acceptodds
Under review as a conference paper at ICLR 2027

Reimagining Meaningful Model Multiplicity

Abstract

Predictive model multiplicity is when multiple models in a hypothesis class disagree on their predictions while having similar overall error to the risk minimizer . This has substantial implications for algorithmic fairness: if such multiplicity exists, then there may be a model competitive with the risk minimizer of the target hypothesis class which does substantially better in terms of target fairness metrics; some argue that searching for such a less discriminatory model is a legal duty. Standard approaches formalize this duty via search over a Rashomon set of models, but this search is computationally expensive and does not easily provide generalization guarantees. We propose a reframing to “meaningful" model multiplicity that avoids both obstacle—rather than asking how to find the most fair model in we ask whether any can outperform on a target group . We provide efficient ensembling techniques over models which witness such multiplicity which simultaneously will outperform on all affected groups, and show that models which are suitably multiaccurate preclude meaningful multiplicity. Finally, for constraint-based fairness metrics for classification tasks, we show that building a multicalibrated predictor for the label probability and then postprocessing it to satisfy the fairness guarantees will Pareto dominate approaches based on searching the Rashomon set of classifiers.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
model multiplicity, algorithmic fairness, model ensembling, multicalibration, machine learning

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

This paper studies how model multiplicity can be used to improve subgroup predictive performance without explicitly enumerating a Rashomon set. It defines meaningful multiplicity through subgroup loss improvements, develops decision-list and linear-ensemble constructions for combining improvement witnesses, and establishes sufficient conditions based on multiaccuracy and groupwise self-orthogonality that exclude such improvements. It further applies existing multicalibrated postprocessing results to fairness-constrained classification. The contribution is theoretical, supported by population-level guarantees and a finite-sample analysis of the decision-list procedure.

Critical strengths
  • An interesting connection between model multiplicity and exploitable subgroup prediction error. The paper focuses on alternative models that improve predictive performance on identifiable groups, giving a concrete interpretation of when multiplicity can support model improvement.

  • Constructive guarantees for combining improvements across overlapping groups. The proposed ensemble procedures address the substantive difficulty that improving performance on one group can affect other overlapping groups. The decision-list construction approximately matches all supplied witnesses on their respective groups. The sequential linear construction provides a complementary approach without explicitly enumerating every group intersection.

  • The analysis connects population guarantees with finite-sample learning and downstream fairness. This work provides a finite-sample generalization guarantee for the decision-list procedure, giving statistical support to the finite-witness setting.

Critical weaknesses
  • The technical contribution beyond existing multigroup learning frameworks is insufficiently distinguished. The subgroup improvement criterion in Definition 2.1 closely resembles the certificates of suboptimality and multigroup optimality targets already studied in the bias-bounty framework (An Algorithmic Framework for Bias Bounties, FAccT 2022), which also supplies the underlying decision-list update mechanism. Moreover, the main downstream classification guarantee in Theorem 6.8 is explicitly inherited from prior multicalibrated postprocessing work.

  • Lack of empirical validation of the proposed algorithms. The paper does not demonstrate whether the methods achieve substantial improvements for underperforming subgroups at a reasonable computational cost, or how their performance changes when subgroup risks must be estimated from limited data and many groups overlap. Comparisons with existing multigroup learning methods and Rashomon-set selection would help establish whether the proposed approach offers a meaningful practical advantage.

  • The claimed superiority over Rashomon-set search depends on changing the admissible model class. The ensemble constructions explicitly allow predictors outside the original hypothesis class, while the cited downstream postprocessing procedure can also produce classifiers outside the convex hull of the comparator class. This limits the replacement claim when the original class encodes substantive requirements such as interpretability, model complexity, or deployment restrictions.

Questions for authors
  1. Which algorithmic or theoretical guarantees are new relative to the bias-bounty framework and prior multicalibrated postprocessing work? Please clarify which results require substantive new arguments beyond adapting these existing frameworks to model multiplicity.

  2. Can the authors provide empirical comparisons with existing multigroup learning methods and Rashomon-set selection to demonstrate subgroup improvements at a reasonable computational cost? How do these benefits vary with group overlap and the amount of data available for estimating subgroup risks?

  3. Does the claimed advantage over Rashomon-set search persist when both approaches must satisfy the same restrictions on the admissible model class? Please clarify how the replacement claim should be qualified when the original class encodes interpretability, complexity, or deployment requirements.

