acceptodds
Under review as a conference paper at ICLR 2027

Certification, Not Search: Deployment-Time Model Sharing under Hard Per-Cell Risk Contracts

Abstract

Deployment under hard per-cell risk contracts must decide which trained predictors to store, which cells may share them, and when local fallback is required, while minimizing serialized model, payload, and routing cost. Existing procedures commit to a sharing structure during training: multi-task models fix parameter sharing, mixture-of-experts learns routing, and clustering fixes partitions before deployment. Deployment still requires a separate decision: whether a sharing structure can be certified against a per-cell risk budget from the calibration data actually available. We propose Certified Shared Decoder Selection (CSDS), a deployment-time framework that builds a finite candidate portfolio, certifies every cell–action pair from calibration data, and selects the shared subset and the per-cell assignment minimizing a declared additive coding-cost proxy over the certified action set. Because the contract is cell-separable, certifying the individual cell–action pairs certifies every deployment plan simultaneously. This property of the contract, rather than of the testing procedure, is what makes a deployment portfolio certifiable at all: testing the plan family directly yields a bound looser than the entire risk budget, certifying nothing. Experiments on five 30-seed tabular contract settings, predictor-family controls, and a four-domain transformer-adapter deployment show that CSDS reduces deployment cost against both all-local deployment and a calibration-tuned clustered router, with no observed contract violation. Matched controls that hold the candidate set and the fitted predictors fixed identify certification, not subset search, as the scarce resource, while the certificate family and its failure-budget allocation remain live deployment decisions. Code is provided in the supplementary material.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
risk control, multiple testing, model sharing, model deployment, model selection, uncertainty quantification

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_w4WEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses the problem of selecting and sharing pretrained models under risk constraints imposed on individual data groups (contract cells). Using a separate (local) model for each cell can lead to substantial storage and deployment costs, whereas sharing models across multiple cells can potentially reduce these costs but introduces the risk of degraded performance on individual data groups. In distributed systems, using a single model for all data groups does not guarantee compliance with risk constraints, while deploying local models for every group may be expensive. The authors formulate the problem of determining which models can be shared across which groups and at what cost. The proposed approach operates at deployment time and uses a pool of trained candidate models. Based on certification results, the method selects a subset of shared models and assigns them to admissible cells. The authors propose a method for grouping cells and assigning a common model to each group based on K-means clustering.

The considered problem is decomposed into two subproblems: selecting a subset of models and assigning the selected models to individual cells. The first subproblem is combinatorial, and the authors propose using exhaustive search or greedy algorithm to solve it. For the second subproblem, the selection of the least expensive certified model for each cell is proposed.

The practical significance of the proposed approach lies in its potential to reduce the costs of storing and deploying models while preserving formal risk constraints for the data groups being served. The problem under consideration is how to certifiably allocate a limited set of models across groups under risk constraints with minimal deployment costs. This direction is relevant to machine learning systems in which reliability requirements are specified separately for different data segments and storage and serving resources are limited.

The scientific novelty of the work lies in formulating the problem of certified model sharing under per-group risk constraints and proposing the CSDS (Certified Shared Decoder Selection) procedure for it. The certification techniques used (Hoeffding risk bounds), candidate-pool generation methods, shared-model subset selection methods, and model to cell assignment procedure themselves are not novel.

The problem considered is practically relevant and the paper presents a mathematically well-structured framework for certified model sharing under per-cell risk constraints. Its value is the formalization of the deployment-time model selection problem and the theoretically justified decomposition of certification into individual cell-action pairs. However, the methodological novelty remains insufficiently established. In particular, the experiments do not demonstrate the evidence that the proposed architecture offers an advantage due to the new algorithmic solutions, and not due to the problem statement itself, the certification requirement, or the quality of the candidate pool.

Overall, the paper presents a coherent and potentially useful framework, but the extent of its contribution beyond the integration of existing certification and optimization techniques remains open. It would be more convincing if it included more evidence demonstrating the clear advantages of the CSDS, as well as a more systematic analysis of the factors determining its efficiency.

