acceptodds
Under review as a conference paper at ICLR 2027

Fixed-point Neural Samplers in Discrete Space

Abstract

Sampling from discrete, unnormalized distributions without access to data is a challenging problem. Neural samplers offer a promising approach by training generative models from density evaluations directly. Despite recent progress, existing discrete neural samplers are prone to mode collapse, come without convergence guarantees when trained via fixed-point iterations, and are often tied to a specific reference process such as masked or uniform diffusion. In this work, we introduce Discrete Gibbs Iterative Neural Sampler, a fixed-point neural sampler that addresses these limitations, enabling efficient, scalable learning, substantially reducing mode collapse in practice. Our framework builds on masked diffusion and also extends to transport between pairs of distributions. We demonstrate that the resulting method scales effectively to high-dimensional systems, supports amortized sampling across different conditions, and enables accurate estimation of alloy phase diagrams.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
neural sampler, sampling, discrete diffusion, masked diffusion, discrete interpolants

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_E24BOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DIG-IN, a neural sampler for discrete unnormalized distributions. It uses exact single-site Gibbs conditionals as soft supervision for a masked predictor, together with tempered importance weighting and damping. The paper establishes population-level consistency and convergence results under specified assumptions, and evaluates the method on Ising/Potts models, amortized sampling across thermodynamic conditions, and alloy phase diagrams. The controlled lattice experiments are encouraging.

Critical strengths

The Gibbs-supervision construction has a principled foundation. Proposition 3.1 establishes consistency and per-site variance reduction under target sampling, and the subsequent analysis distinguishes preserving the target as a fixed point from convergence toward it.

The controlled lattice experiments provide meaningful evidence. Tables 1 and 8 evaluate complementary metrics and include controls for Gibbs labels, importance weighting, and damping. These results support the usefulness of the combined method and should be distinguished from the less conclusive alloy comparison.

Critical weaknesses

N/A

Questions for authors

Clarification of theory versus implementation: What range of tempering exponents is selected during ordered-phase training? What exactly is the multi-site reveal procedure in Table 4, and which fixed-point or convergence statements apply to it? Does the damping-dependent behavior in Table 8 remain when generation uses a single fixed order?

Clarification of reference accuracy: How were the reference log-partition values obtained for the nonzero-field amortized lattices and the fixed-composition alloy tables? Equations (134)–(135) define and transform the alloy quantities but do not describe how the sums were computed.

2 comments
AuthorsOpenReview ↗

Reply to Reviewer E24B

We thank the reviewer for their careful reading and constructive feedback. Below, we address each point in turn and indicate the corresponding changes in the revised manuscript.


What range of tempering exponents is selected during ordered-phase training?

To avoid any confusion, we are not selecting the tempering exponents manually; it is automatically selected using the ESS control in in Eq. (13). Nevertheless, we added an additional experiment in the appendix (App. F.4, Tempering exponent during training) that shows the tempering exponent, which the ESS control sets automatically over the course of training. In the beginning ν\nu is small and therefore smooths the importance weights a lot whereas in later stages when the model approaches the target ν\nu goes to 1 (no smoothing).


What exactly is the multi-site reveal procedure in Table 4, and which fixed-point or convergence statements apply to it?

The 50×5050\times50 amortized sampler generates a configuration in 512 instead of 2,500 network calls: we first reveal one by one for the first 64 sites, and then randomly reveal 5-6 sites together. The probability of the sample is accumulated exactly, so the importance weights remain exact for this procedure, such that the ln⁡Z\ln Z estimates stay unbiased and reweighted observables stay consistent.

This can be viewed as a coarser discretization for the underlying CTMC process. Our Theorem 3.2 does not consider the discretization error, and hence it holds up to the discretization error.

To quantify the effect, we sampled the trained 50×5050\times50 models both with the multi-site reveal and with exact single-site reveal. The two procedures give the same mode balance in every setting. The block reveal therefore does not cause or hide mode collapse. It lowers the effective sample size, most strongly in the disordered phase, in exchange for roughly five times faster generation. We added the procedure and this comparison to the appendix (App. F.6, Multi-site reveal in the amortized Ising sampler).


