM. On Ising 4×44\times4 at β=0.6\beta=0.6, IEDG achieves the lowest distributional errors among the evaluated neural samplers (Table 1). On the larger lattices, it improves over one-shot DGM and is competitive with the other neural samplers on several metrics (Table 2).

Critical weaknesses
  1. My main concern is the novelty of IEDG relative to Discrete Tilt Matching (DTM) [1]. DTM already learns a sequence of smaller tilts by directly training the denoising posterior at each stage. Taking negative energy as the reward gives the same exponential reweighting used here. This overlaps with the first claimed contribution (lines: 64-67), but DTM is neither cited nor compared against. The comparison with one-shot DGM shows that IEDG improves on learning the full correction at once, but does not establish what is gained over DTM's incremental approach. A natural baseline is therefore to train the posterior in Eq. (4.3) directly with weighted denoising, as in DTM, instead of learning the guide through Eq. (4.6).

  2. I had trouble understanding what the model is learning in Sections 4.2 and 4.3. The guide and several intermediate quantities are introduced before I could see how they lead to the training objective. I would find the construction easier to follow if the authors started from the desired posterior in Eq. (4.3), identified the correction relative to the fixed reference, and then introduced the guide to represent that correction. Writing it as a conditional expectation would then motivate the regression response and loss. The identity after Eq. (4.5) (lines: 234-235) is helpful, but I think it should lead the explanation.

    For example, the authors already state the aim of using a fixed analytic reference posterior (lines: 214-215). This could be connected directly to Eq. (4.3) by writing, as a function of zz,

    q_{k,1|t}^d(z\mid x_t) &\propto p_{k-1,1|t}^d(z\mid x_t)\\,h_{k,t}^d(z,x_t) \\\\ &\propto p_{\mathrm{ref},1|t}^d(z\mid x_t) \frac{p_{k-1,1|t}^d(z\mid x_t)} {p_{\mathrm{ref},1|t}^d(z\mid x_t)} h_{k,t}^d(z,x_t). \end{aligned}$$ The factor multiplying $p_{\mathrm{ref},1|t}^d$ is the positive correction to represent with a network $G_{\psi_k}$. This gives the parameterization $q_{\psi_k,1|t}^d(z\mid x_t)\propto p_{\mathrm{ref},1|t}^d(z\mid x_t)G_{\psi_k,t}^d(z,x_t)$, normalized over $z$. In practice, replacing the exact source posterior with its frozen estimate $\widehat p_{k-1,1|t}^d$ makes the target correction $A_{k-1,t}^d h_{k,t}^d$, yielding Eq. (4.4) after rescaling by $1/\widehat c_k$. The previous guide can then be introduced as the way to construct this frozen estimate. This explains what the guide represents and where the frozen-posterior approximation enters. To motivate the regression objective, for fixed $t,d$, draw $X_1\sim\mu_{k-1}$ and $X_t\sim p_{t|1}(\cdot\mid X_1)$, keeping $A_{k-1,t}^d$ and $\widehat c_k$ frozen. Substituting the definition of $h_{k,t}^d$ into Eq. (4.4) gives $$\begin{aligned} G_k^{\star,d}(z,x_t) &= \frac{A_{k-1,t}^d(z,x_t)}{\widehat c_k} \mathbb E\\!\left[w_k(X_1)\mid X_1^d=z,\\ X_t=x_t\right] \\\\ &= \mathbb E\\!\left[ \frac{A_{k-1,t}^d(X_1^d,X_t)\\,w_k(X_1)}{\widehat c_k} \\,\middle|\\, X_1^d=z,\\ X_t=x_t \right]. \end{aligned}$$ The quantity inside the expectation is computable from each training pair. Defining it as the response in Eq. (4.5), $$R_{k,t}^d(X_1,X_t) := \frac{A_{k-1,t}^d(X_1^d,X_t)\\,w_k(X_1)}{\widehat c_k},$$ makes clear that its conditional mean is the desired guide. The Bregman loss then provides the regression objective, since $$\begin{aligned} \underset{g > 0}{\operatorname{arg\\,min}}\\; \mathbb E\\!\left[ g-R_{k,t}^d(X_1,X_t)\log g \\,\middle|\\,X_1^d=z,\\ X_t=x_t \right] &= \mathbb E\\!\left[R_{k,t}^d(X_1,X_t)\mid X_1^d=z,\\ X_t=x_t\right] \\\\ &= G_k^{\star,d}(z,x_t). \end{aligned}$$ Using $G_{\psi,t}^d(X_1^d,X_t)$ as the prediction and averaging the loss over training pairs, $t\sim\tau$, and coordinates $d$ gives Eq. (4.6). This ordering would let the reader follow the construction from the desired posterior to the guide and then to its training objective.