Critical strengths
  1. A rigorous and practically relevant problem formulation.
  2. A clear decomposition of the CSDS framework into distinct stages: candidate pool generation, statistical certification of the cell-model assignments, and cost-aware selection of the model for its assignment to cells.
  3. A theoretically justified certification strategy. The authors show that the risk constraints and the objective function are separable across contract cells. This allows the feasibility of an deployment configuration to be guaranteed by certifying individual cell-action pairs.
  4. A formal proof of optimality for the assignment subproblem. The authors establish that, for a fixed subset of shared models, the optimal assignment can be computed in polynomial time by independently selecting the least costly feasible model for each cell.
Critical weaknesses
  1. It remains unclear how significant the combination of known methods in the proposed CSDS architecture is. In particular, the paper could better clarify which aspects of the proposed approach go beyond a direct application of existing principles of statistical certification and risk control to the model-sharing problem.

  2. The benefit of subset optimization over simpler selection policies is not demonstrated. The experimental results show no advantage of the proposed subset-selection procedure over a simpler fixed-K policy on the same candidate pool: both achieve identical results in the reported comparisons. So, the practical or theoretical benefit of optimizing the number of shared models remains unclear. The authors should discuss about the main contribution whether lies in the subset-selection algorithm itself or in candidate generation and certification.

  3. The paper does not address how the proposed certification guarantees and deployment decisions would behave under distribution or conceptual shift. In particular, it remains unclear how the statistical guarantees should be adjusted, whether model assignments should be updated dynamically, and how such updates could be certified under risk constraints.

  4. The performance of CSDS appears to depend substantially on the quality and diversity of the source candidate pool generated before deployment. Since CSDS selects and assigns existing predictors rather than improving their predictive capabilities, the CSDS quality is inherently constrained by the candidates. However, the paper does not investigate how the CSDS performance depends on the performance of the candidate models, particularly weak models. The additional experiment should distinguish the effect of individual model performance and effect of model diversity within the pool.

  5. It is unclear how the proposed cost function in Equation (2) relates to actual costs in industrial deployment. In particular, the relationship between the theoretical coding-cost objective and practical considerations such as storage requirements, inference latency, computational resources, and serving costs should be clarified. Without such an analysis, the CSDS grounding to industrial deployment infrastructure remains uncertain.

  6. In the CSDS, a cell is admitted for service only if its local model is certified. If the local model fails certification, the cell is referred for further processing even when a shared model has been successfully certified for that cell. Such restriction is the authors’ choice and is not a shortcoming of the work. Nevertheless, an ablation study relaxing this restriction would be valuable. In particular, how allowing a cell to be served whenever at least one shared model is certified for it affects service coverage, deployment cost, and the resulting risk guarantees.

  7. Insufficient practical guidance on where the architecture's improvements should be focused. From the perspective of researchers seeking to apply CSDS to their own problems, it remains unclear which component deserves the greatest attention: candidate-pool generation, the choice of statistical certification procedure, or the optimization algorithm used to select the shared models. The paper would benefit from more explicit practical recommendations, supported by the experimental results, on which components are most influential and are likely to yield the greatest improvements.

Questions for authors

How important is the combination of known methods in the proposed CSDS architecture?

What measurable benefits does the subset optimization procedure provide?

What assumptions underlie the statistical certification guarantees when the data distribution changes after deployment? Does CSDS require recertification before updating model assignments?

Since CSDS selects and assigns existing predictors rather than improving their predictive capabilities, how sensitive are its results to the quality and diversity of the initial candidate pool?

What is the intended interpretation of the cost function in Equation (2) in terms of actual deployment infrastructure? Is it intended as a theoretical proxy for deployment complexity?

See Critical Weaknesses for more details.

2 comments
AuthorsOpenReview ↗

Response to Reviewer w4WE: component ranking, admission ablation, candidate-pool analysis

Dear Reviewer w4WE,

Thank you for the careful and well-organized review. It prompted three analyses that we think settle your main questions: a ranking of components, an ablation of local admission, and a test of candidate-pool quality and diversity.

Components and subset optimization (W7, W2/Q2). In order of size (bits/sample, Covertype unless stated): candidate columns 1.7–2.2 (choosing by tune risk rather than cost gain; random columns 0.9–1.2); certificate and admission 21–35% of cost on Covertype and 37–41% on Electricity; subset search at most 0.18; Holm instead of Bonferroni at most 0.03. The fixed-K policy in the submitted paper coincided with exact enumeration, so it could not show a benefit; we replace it with controls that do not enumerate: greedy keeps 99.6% of the exact solver's gain over the tuned router (+2.699 vs +2.708; same subset in 123/150 runs; ~3 ms vs ~0.1 s) and a frozen ranking 94.4%. The main contribution therefore lies in candidate generation and certification.