Does the damping-dependent behavior in Table 8 remain when generation uses a single fixed order?

We added an experiment in which DIG-IN is trained and evaluated with a single fixed reveal order, with and without damping (App. F.5, Damping under a fixed reveal order). Damping helps under a fixed order as well. More seeds keep both modes of the ordered Ising model with damping than without. The random-order model nevertheless performs better and is less prone to mode collapse. With damping, it keeps both modes in every seed and reaches a higher effective sample size than the fixed-order model.


How were the reference log-partition values obtained for the nonzero-field amortized lattices and the fixed-composition alloy tables? Equations (134)–(135) define and transform the alloy quantities but do not describe how the sums were computed.

Amortized lattices: We noticed a minor error: The column header shows that the metrics are computed in expectation over inverse temperature β\beta and external field hh. However, the log⁡Z\log Z error was only computed in expectation over β\beta for zero-field ( h=0h=0), while other metrics were computed correctly in expectation over both β\beta and hh. To avoid any confusion, we removed the log⁡Z\log Z error from the table of the amortized runs.

Fixed-composition alloys. The reviewer is right that the appendix did not describe how the sums are computed. The fixed-composition ln⁡Zn(T)\ln Z_n(T) are computed with a standard high-precision Monte Carlo method from statistical physics, which samples all compositions evenly and therefore yields ln⁡Zn(T)\ln Z_n(T) for every composition at once (multicanonical sampling [1, 2] combined with replica exchange [3]). We added these details to the appendix and replaced "exact" with "reference" throughout.

We included these clarifications in the revised version of the manuscript.

[1] Berg & Neuhaus. "Multicanonical ensemble: A new approach to simulate first-order phase transitions."

[2] Wang & Landau. "Efficient, multiple-range random walk algorithm to calculate the density of states."

[3] Hukushima & Nemoto. "Exchange Monte Carlo method and application to spin glass simulations."


We hope these clarifications and additional experiments address the reviewer's concerns. We would like to thank the reviewer again for helping us improve our manuscript with the insightful comments and feedback. We are happy to answer any further questions during the rebuttal period.

Reviewer_E24BOpenReview ↗

Thanks for the response

Thanks the authors for the response, my questions are addressed. I keep my positive score of the paper.

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

This paper proposes Discrete Gibbs Iterative Neural Sampler (DIG-IN), a neural sampler for unnormalised distributions in discrete spaces, trained by a fixed-point iteration on its own importance-weighted samples. The key idea is Gibbs soft labels that Rao–Blackwellise the one-hot denoising target using the target conditionals at each site obtained by enumerating single-site energy differences. To stabilise the fixed-point iteration involving importance weighting, the authors use a weight-tempering scheme where a temperature is chosen to match a target effective sample size. The authors prove convergence of this iteration under some assumptions and extend the construction to general discrete interpolants between two distributions. The proposed method was evaluated in standard Ising and Potts models, an alloy phase diagram, and also in an amortised sampling setup where a single sampler tries to model multiple targets from the same family but with different configurations.

Critical strengths
  1. This paper is overall easy to read.
  2. The proposed algorithm is mostly well justified (especially the lower variance from Rao-Blackwellisation and convergence in the ideal case). It also outperforms the baselines (which I believe are sufficiently recent and strong) by large margins.
Critical weaknesses

My only major concern is that the paper lacks a detailed comparison with MDNS and PDNS. I think the proposed algorithm is slight modification of the weighted denoising cross entropy (WDCE) loss as in MDNS using Rao-Blackwellisation by computing soft labels. Although this is discussed in some places in the appendices (e.g., one sentence in App. D), it should be discussed in the main body with full details. I believe the fixed-point iteration part is from MDNS and the tempering corresponds to PDNS’s proximal step. The paper's contributions could be misleading without the discussion. In this sense, I also think the ablation study in App. F is important, and it would be helpful if the authors mention when the proposed algorithm becomes equivalent to MDNS or PDNS.

Related to this, clarifying when DIG-IN is a better choice than the two baselines and when it is not (e.g., energy is a black-box) could strengthen the paper. Additionally, although I think the experiments seem fair overall, MDNS and PDNS use half as many energy calls, and it would be helpful to report results under the same energy-call budget (e.g., by doubling the batch size).