[1] Yuyuan Chen, Shiyi Wang, Peter Potaptchik, Jaeyeon Kim, and Michael S. Albergo. Discrete Tilt Matching. 2026.

Questions for authors
  1. Echoing weakness 1, why learn the guide using Eq. (4.6) instead of directly training the denoising posterior in Eq. (4.3) with weighted denoising, as in DTM [1]? Could the authors compare these alternatives using the same annealing schedule and comparable training and sampling budgets? I ask because this would help establish whether the guide objective contributes to the improvement beyond the use of incremental tilts.

  2. How much of the improvement over one-shot DGM comes from iterative training? Does the DGM baseline use the same preconditioning and auxiliary loss? Tables 12 and 13 vary their strengths but keep both enabled, so I cannot tell how much they contribute. Could the authors match these components in the DGM comparison and evaluate IEDG with either or both disabled, keeping the annealing path fixed for the ablations and using comparable budgets?

Reviewer_gwSqOpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

This paper proposed Iterative Exact Discrete Guidance to learn neural samplers in the discrete space. The algorithm start from uniform and then learns to progressively tilting itself to the target. The paper evaluates the algorithm on several discrete targets, including Ising/Potts and MaxCut, achieveing a promising performance.

Critical strengths

The proposed method achieves the SOTA performance across several discrete target, and presents advantages over baselines including similar construction PDNS. However, the size of the target is limited, and this is also related to my concern discussed below.

Critical weaknesses
  1. One weaknes is that this approach seems to need the assumption that the current sample learns well. For each iteration, the target objective is only exact when the current sampler learns perfectly, which never holds in practice. This can lead to error accumulation, especially when target is complicated, higher dimension and requires more anneal stages to learn (e.g.,, when dimension is large, ESS is typically very small between even close targets, and it may require much more stages). This is not an issue for PDNS, or other trust-region based, annealing-based methods [1, 2].

  2. Missing related works. There are many related works using the idea of annealing path to learning samplers [1,2]. Particularly, tilting matching and discrete tilting matching is highly related [3, 4]:Both DTM and IEDG have the same tilting path and feature the same incremental posterior-tilting construction. However, the author did not cite any of these works.

[1] Trust Region Constrained Measure Transport in Path Space for Stochastic Optimal Control and Inference

[2] Learning Boltzmann Generators via Constrained Mass Transport.

[3] Tilt Matching for Scalable Sampling and Fine-Tuning.

[4] Discrete Tilt Matching.

Questions for authors
  1. Did the number of stages needed increase when target increases? Is this relation closer to linear, sublinear or more expensive?

  2. When you decide to finish 1 stage training and move to next? I guess this matters a lot to avoid error accummulation?

Overall the idea of annealing is promising, and the proposed method seem to work well in the tested settings. However, this idea was already verified in my previous works and it is also unclear if this method has its own advantages over them (not necessary better all the time, but whats the uniqueness of it that can benefit the committee and the field is stilll unclear). I would increase my score to 2 if the related works are properly discussed, and I will increase my score further if the author can also demonstrate / explain how this method is more effective given the similar framework and ablate different components of the method and identify which is more useful over previous method like DTM (particularly, is the loss better? is the rule of usig ESS to progress better?) For acceptance, I would also strongly encourage the author further discuss and show the weakness of the method (like assumption of the perfect sampler). This is to avoid the misleading to potential readers.