1 comment

Thank you for your review and thoughtful response! To your questions:

  1. You are right that our work is building off of existing literature, though the actual analysis of the bias bounties-esque framework we use is quite modified and has a new analysis---the approach in that paper does not by itself work for our setting; in that work the authors only track the models that are accepted into the model, while we have to track all of them while maintaining good sample complexity guarantees.

More genrally, we think of this work as a translational one: in our minds the interesting thing right now for ML-theory adjacent work is how it can be leveraged to do something useful, especially given how human-written technical work is getting swiftly suplanted. I.e., the thing that is valuable to us in this moment is making connections between theory and other areas and asking the right questions. The model multiplicity literature is actively shaping government policy and legal conversations around AI, so it seems very worthwhile to us to point out that the idea of predictive diversity within the Rashomon set that underpins it is itself suspect.

  1. Your point about empirical validation is well-taken; we have some simple simulated results we can include in an appendix. As mentioned in the response to reviewer MDo8, the statements hold, for instance, any time a minority class has a different relationship to the label than the majority. In terms of computational cost: the existing guarantees for Rashomon set search do not have sample complexity guarantees, because there can be an exponential number of models in the set, it is NP hard to search over, and the search cannot be reduced to a standard squared error oracle call as our approaches can be. In terms of comparison to other multigroup learning methods: when groups are disjoint they are generally equivalent to learning separate minimizers per group. When they are overlapping, however, we get 1/eps^2 complexity with low swap regret per group, while naive ensembling will not get you any form of swap regret. (ie you have to handle the intersections of groups, and naive approaches will lead to a sample complexity blow-up.)

  2. You are correct that the results do require increased complexity in the model, though we emphasize that our algorithms are all on the simple side (i.e., adding in a decision list/tree over models)---as we also mention in the discussion to reviewer MDo8, in practice even the most complex multicalibration processes do not greatly increase model complexity [1] Fig. 5 (e.g., looking at the decision tree of depth 6 the boosting process has created after round 6 to round, and how much the MSE and MSCE drop).

We can add discussion on this to the appendix with a specific example demonstrating how explainability can be maintained to some degree---i.e. generating a concrete example demonstrating the effect seen in Fig 5 of the above.

As a side note: one of the reasons why in practice one might not want to use a more complex model is not just that it may not be explainable, but also that very simple models tend to in practice be more robust to distribution shift etc---anecdotally, this is the case in medical settings where a model may have to be deployed across very different hospitals and patient populations. A nice thing about the multicalibration-style approaches we use here is that they are also more robust to these kinds of effects---this is the observation underpinning the omniprediction literature, and is one reason why these approaches have been used in medical settings [2].

[1] Globus-Harris et al, Multicalibration as Boosting for Regression, ICML 2023. https://arxiv.org/pdf/2301.13767

[2] Barda et al, Performing risk stratification for Covid-19 when individual level data is not available - the experience of a large healthcare organization https://www.medrxiv.org/content/10.1101/2020.04.23.20076976v1.full.pdf

Reviewer_2KBTOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper reframes model multiplicity in the context of fairness: instead of looking for an alternative model that is more fair while remaining in the Rashomon Set, simply search for a model that performs better on the selected subgroup of data. We can then mitigate multiplicity by proving that no such model exists within the hypothesis class. The authors finally propose algorithms with statisical guarantees to aggregate diverse models into a single one while ensuring that no singular model can be better than the whole on any group.

Critical strengths

The paper is a joy to read. It is very well written and every idea naturally flows to the next in a steady increase in complexity. The paper also clearly positions itself within the fairness and model multiplicity literature.

The theoretical results presented are correct and rigorous (except some minor details pointed out below). Guarantees are presented for the MultiplicityListEnsemble using probability distributions and finite sample estimates, which are necessary to make the algorithm practical.

Critical weaknesses

My problems with the article are minor.

  • Most theorems work with the theoretical distribution D\mathcal{D} and not finite samples. From this perspective, it makes sense to assign a separate model to each group if they happen to perform better than the risk minimizer h⋆∈Hh^\star \in\mathcal{H} on said group. There is no trade-off. However, when working with finite samples drawn from the distribution, there will be a trade-off seeing as fitting a separate model per group will likely overfit smaller groups. The authors aknowledge with fact at line 281 and also provide a finite sample bound (Theorem D.4), but this bound assumes that the group-wise models hg1,hg2,…hgkh_{g_1}, h_{g_2}, \ldots h_{g_k} are fixed before seeing the data, allowing them to use a union bound over a finite number of models. This can work if the models are fitted on training data and the MultiplicityListEnsemble is run on validation data, but it is important to clarify this point to the reader.