I'm open to raising my score to 3 if the concern mentioned above is resolved.

(Minor points)

  • I believe a similar or the same ESS-based tempering scheme is considered in [1]
  • While Section 3.3 has the most math-heavy part, it doesn't seem to contribute much to the paper.
  • No baselines for the amortised training setup (5.2)

[1] Choi, Sanghyeok, et al. "Reinforced sequential Monte Carlo for amortised sampling." ICML 2026

Questions for authors
  1. I think PDNS already shows that the exact proximal iterations converge to the target. Is Theorem 3.2 similar to that, or does it have something more?
  2. Is the condition 2D(1−ν)<12D(1 − \nu) < 1 in theorem 3.2 satisfied in practice?
  3. When would you use general interpolants instead of masked diffusion?
4 comments
AuthorsOpenReview ↗

Reply to Reviewer cBQG (1/2)

We thank the reviewer for their careful reading and constructive feedback. Below, we address each point in turn and indicate the corresponding changes in the revised manuscript.


My only major concern is that the paper lacks a detailed comparison with MDNS and PDNS. […] Although this is discussed in some places in the appendices (e.g., one sentence in App. D), it should be discussed in the main body with full details. […] it would be helpful if the authors mention when the proposed algorithm becomes equivalent to MDNS or PDNS.

We thank the reviewer for this comment and agree that the relation to MDNS and PDNS should be stated explicitly in the main text. DIG-IN builds on the weighted denoising cross-entropy objective of MDNS, and its ESS-tempered weights correspond to the proximal step of PDNS. However, DIG-IN differs in two ingredients: Gibbs soft labels, i.e., the Rao-Blackwellized cross-entropy, and damping of the labels towards the previous model. The exact equivalences are:

  • MDNS = DIG-IN with hard (one-hot) labels, full importance weights (ν=1\nu = 1) and no damping (η=1\eta = 1).
  • PDNS = DIG-IN with hard labels, ESS-tempered weights and no damping (η=1\eta = 1).

This is how both baselines are implemented in our experiments: same codebase, network, sampler, buffer and optimizer. The comparisons in the main tables are therefore controlled ablations of these components. In the component ablation (App. F), "no Gibbs score" with tempered or with full weights corresponds to PDNS or MDNS with damping added.

The soft labels matter beyond variance reduction. Under a biased buffer (ν<1\nu < 1), hard labels fit the buffer's own conditionals, so at ν=0\nu = 0 every model is a fixed point and training does not move. Gibbs labels instead supply the target's conditional at every training context. This is why our convergence result also covers the label-only iteration (ν=0\nu = 0, Theorem B.1(iv)). Empirically, at matched budgets, hard-label training loses modes that the Gibbs labels keep, as shown by our ablation study. As suggested by Reviewer tX69, we added new controlled multimodal toy experiment (App. F.2 Mode collapse on a controlled multimodal target) which also confirms that DIG-IN outperforms both baselines.

Moreover, in the revision we add a paragraph "Relation to MDNS and PDNS" to Section 3 that states these equivalences, introduces the damping in the main text, and attributes the fixed-point iteration and the tempering to MDNS and PDNS, respectively. We will also revise the contribution statement accordingly.


Additionally, although I think the experiments seem fair overall, MDNS and PDNS use half as many energy calls, and it would be helpful to report results under the same energy-call budget (e.g., by doubling the batch size).

We thank the reviewer for this suggestion. We retrained MDNS and PDNS with twice as many fresh samples per exploration round (1024 instead of 512), with all other settings unchanged. This exactly matches DIG-IN's number of energy evaluations, counting each local table as one evaluation (App. F.3, Baselines at matched training budget). On the ordered Ising and Potts models, both baselines still collapse to a single mode on every seed.


Related to this, clarifying when DIG-IN is a better choice than the two baselines and when it is not (e.g., energy is a black-box) could strengthen the paper.

