- 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
-
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).
-
The experiments show big improvements over one-shot DGM. On Ising at , 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
-
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).
-
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 ,
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
-
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.
-
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?
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