Known methods (W1/Q1). We do not claim a new statistical tool. What goes beyond direct application is cell separability: m(K+1)m(K+1) pair-level bounds cover every deployment plan at once, so subset choice and assignment can use the calibration data freely, with local admission inside the same failure budget.

Local admission (W6). Thank you for this ablation; it produced a clear improvement. Letting a cell without a certified local predictor be served by a certified shared action, with voluntary referral, is never worse than the original rule at referral prices 4, 8, 16: 0.12–0.65 bits lower on Covertype and 0.76–1.21 on Electricity, with no held-out violation among the newly served cells (1–2% and 7% of cells); only certified pairs are served, so the guarantee keeps its form. We recommend it as the default.

Candidate pool (W4/Q4). Selection by lowest tune risk costs 1.7–2.2 bits more than ranking by cost gain (3.0 on Electricity), random columns 0.9–1.2 more, and a coverage rule meant to add diversity performs like the tune-risk rule: what repairs the dependence is the ranking criterion, not diversity. Deliberately degraded candidates are a natural extension.

Cost function (W5/Q5). Eq. (2) is a description-length proxy: bits for the shared and local models, the routing index and the data, plus a referral price κ\kappa; latency, compute and serving cost are not modeled. On Covertype CSDS stays ahead at every price we tried; on Electricity up to κ≈16\kappa\approx16, and 18.7 with voluntary referral.

Distribution shift (W3/Q3). The guarantee assumes calibration and deployment data are exchangeable within each cell; after a shift, assignments are recertified on fresh labelled calibration data, which is cheap (m(K+1)m(K+1) bounds, millisecond selection). Descriptively, served pairs have median margin τ−\tau- test risk ≈0.15 (Covertype) and 0.23 (Electricity), and the smallest uniform risk shift producing a first violation is ≈0.06 and 0.11; this measures slack, not robustness.

Main results. On Covertype, CSDS beats the tuned router in all 30 seeds of each setting, by 2.68–3.70 bits, with no held-out violation under the headline certificate. Electricity, with the least calibration data per cell (about 130 samples), is the closest comparison and the one where the tighter certificate helps most (37–41%). After recomputing the main tables from the cached calibration artifacts, CSDS uses about 56% fewer bits than all-local deployment there (all 30 seeds) and about 15% fewer than the tuned router (+0.708 bits, 25 of 30 seeds), and is level with the certified MoE router (0.090 bits, 18/30).

Thank you again for the detailed engagement.

Thank you for the response and the additional analyses. The new experiments and clarifications address several of my concerns and make the paper better and the contribution of CSDS clearer. The additional comparisons of subset-selection strategies clarify the role of the optimization component that subset-selection procedure is not the primary source of the reported improvements. So the main contribution lies primarily in candidate generation and statistical certification, rather than in the subset-selection algorithm itself. If so, my concerns are related to the methodological novelty, contribution to the new knowledge and the limited evidence regarding candidate-pool quality and practical deployment costs. The paper is well-organized but I'm not sure that the theoretical and methodological ideas are sufficient for the conference. So, my decision is still between weak rejection and borderline.

Reviewer_3KxYOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies deployment-time model sharing under a risk limit for every contract cell. The proposed method, CSDS, builds and freezes a finite pool of fitted predictors, certifies all cell–action pairs including local fallback, and selects a shared subset and cell assignments using a coding-cost objective. Simultaneous bounds allow the calibration data to be used for selection while keeping the population risk guarantee. I agree that the separation of candidate fitting, certification, and deployment selection is useful. The inclusion of local admission in the failure budget and the reporting of referral coverage are also strengths.

My concerns are followed below.

Critical strengths

The work addresses a practically meaningful issue by combining finite-sample risk guarantees with deployment-time model sharing. I appreciate the clear problem formulation, the explicit consideration of deployment costs, and the careful experimental analysis. In particular, the matched-candidate comparisons and certification ablations provide useful insights into where the performance gains originate.