DIG-IN's Gibbs labels require the energy differences of all D(q−1)D(q-1) single-site changes. For local energies (lattice models, cluster expansions, pairwise interactions), computing these costs about as much as a single energy evaluation. Moreover, even at a matched budget of energy evaluations, DIG-IN outperforms both MDNS and PDNS in our newly added experiments. We therefore argue that DIG-IN is preferable to the baselines, particularly for multimodal targets.


I believe a similar or the same ESS-based tempering scheme is considered in [1]

We thank the reviewer for pointing us to this reference. We added the respective reference when introducing tempered importance weighting to credit their work.


AuthorsOpenReview ↗

Reply to Reviewer cBQG (2/2)


While Section 3.3 has the most math-heavy part, it doesn't seem to contribute much to the paper

We agree with the reviewer that additional experiments that show the usefulness of the more math-heavy construction that allows for bridging two arbitrary distributions would greatly benefit the paper. We, therefore, added additional experiments where we compute the free energy difference between two biomolecular systems, specifically the two alanine dipeptide tasks from [1] (solvation free energy and the free-energy difference between two conformational states). Notably, unlike FEAT [1], which requires training samples from both endpoint distributions, our method only requires samples from one endpoint and generates samples from the other. These experiments demonstrate both the effectiveness and practical utility of the general bridging construction in Section 3.3.

We kindly refer the reviewer to Section 5.5, Free energy estimation, for details on the experiments and results.

[1] He et al. “Feat: Free energy estimators with adaptive transport.” NeurIPS 2025


No baselines for the amortised training setup (5.2)

We agree that baselines for the amortized setting would strengthen the paper, and we are currently running them. However, these runs are expensive, so we cannot guarantee that they will finish within the rebuttal period. We will post the results here as soon as they are available and include them in the final version.


I think PDNS already shows that the exact proximal iterations converge to the target. Is Theorem 3.2 similar to that, or does it have something more?

PDNS solves a KL-regularized problem at every step, so its convergence follows from the theory of proximal-point methods.

DIG-IN’s Gibbs labels come from fixed-point map perspective and do not have a proximal point interpretation (it cannot be interpreted as a KL-regularized update). So the PDNS argument does not apply, and Theorem 3.2 establishes convergence for this fixed-point map. Only with one-hot labels does our update coincide with the PDNS proximal target. With Gibbs labels, each update is pulled directly towards the target. Convergence therefore holds even without importance weighting (ν=0\nu = 0), where the proximal iteration of PDNS does not move.


Is the condition 2D(1−ν)<12D(1-\nu) <1 in theorem 3.2 satisfied in practice?

As noted in the Limitations and future work paragraph, this condition does not hold throughout training. It is a sufficient condition, not a necessary one, and it only concerns random reveal orders. With a fixed reveal order, Theorem 3.2 guarantees convergence for every value of ν\nu. We nevertheless consider it a useful result, since convergence guarantees for fixed-point neural samplers of this kind are rare.

To show when the condition holds in practice, we added an experiment (App. F.4, The tempering exponent during training) that reports the exponent ν\nu selected by the ESS rule over the course of training. Early in training, ν\nu is small, so the importance weights are strongly smoothed and the condition is not satisfied. As the model approaches the target, ν\nu rises, and from 50k steps on it equals 11 at most steps, where the condition holds.


When would you use general interpolants instead of masked diffusion?

The general interpolant is needed when the goal is to learn a transport between two arbitrary distributions. Masked diffusion always starts from the fully masked state, so it can only transport from that fixed reference to a single target. One application is estimating the free-energy difference between two systems, where both endpoints are Boltzmann distributions (Section 5.5).


We hope these clarifications and additional experiments address the reviewer's concerns. We would like to thank the reviewer again for the detailed review, which has helped us improve our manuscript a lot. We are happy to answer any further questions during the rebuttal period.

I appreciate the authors' response. I'm happy to increase my score to 3.