  • It would be better to attribute numbers to equations in proofs, otherwise it is hard for reviewers to ask for additional clarifications on specific steps or equations.

  • In the proof of Lemma C.1, one step consists of 'taking the derivative of both sides [of an inequality] with respect to lambda and then setting lambda equal to 0'. I don't think you are allowed to take derivative of two sides of an inequality while keeping the same inequality. The correct reasoning is that the right hand side is a quadratic function of lambda whose minimum occurs at lambda=0. Thus, the derivative of this function w.r.t to lambda must be non-negative at zero. The conclusion is the same, but the reasoning is different.

  • Theorem D.4 involves a parameter μ^\hat{\mu} (without a subscript ii) that it is not defined anywhere. I suspect it is a typo for μ∗\mu^*.

Questions for authors

Would it be realistic to obtain finite samples statistical guarantees when the group-wise models hgih_{g_i} are the empirical minimizers on their own group (hgi∈argminf∈HL(D,f,gi)h_{g_i} \in \text{argmin}_{f\in \mathcal{H}} L(D, f, g_i))? Assuming that ∣H∣|\mathcal{H}| is finite and groups are fixed beforehand and disjoint, could it possible to guaratee that, under certain conditions, the true risk of the MultiplicityListEnsemble model will not exceed that of the optimal rule list over models from H\mathcal{H}. This guarantee would not relate to the notion of ϵ\epsilon-meaningful multiplicity but would instead investigate the true risk of the learned model. Such a guarantee would improve the work since a practitionner could reject the rule-list model (although it gets rid of meaningful multiplicity) because its generalization suffers.

1 comment

Thank you for your close reading! We will clarify the exposition, e.g. for the necessity of having a separate validation set, renumbering equations/fixing the C.1 exposition, and typo-checking.

Re the finite sample guarantees of these algorithms and how it would compare to the optimal combination of models in H if you ran a separate risk minimizer for each of the groups:

  • If the groups are disjoint, you'll never have patches, so you will get a model that is equivalent to the ensemble of the risk minimizer. There can't be any better list over models: say that such a list existed, and that within it, group g is routed to model h_1 instead of the ERM h_g. Since g is disjoint from any other group, no point in g will be routed to a different model within the list. So h_1 must have group error lower than h_g, which is a contradiction (whp, since the total number of patches in the list will be small, so we can bound our sample complexity over them).

  • Even if the groups aren't disjoint, we still will be competetive. There the guarantee works because the total number of patches stays small, so we don't need to train on each individual intersection. As a side note, you also don't need finite H, it just needs to have finite VC/Natarajan/etc dimension.

Reviewer_MDo8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work explores the question of model multiplicity — where multiple models in the same hypothesis class can have the same/similar error to an optimal predictor — in the context of algorithmic fairness. A Rashomon set R(H) for hypothesis class H is a set of models that are ϵ\epsilon-close to the risk minimizer. In the context of algorithmic fairness, legal arguments have been made that it is a practitioner’s legal duty to (at least) perform a search over the Rashomon set for the most fair model, as no performance will be sacrificed. The authors introduce a refinement to the notion of multiplicity for (an optimal) model h* in the context of group fairness that they call “meaningful multiplicity,” defined as the existence of a model hgh_g whose performance is ϵ\epsilon-close to h* on a group g. With a focus on ‘meaningful’ multiplicity, instead of searching over R(H) for a more fair model, the authors propose to ensemble a set of group-optimal models hg1,…,hgkh_{g_1}, \ldots, h_{g_k} into an aggregate model that Pareto dominates h* and also no longer demonstrates ‘meaningful’ multiplicity. They further show this procedure is optimal for other group-based fairness metrics that may be in tension with accuracy, achieving an optimal fairness-accuracy tradeoff. Finally, because finding a set of group-optimal models hg1,…,hgkh_{g_1}, \ldots, h_{g_k} might be infeasible in the first place, the authors also develop a method to directly learn a ‘multicalibrated’ predictor that has no meaningful multiplicity and then postprocess this model for Pareto optimal fairness.