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

The paper presents a neural sampling approach for discrete distributions with known energy functions. The paper is based on an annealing trajectory connecting a simple distribution with the target. The approach proceeds in stages, using the last learned distribution in the schedule to produce weighted samples from the next distribution in the schedule, until the target is reached. If the increments are sufficiently small, the ESS of the weighted samples will not collapse. The weighted samples essentially reduce the problem at each stage to a discrete generative modeling problem, tackled using discrete flow-matching machinery, with the “Bregman loss” from Discrete Guidance Matching (DGM). The authors explain how it yields an error-propagation bound in terms of the population excess Bregman risk.

Critical strengths

I think it is an interesting topic, and the authors’ approach makes sense and demonstrates great understanding. The implementation appears to be careful, and the results look okay. The theory appears sound, and while it is not exactly a posteriori control, I don’t think it’s reasonable to expect more.

I am happy to see that the authors attempt to generalize across random graphs for Max-Cut sampling.

Critical weaknesses

(1) Exposition. There is a lot of internally defined, slightly buzzwordy jargon (population-exact, trajectory-wise, stage-local, posterior-exact, same-base accumulated bridge, ESS-controlled thermodynamic path) that I think really obscures what is going on. The description I just gave in my summary seems to me to be far more transparent. I would appreciate if things were written in a more down-to-earth way. It’s pretty hard to understand the contribution based on a glance at the abstract and introduction, and the organization of theoretical results could use some cleanup to clarify what they are doing.

(2) I think the experiments are probably too small and don’t really demonstrate a systematic advantage. I’m not sure I’m familiar with all alternatives, but the PDNS comparison on Max-Cut does not seem totally apples-to-apples, and LEAPS would offer another point of comparison based on an annealing schedule. At this point I think ideally we would be seeing more than proof of concept, either real-world applications or, if testing on lattice physics models, real calculations of critical exponents, etc. I think here the authors do not have the benefit of being early in the field, which is annoying, but it is hard to see the purpose of more algorithms without seeing that they really extend the reach of the field.

(3) It could be nice to see some argument / extension that establishes how to remove parametric fitting error during downstream sampling via trajectory importance weighting (with possible resampling along the way). Other diffusion based samplers have shown how to do this.

(4) My last point, which I recognize is somewhat over-broad, concerns the position of neural samplers more generally - as opposed to generative modeling. Especially for samplers based on annealing paths (i.e., most contemporary cases), I think I’d like to see a more honest comparison to classical approaches. Spiritually, I am kind of skeptical that this model could ever cost fewer samples to “crack” the distribution than would properly implemented SMC along the same annealing trajectory, with standard spin-flip mutations. In fact, for the larger experiments, the authors introduce an auxiliary objective that requires some Gibbs sampling during training. Meanwhile one has to accept some parametric fitting error.

So the argument for all this complication and training has to be that you (A) amortize the cost for downstream samples, or (B) can generalize the training across problems. Given that Monte Carlo accuracy hits seriously diminishing returns after say 100k (or fewer) samples, the argument for (A) to me feels shaky - you never really need the billionth sample, unless you are interested in rare events, where I am concerned that the fitting error based on equilibrium training would be especially bad. I understand that my point here is a bit ambitious, but at this point I think the field has enough papers such that it gets harder to accept ignoring some argument for (A).

But relatedly, I do think that (B) presents the real opportunity, though it has to be proved. I am happy to see that the authors attempt to generalize across random graphs for Max-Cut sampling, though I’m not sure they reach the scale of DiffUCO and SDDS. For lattice models, what about an architecture that seeks to generalize across lattice sizes, or magnetic fields? Then you could use the trained model to perform finite size scaling to understand the critical behavior.

In any case, I think my deeper criticism for (A/B) is that it’s not clear to me that the sampling and the neural training have to be coupled, again most especially if the approach is based on an annealing trajectory. We could use our favorite classical method to more cheaply generate a bunch of data, then just do generative modeling on the data we obtained. See, e.g., "On scalable and efficient training of diffusion samplers" by Kim et al for a recent paper taking these points seriously.