Here are some remaining concerns/suggestions:

  1. PDNS tempers weights but chooses the exponent via a KL trust region or schedule, so "PDNS = DIG-IN with …ESS-tempered weights" describes your reimplementation, not the original PDNS, I guess. Correct me if I'm wrong.
  2. Regarding your "We retrained MDNS and PDNS with twice as many fresh samples per exploration round (1024 instead of 512)”, I guess the extra exploration samples just refresh the buffer faster rather than contributing to the updates, so I'm not sure this setup favors MDNS/PDNS. Could you also scale the batch size so the baselines' extra energy calls result in more samples per update?
  3. In the "Relation to MDNS and PDNS" paragraph in the revised manuscript, it might be nice to add a citation after "damps them toward the previous model ..." disclosing that the damping idea is similar to the one used in Blessing et al. (2026c).
  4. I think the placement of the citation of [1] may not be ideal. The ESS-based selection rule (Eq. 13) is the adaptive importance weight tempering of [1], isn't it? Also, correct me if I'm wrong.
  5. I still think at least some of Section 3.3 can go to the appendix. I'm also not sure interpolants are necessary for free-energy differences. Couldn't you train two masked samplers and take the difference of their ln Z estimates? Feel free to ignore this point.

Sorry for the late reply!

AuthorsOpenReview ↗

Follow-up reply to Reviewer cBQG

We thank the reviewer for the thoughtful follow-up and sincerely appreciate the increased score.


PDNS tempers weights but chooses the exponent via a KL trust region or schedule, so "PDNS = DIG-IN with …ESS-tempered weights" describes your reimplementation, not the original PDNS, I guess. Correct me if I'm wrong.

You are right. The original PDNS chooses its tempering step either by a fixed schedule or by a KL trust region. Our baseline keeps the PDNS objective and changes only the step selection to DIG-IN's ESS rule, so that the comparison isolates the effect of the labels and the damping. We will call it "PDNS with ESS-based step selection" in the revision.

We expect this change to matter little. Both rules pick the largest tempering exponent for which the tempered target stays close to the current model, and both compute this from the same importance weights. The only difference is how closeness is measured: PDNS uses the KL divergence, while the ESS corresponds to the χ² divergence. When the two distributions are close, which is what a high ESS target enforces, χ² is approximately twice the KL.


Regarding your "We retrained MDNS and PDNS with twice as many fresh samples per exploration round (1024 instead of 512)”, I guess the extra exploration samples just refresh the buffer faster rather than contributing to the updates, so I'm not sure this setup favors MDNS/PDNS. Could you also scale the batch size so the baselines' extra energy calls result in more samples per update?

Thank you for this suggestion. We have started new runs in which the batch size is scaled accordingly, so that the baselines’ additional energy evaluations translate into more samples per update. Unfortunately, these runs will not finish before the end of the rebuttal period, but we will include the results in the camera-ready version.


In the "Relation to MDNS and PDNS" paragraph in the revised manuscript, it might be nice to add a citation after "damps them toward the previous model ..." disclosing that the damping idea is similar to the one used in Blessing et al. (2026c).

We thank the reviewer for the comment. We added the citation.


I think the placement of the citation of [1] may not be ideal. The ESS-based selection rule (Eq. 13) is the adaptive importance weight tempering of [1], isn't it? Also, correct me if I'm wrong.

We agree. We made it more explicit that the adaptive importance weight tempering is equivalent to the one in [1].


I still think at least some of Section 3.3 can go to the appendix. I'm also not sure interpolants are necessary for free-energy differences. Couldn't you train two masked samplers and take the difference of their ln Z estimates? Feel free to ignore this point.

Section 3.3 makes two separate contributions: it establishes a way to construct interpolant models in discrete state spaces and proposes methods to train them directly from energy evaluations. You are correct that training two separate masked diffusion samplers and taking the difference between their estimates of ln Z can also yield a free-energy difference. However, when the two endpoint distributions are closely related, as in our solvent-vacuum example, learning a direct transport between them can be easier. The transport path is potentially much shorter than learning both distributions independently from a common reference distribution.


We hope this addresses your remaining questions, and we thank you again for your positive assessment of our work.

Reviewer_tX69OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a sampling method for discrete unnormalized lattice distributions. The model training does not depend on true samples, but iteratively proceeds by using importance-weighted samples from the current model. The method is shown to work for masked diffusion and transport between 2 distributions. Experiments over Ising, Potts and alloy models show more reliable sampling performance as compared to some state of the art samplers.