Critical strengths
  • The paper is well written and clear.
  • The arguments made are well supported by theory and appropriate formalism. Corresponding proofs are provided for all the important claims.
  • The proposed reframing of the multiplicity problem is well-motivated and shown to lead to a corresponding alternate (and more feasible) approach to maximizing fairness in the multiplicity setting.
Critical weaknesses
  • It would be useful to ground the theory in one or more concrete examples; as the purpose of the work is to make algorithmic fairness under multiplicity more tractable than searching over the Rashomon set, an example and/or discussion of when the key assumptions made in this work would or would not apply in practice feels an important missing part of the discussion.
  • Related to the above, a discussion of key limitations feels missing from this work.
Questions for authors
  • In Section 6, the focus is on group fairness metrics that compare performance on group g to performance on the overall population. Is this an inherent limitation in focus? What relevant notions of fairness, if any, do not fall into this class?
  • Regarding ensembling via decision lists, if I am understanding correctly, the main motivation for Algorithm 2 is that the naive implementation doesn’t work in the case where, e.g., group gig_i is contained within group gjg_j — “any decision list that has gjg_j appear as a predicate before gig_i will route points in gig_i to hgjh_{g_j} rather than hgih_{g_i}”. Isn't there a naive way to solve this routing issue? Would we not consider this just a "bug" where the incorrect function was chosen, as opposed to some fundamental issue? Is the underlying claim here that we are not actually able to detect cases where gjg_j appears as a predicate before gig_i? Perhaps an example would help me better understand how this is a fundamental issue.
1 comment

Thank you for your response! Regarding your comments: we will happily add a more thorough limitations sction.

The main critical assumption is that the setting is one where we can expand to a slightly more complex hypothesis class than we initially optimized over---this ensures that we can run the multicalibration process. In practice, multicalibration and similar approaches often only require a very small increase in model complexity, equivalent to adding a few layers to a decision tree. (See, for instance, the curves denoting the MSE and MSCE drops per round of boosting in fig 5 of [1])

More generally, the fact that (a) search outside of the Rashomon set gives you substantial gains over only searching the Rashamon set, and (b) that the Rash. set itself is intractable to search over is generally the case.

To see (a): consider a simple class like decision stumps, and say our distribution has a majority and minority subpopulation with very different label-feature relationships. E.g. you have a majority class where the label is perfectly described by a thresholding over feature A, but for the minority class, a threshold on feature B determines the label. Then for any ϵ\epsilon less than the weight of the minority class, the decision stump that learns on the minority class will not be contained in the Rashomon set. Of course, this is an extremely simple example, but the general principle holds any time you have a minority class whose label has a different correlation to the label than the majority class'. For (b), note that unless one were to make very restrictive assumptions about the smoothness of the class' parameter space, it is not clear how one could search the Rashomon set for group-improving models, and that moreover it is not clear what guarantees one can get from trying to approximate it by random sampling.

You also asked about the fairness definition in section 6, and how we compare group metrics to overall ones. You can convert this into a guarantee about any two groups and their weights, since each independently will have the guarantee hold with respect to the overall metric. We write it with respect to the average mostly as a matter of convenience, as it reduces needing to juggle multiple group densities. (As a side note, another way we could present the result is without the group weight term explicit, but then the guarantee would only be able to hold whp on sufficiently dense groups; here instead we make it an explicit tradeoff, but it's potatoes-potatahs. We do always need to consider group weight in some way, unless you don't want statistical guarantees.)

Re decision lists and alg 2: We could avoid the repairs if the grouping functions were closed under intersection. I.e. if we could guarantee that there is a patch for group A, group B, and group A \cap group B, for all groups and all intersections. Then we could greedily patch by order of group density and things would work out. But that's a very big assumption, and would greatly impact our sample complexity guarantees! If we don't assume this kind of closure, then the problem is that group B's patch/routing can mess up in the intersection of group B and group A, resulting in a patch. But since there can only be a polynomial number of such patchings, we don't have to take the same sample complexity hit that we would if we had to train 2^\vert \mathcal{G} \vert models like we would with the naive approach.

As an aside, one interesting thing that this means is that if you're running an audit on an ML model and you are repairing it according to the audit, then the order you choose to run your audit tests in matters. A nice new paper that thinks about this in the context of LLMs is [2] (In section 3).

[1] Globus-Harris et al, Multicalibration as Boosting for Regression, ICML 2023. https://arxiv.org/pdf/2301.13767

[2] Federova, Morosini, Podimata. Reasoning-Aware Multicalibrated Reliability Predictors for Language Models. 2026. https://arxiv.org/pdf/2610.26645

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.