Questions for authors

Why not follow the strategy outlined in my summary paragraph, but instead use the ordinary (weighted) denoising cross-entropy loss for the generative modeling problem at each stage? I think this could even yield cleaner error accumulation bounds, since this loss bounds TV via Pinsker’s inequality, without an extra factor involving the ESS. I am wondering if your approach offers some practical advantage.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions and evidence for Area Chairs and Reviewers

Dear Area Chairs and Reviewers,

We thank the AC and reviewers for their careful assessment and constructive discussion. The reviews highlighted IEDG’s gains over one-shot DGM and neural sampler baselines, and the usefulness of its error analysis, noting the practical appeal of retaining a single guide for inference while raising questions about novelty, objective and component attribution, and practical value. Below, we summarize the manuscript revisions and additional rebuttal experiments trying to address these concerns.

1. Contribution and exposition — gwSq, 91dH, TwPC.

Addressing gwSq/91dH on novelty, Section 2 and Appendix B.4 now discuss Discrete Tilt Matching (DTM), Tilt Matching, and constrained-transport methods. We distinguish shared incremental tilting from IEDG's construction: IEDG learns a positive accumulated correction to a fixed analytic reference posterior rather than using DTM’s control-variate-augmented cross-entropy formulation. The preceding model supplies both the next stage’s source samples and a posterior-to-reference ratio, which is multiplied by the incremental energy weight to form the regression response. This conditional-mean construction retains successive corrections in one guide and connects rESS-based increment selection directly to the noise in supervision.

For 91dH/TwPC on clarity, we simplified the Abstract and Introduction and reordered Section 4.3: desired posterior, correction, regression response, then loss. Section 5 connects rESS to scaled response noise. The contribution is thus a specific regression construction and its coupling to rESS, rather than annealing or Bregman learning alone.

2. Controlled evidence for the learning construction — gwSq, 91dH, TwPC.

All three reviewers asked why use regression rather than weighted cross-entropy (CE), so we designed a controlled experiment: revised Table 3 compares IEDG, CE with and without CR, and DTM-style CE with a control variate. Networks, paths and update/sampling budgets are matched. IEDG's near-critical Ising/Potts correlation errors are 0.150/0.123, versus 0.419/0.228 for the stronger CE variant on each task. Additional ordered-stage comparisons also demonstrate better correlation and phase coverage when disabling CR. For 91dH, our matched one-shot comparison gives DGM the same auxiliary training components, total updates and shows iterative learning has much better performance than one-step learning on lattice models.

Together, these tests support an overall advantage for IEDG's regression-based design across the tested Ising/Potts transports, and isolate gains from iteration beyond the auxiliary components.

3. Model dependence and correction — gwSq, TwPC.

For gwSq, revised Sections 5 and 7 separate regression error, posterior inaccuracy, and inherited source error. Accurate local fitting need not remove inherited bias, and the fixed-budget stage-count ablation shows that more stages can worsen accuracy. This makes the trade-off explicit: posterior reuse reduces incremental supervision noise but makes learning dependent on the preceding posterior's accuracy.

For TwPC's request for downstream correction, the rebuttal implements trajectory importance weighting using transition probabilities. It yields consistent target estimates as particle count increases. On Ising 4-by-4, empirical total variation falls from 0.0378 to 0.0125, with substantially higher ESS for IEDG than one-shot DGM. On Ising 16-by-16, sequential Monte Carlo with Metropolis–Hastings moves improves phase weights, while raw sampling retains lower correlation error.

We also test training-time source refinement: eight Swendsen–Wang (SW) sweeps on generated buffers improve subsequent CE correlation recovery on both lattices. These tests establish practical correction routes and their accuracy–cost trade-offs.

4. Practical value alongside classical sampling — TwPC.

The rebuttal tests TwPC's proposed ordinary CE training on independent target SW data. It recovers Ising phase weights better, while IEDG retains lower correlation error. Our additional QM9 Bernoulli adaptation tests another use case: a frozen generator with reusable guidance learned from endpoints and binary joint-constraint checks. Joint success rises from 3.61% to 70.54% at guidance strength one. Including about two hours of network training, projected cost for 50,000 successful attempts is 2.39 h, versus 4.44 h for frozen-generator rejection under the reported hardware conditions. The cost follow-up includes other guidance baselines. IEDG's value here is reusable constraint-specific guidance: higher success and throughput can offset training cost under sufficient repeated use.

We believe these revisions clarify IEDG's contributions and limitations while strengthening its empirical support. We sincerely appreciate the reviewers' feedback and hope our responses address their concerns.

Best regards, The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.