Critical strengths

Training does not depend on true samples, but uses Gibbs soft labels. Proposition 3.1 further supports this choice theoretically.

Experiments with amortised sampling demonstrate potential scientific usefulness for studying phase diagrams.

Critical weaknesses

The paper (abstract) aims at reducing mode collapse, but needs a more rigorous analysis of mode collapse, at least empirically. Toy problems may help (where modes can be controlled and are known for evaluation, e.g., MoG-40 in continuous distributions). Using appropriate metrics, such as NLL and R-NLL [FUND: Density Flow for Sampling Unnormalised Distributions, TMLR 2026] can help quantify mode-covering and mode-collapse behaviour.

The lattice sizes used in the experiments, i.e., 16x16 Potts model and 5x5x5 alloy model, are small. There is a class of algorithms aiming a scaling with size; e.g., [Multilevel Generative Samplers for Investigating Critical Phenomena, ICLR 2025] shows results on 128x128 lattice. So the scaling claim could be softened.

The novelty of the proposed approach needs clarification. E.g., [Zhang et al, ICML 2025] (refered to in the paper) also uses KL divergence loss, which takes a similar form as the loss proposed in this paper.

Questions for authors

Sec 1:

  • What is meant by “fixed point” sampling? Does it mean Gibbs sampling where x−dx^{-d} is fixed? Or does it mean the target distribution is fixed, and we want our model distribution to converge to it?

Sec 2:

  • A minor point: μθ\mu_\theta is not defined but used in Eq. (6).

Sec 3:

  • The proof that the sampler converges to the target distribution is not clear. (Theorem 3.2) Please elaborate.
  • Theorem 3.2: (iii) requires ν>1−1/2D\nu > 1-1/2D, i.e., for large lattices, ν\nu is already close to 1. Does that mean no tempering should be done?

Sec 4:

  • It would be good to discuss hierarchical samplers, such as [Multilevel Generative Samplers for Investigating Critical Phenomena, ICLR 2025], which aim at reliable sampling of large discrete lattices.

Sec 5:

  • An analysis of computational costs will help. The training steps for DIG-IN and FrOGS are the same, but that does not imply equal computational cost. It would be better if training and generation times and computational costs could also be reported and compared across different methods.

Further, please address the points in the "Critical weaknesses" above.

4 comments
AuthorsOpenReview ↗

Reply to Reviewer tX69 (1/2)

We thank the reviewer for their careful reading and constructive feedback. According to your review and suggestions, we have modified our manuscript with additional experiments and discussions. Below, we address each point in turn and indicate the corresponding changes in the revised manuscript.


The paper (abstract) aims at reducing mode collapse, but needs a more rigorous analysis of mode collapse, at least empirically.

We have added such an experiment in our revised manuscript. The target is the discrete analogue of the Gaussian-mixture benchmarks: 16 binary variables, 16 modes at known locations with different mode weights. The state space is small enough that the model density can be evaluated on all 2162^{16} configurations. Every metric is therefore computed exactly, including the NLL and R-NLL metrics of Kanaujia & Arora (TMLR 2026). We also report the total variation per basin and the number of modes hit.

Across all metrics, our method achieves the best performance compared to the baselines. We kindly refer the reviewer to Appendix F.2 (Mode collapse on a controlled multimodal target) for details on the experiment and the results.


The lattice sizes used in the experiments, i.e., 16x16 Potts model and 5x5x5 alloy model, are small. There is a class of algorithms aiming a scaling with size […] It would be good to discuss hierarchical samplers, such as [Multilevel Generative Samplers for Investigating Critical Phenomena, ICLR 2025], which aim at reliable sampling of large discrete lattices.

We thank the reviewer for highlighting this highly relevant work. We agree that RiGCS is an important example of scaling neural samplers to large lattices. We have added a discussion of it to the revised manuscript and softened our scaling claim accordingly.

