acceptodds
Under review as a conference paper at ICLR 2027

Count the Groups, Not the Configurations: Exact Inference and Learning at Fixed Totals

Abstract

A fixed total couples every variable in an integer allocation model. For interactions determined by group totals, we turn this constraint into a computational resource: sum local allocation weights into group-count tables, remove a redundant count coordinate, and sample exact configurations from the resulting small distribution. We reuse these tables to score whole batches of changes to group membership through exact partition-function ratios. For two groups, compilation costs for variables, maximum capacity , and total , versus for an explicit total-and-signed-count recursion. A shared conditional allocation law also makes model comparison lossless at the count level. Controlled synthetic experiments with heterogeneous local weights yield – more effective samples per CPU second than the best tested positive auxiliary-field mixtures, and count-compressed membership scoring is faster than its expanded implementation. Learning studies separate this computational benefit from the choice of statistical objective. Exact residual calculations further reveal how an omitted interaction can amplify rare configurations, damaging importance sampling before substantially changing the bulk distribution. The result is a reusable count representation for sampling, comparison, and discrete structure refinement.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
probabilistic time series forecasting, hierarchical forecasting, count data, forecast coherence, distribution shift

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 2 reviews, above 85% of the others.
Reviewer_YihDOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

Suppose one wishes to sample from all possible allocations of rr objects into JJ boxes, where each box has a capacity of at most ci,i=1,…,Jc_i, i=1,\ldots,J. (Each of these numbers are integers). The authors consider a version of this problem which includes importance sampling terms inspired by exponential tilting and Ising/spin models. They observe that if the interaction between boxes depends only on some group totals in a manner which can be learned, then it is sufficient to first sample those totals and subsequently sample individual allocations conditional on them. They devise an algorithm which, conditioned on the group dependence, samples the allocations. Then they discuss how to learn the group dependence.

Critical strengths

Ultimately, the running time reduction is substantial, under the assumption. This is a major strength of the paper. The simplicity of Algorithm 1 is also a plus point, since its analysis was particularly easy.

Critical weaknesses

A critical weakness of the paper is that it fails to motivate practical scenarios under which the group dependence actually occurs. The experiments all come with the allocations that need to be learned, but there is no real-life scenario explained in the paper where the group dependence exists, but is not known beforehand. This applies, in particular, to situations where binary grouping is used.

In light of this, the contribution feels tangential to ongoing research. Furthermore, the introduction is tersely written and does not really bring out the separation between learning the group dependence, and then sampling once the group dependence is known: the writing needs to reflect this better.

Terms such as compiled interaction, remaining energy, macrostate, group-count law, source total, and membership swap are introduced rapidly and not within context.

These concerns have been addressed by the author's comments below

Questions for authors

a) Typographical error : in the graphics in page 2, the subscripts have not rendered, so it looks like x1x1 instead of x1x_1. Similar for A+,A−A+, A-.

b) I think the introduction and abstract contain a number of terms that require definition or explanation. For example, "A fixed total couples every variable in an integer allocation model" is a very terse statement, which can be replaced by "in an integer allocation model, under the constraint that the total number of slots are fixed, the occupancies of each slot become highly dependent on one another". I think you should be very clear about what you mean by "groups", and isolate the learning of group dependence from the actual sampling conditioned on the dependence itself.

c) I think you should check out Ankur Moitra et. al. "Steering diffusion models with quadratic rewards: a fine-grained analysis" : you will find some hardness results on generating from tilted distributions. Do see both the result and linked literature.

These questions have been addressed by the authors below. Since their addressal of the problem is sufficient, I have decided to raise my rating from 2 to 3.

d) Is there some theoretical result which can be provided on the accuracy of the ML scheme? Right now, I'm thinking why is this particular scheme chosen, so it could be validated not only by experiments but by basic results establishing asymptotic correctness and a rate of convergence.

Right now, I am marking this for a borderline rejection due to its writing primarily. I like the idea and I think it can be taken further, but need convincing that the rewrite required at this point for acceptance is not substantial.

Reviewer_XQR6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a group-count compiler for exact inference and learning in fixed-total discrete allocation models, avoiding explicit enumeration of exponentially many configurations. By representing allocations through macrostate variables and polynomial coefficient tables, it reduces computational complexity while preserving the original probability distribution. The method supports exact sampling, partition-function computation, and likelihood fitting, and uses importance sampling to account for residual interactions omitted from the main model. For group-structure learning, it efficiently evaluates multiple membership swaps by reusing the current macrostate distribution and computing exact likelihood ratios in batches.

Critical strengths
  • The key strength is the use of the fixed-total constraint to eliminate one count dimension and represent the problem through a compact macrostate. The group-count compiler preserves the original probability distribution while reducing the complexity of the two-group recurrence.
  • The paper extends the compiler beyond sampling by using the current macrostate distribution to evaluate many candidate membership-swap ratios in batches. This provides a coherent connection between exact inference and structure learning, avoiding the need to rebuild a full partition function for every proposed swap.
Critical weaknesses
  • The experimental validation primarily uses controlled synthetic models and known or specified model components, including unary fields and group-class constraints. Although these experiments enable exact reference probabilities and careful correctness checks, they provide limited evidence that the proposed framework will be effective in more complex or realistic settings.
  • Prototype learning, global representation modeling, and gating mechanisms are all existing techniques, so the approach does not constitute a completely novel breakthrough.
Questions for authors
  • The paper reports substantial efficiency gains for group-structured interactions, but the results for random dense controls are close to parity with the baseline. Could the authors clarify the structural conditions under which the group-count compiler is expected to provide meaningful computational gains, and whether these conditions can be identified before running the compiler?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.