We would nevertheless like to point out a distinction between the two approaches. RiGCS achieves its impressive scaling through a specialized, renormalization-informed architecture and a sequential training curriculum, while the sampler itself is still trained with the reverse KL divergence and REINFORCE. DIG-IN instead addresses the training principle: it is a general fixed-point learning framework that does not rely on problem-specific structure. Therefore, we see the two directions as complementary. Combining DIG-IN with multilevel architectures such as RiGCS, bringing together efficient multiscale representations and robust fixed-point learning, is a promising direction for future work.

Finally, DIG-IN is also not limited to lattice systems. In the revised manuscript, we additionally evaluate biomolecular sampling and extend the framework to interpolant-based transport for free-energy estimation, which shows that the training framework applies beyond lattice models.


The novelty of the proposed approach needs clarification. E.g., [Zhang et al, ICML 2025] (refered to in the paper) also uses KL divergence loss, which takes a similar form as the loss proposed in this paper.

We thank the reviewer for highlighting TCSM. We agree that it is closely related to our work and have expanded the discussion in the revised manuscript (App. D, Further related work).

Both methods exploit local information about a discrete target distribution, and the objectives are closely related: TCSM's distribution-based loss and our Gibbs-label loss both match single-site conditional distributions. They differ in setting and purpose. TCSM is a framework based on concrete-score ratios for data-driven training and for reward-guided fine-tuning of pretrained discrete diffusion models. DIG-IN trains a neural sampler from an unnormalized energy alone, with no data and no pretrained model, and uses Gibbs conditionals computed from that energy as training labels.

Because no target samples are available in our setting, the labels alone are not sufficient. DIG-IN therefore combines them with an iterative fixed-point procedure and ESS-tempered importance weighting, for which we provide convergence guarantees. It further supports discrete interpolants between arbitrary endpoint distributions. In the revised manuscript, we demonstrate biomolecular sampling directly from energy evaluations and apply the interpolant to solvation and conformational free-energy estimation.

AuthorsOpenReview ↗

Reply to Reviewer tX69 (2/2)

What is meant by “fixed point” sampling?

"Fixed point" refers to the training procedure, not to the sampler, so the second interpretation is the closer one. The target distribution is fixed, and we want the model distribution to converge to it.

Concretely, each training round draws samples from the current model. It computes, from the energy alone, the target's exact conditional distribution of each site given all the others, and fits the model to these labels. This defines an update map from the current model to the next one. The target is a fixed point of this map: if the model already matches the target, an update leaves it unchanged (Theorem 3.2(i)). Training repeats the update until the model reaches this fixed point, and Theorem 3.2 states when that is guaranteed.


A minor point: μθ\mu_\theta is not defined but used in Eq. (6)

We thank the reviewer for bringing this to our attention. We adjusted the revised manuscript accordingly.


The proof that the sampler converges to the target distribution is not clear. (Theorem 3.2) Please elaborate.

We thank the reviewer for raising this. We would appreciate it if the reviewer could point to the specific part that seems unclear. That would let us address it directly. In the meantime, we give a short overview of the argument below.

(1) The target is a fixed point. When the sampler matches the target, its generated samples follow the target distribution and the update leaves the sampler unchanged.

(2) The error converges to zero. For fixed reveal orders, the last site's conditional becomes exact after one update. Also, we show that when later sites converge, earlier sites will also converge recursively.

For general reveal orders, the errors accumulate across at most DD sites. Tempering reduces the resulting log-ratio error by a factor of 1−ν1-\nu. Together, these give

Δ(gk+1)≤2D(1−ν)Δ(gk),\Delta(g_{k+1})\le 2D(1-\nu)\Delta(g_k),

so the error converges to zero when 2D(1−ν)<12D(1-\nu)<1.

(3) Convergence of the sampler. Since all conditional predictors converge to the target conditionals, we have convergence of the generated distribution.

Full proofs are provided in Appendix B.2-B.6.


Theorem 3.2: (iii) requires ν>1−1/2D\nu>1-1/2D, i.e., for large lattices, is already close to 1. Does that mean no tempering should be done?

As noted in the Limitations and future work paragraph, this condition does not hold throughout training. It is a sufficient condition, not a necessary one, and it only concerns random reveal orders. With a fixed reveal order, Theorem 3.2 guarantees convergence for every value of ν\nu. We nevertheless consider it a useful result, since convergence guarantees for fixed-point neural samplers of this kind are rare and this opens a potential direction.

To show when the condition holds in practice, we added an experiment (App. F.4, The tempering exponent during training) that reports the exponent ν\nu selected by the ESS rule over the course of training. Early in training, ν\nu is small, so the importance weights are strongly smoothed and the condition is not satisfied. As the model approaches the target, ν\nu rises, and from 50k steps on it equals 11 at most steps, where the condition holds.


An analysis of computational costs will help. The training steps for DIG-IN and FrOGS are the same, but that does not imply equal computational cost. It would be better if training and generation times and computational costs could also be reported and compared across different methods.

We agree that equal step counts do not imply equal cost. For the alloys, we do not yet have a fair wall-clock comparison, because the runs used different hardware.

However, FrOGS is expected to be less efficient when growing with system size. FrOGS starts its transport from a learned autoregressive prior. Its training objective needs, at every state along every path, how the prior's probability changes when any single site is changed. An autoregressive model has no shortcut for this, because changing one site changes all the conditionals that come after it. Its per-step cost thus grows quadratically with the number of sites. DIG-IN's labels require only the target's single-site energy differences. For local energies, these cost about as much as one energy evaluation, and training needs one network pass per sample. This is what allows DIG-IN to scale to lattices with 2,500 sites. At generation time, FrOGS also has to draw from its autoregressive prior site by site before running its transport, so it is not cheaper to sample from than DIG-IN.


We hope these clarifications and additional experiments address the reviewer's concerns. We would like to thank the reviewer again for the insightful review, which has helped us largely improve our manuscript. We are happy to answer any further questions during the rebuttal period.

Thanks for providing additional clarifications. For reproducibility, it would be nice if the implementation details and, if possible, the code could be shared.

We thank the reviewer for the positive reception of our paper. We will publicly release the code for all experiments, including scripts to reproduce our results, upon publication.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We thank all reviewers for their time and their constructive and detailed feedback, which has helped us improve the paper substantially. Following the reviewers’ constructive feedback, we have provided more experiments, ablations, and discussions, which have largely improved our manuscript. Below we summarize the main changes.


New experiments

To highlight DIG-IN’s general applicability and the usage of the general interpolant construction, we add the following experiments:

  • Biomolecular ensemble generation (Sec. 5.4, App. E.4, F.8). We discretize the internal coordinates (bond lengths, angles and dihedrals) of alanine dipeptide and alanine tetrapeptide into bins and train our method from the energy alone, without any MD data. DIG-IN matches the MD reference on both peptides, including the Ramachandran basins of every residue, and reaches a higher ESS than the same network trained by maximum likelihood on MD data.
  • Free energy estimation with the general bridge (Sec. 5.5, App. E.5, F.7). We use our general discrete interpolant construction to compute the solvation free energy and a conformational free energy difference of alanine dipeptide using samples from only one endpoint and the energy of the other.

Ablations, further analyses and related works

  • Mode collapse on a controlled multimodal target (App. F.2). We use a target with 16 modes of known weights where all quantities are exact, and report forward and reverse KL divergences. At a budget that counts every energy evaluation, DIG-IN recovers nearly all modes, while the baselines lose or underweight light modes.
  • Baselines at matched training budget (App. F.3). The baseline MDNS and PDNS, retrained with the same number of target evaluations as DIG-IN, still collapse to a subset of modes.
  • Tempering exponent during training (App. F.4). We report the exponents selected by the ESS rule over the training iterations.
  • Fixed reveal order (App. F.5). We train DIG-IN with a single fixed order, the setting of Theorem 3.2(ii), and show that damping also helps there.
  • Multi-site reveal (App. F.6). We compare the faster block reveal used for the 50×50 lattice with exact single-site reveal. It trades ESS for speed and leaves mode coverage unchanged.
  • Related works. We added a new paragraph (Sec. 3.2) that makes the connection to MDNS and PDNS precise. Additionally, we extended the discussion of related work, including TCSM and multilevel samplers.

We hope these changes address the reviewers' concerns and are happy to answer any further questions during the discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.