Critical weaknesses
  • Proposition 1 appears to follow directly from standard simultaneous confidence bounds and the separability of per-cell constraints. The authors should clarify what new statistical insight this result provides beyond existing multiple-testing and risk-controlling selection methods.
  • The fixed-K control is not defined well enough to support the main claim that subset search is unnecessary ('not search' in the title?). If it optimizes subsets at each size and then selects the best size on calibration, the tie with unrestricted optimization is an identity. A fixed ranking or genuine greedy policy on the same certified pool would test the claim more directly.
  • Theorem 3 omits the within-group risk slack epsilon in the certification condition and drops the routing bits from the cost boundary. The conditions allow a within-group risk above the contract, and fixed routing-width jumps can make partial sharing better than either endpoint. The stated conditions appear insufficient. For example, (ρ0,ϵ,Δ,r,τ)(\rho_0,\epsilon,\Delta,r,\tau)=(0,0.25,0.5,0.1,0.22) satisfies the assumptions, yet the within-group risk may exceed \tau, preventing certification. Moreover, the break-even condition omits routing costs, potentially invalidating the claimed optimality of full sharing versus all-local deployment.
Questions for authors
  1. What is the exact fixed-K algorithm: a ranking frozen on tune, or subset optimization at each size followed by a calibration-selected size? Can you report an actual non-enumerating control with its cost gap and runtime? And explain the meaning of the 'not search'.

  2. Can you correct Theorem 3 to include epsilon and the fixed routing-width cost, and either allow partial sharing or state the simpler cost model used by the theorem?

  3. [Extension suggestion] The Bonferroni-corrected Hoeffding bounds appear substantially conservative, especially for cells with limited calibration samples. Since certified coverage is central to the proposed method, the authors should further investigate whether less conservative procedures, such as Holm step-down testing [1], can improve coverage while maintaining the same family-wise error guarantee. Moreover, increasing the candidate pool also increases the multiplicity penalty, potentially reducing the number of certified actions. This trade-off deserves further analysis.

Despite my concerns, I find the overall framework promising. I would be inclined to raise my score if the authors can satisfactorily address the theoretical issues, clarify the novelty of the certification results, and better justify the fixed-K comparison.

[Reference]

Holm, S. (1979). A Simple Sequentially Rejective Multiple Test Procedure. Scandinavian Journal of Statistics, 6(2), 65–70. https://www.jstor.org/stable/4615733

2 comments
AuthorsOpenReview ↗

Response to Reviewer 3KxY: non-enumerating controls, corrected Theorem 3, tighter certificates

Dear Reviewer 3KxY,

Thank you for the careful and constructive review. Your three points led to three concrete gains: (1) a direct test of "not search", in which policies that do not enumerate retain 94–99.6% of the exact margin; (2) a corrected Theorem 3, verified by exact enumeration with no disagreement; (3) a certificate improvement, prompted by your Holm suggestion, that lowers cost by a further 21–41% within CSDS.

Fixed-K and "not search". Thank you for pointing out that fixed-K as we ran it (best subset per size, then best size on calibration) coincides with exact enumeration minus the empty set; it cannot test the claim, and we no longer use it as evidence. On the same certified pool and 150 paired runs (5 settings × 30 seeds), the controls you suggest give, as gain over the tuned router in bits/sample [95% CI]: exact +2.708 [1.650, 3.469]; greedy +2.699 [1.636, 3.462] (99.6% of exact; same subset in 123/150 runs; ~3 ms vs ~0.1 s); a ranking frozen on the tune split +2.558 [1.482, 3.321] (94.4%). So "not search" is now measured: the policies differ by at most 0.18 bits against a 2.7-bit margin, and the margin comes from certified coverage.

Proposition 1. We agree it is the union bound plus per-cell separability, not a new statistical result. Its value is structural: the bound covers all m(K+1)m(K+1) cell–action pairs including local fallback, so any subset and assignment chosen from calibration data inherits the guarantee.

Theorem 3. Thank you for the precise counterexample: reading it as (ρ0,ε,Δ,r,τ)=(0,0.25,0.5,0.1,0.22)(\rho_0,\varepsilon,\Delta,r,\tau)=(0,0.25,0.5,0.1,0.22), the hypotheses hold but ρ0+ε>τ\rho_0+\varepsilon>\tau. Corrected: (i) if Δ−ε>2r\Delta-\varepsilon>2r and ρ0+ε+2r≤τ<ρ0+Δ\rho_0+\varepsilon+2r\le\tau<\rho_0+\Delta, the certified graph equals the planted groups on the certificate event, and any τ\tau outside this window admits an outcome that breaks it; (ii) given that structure, sharing gg of GG groups costs J(g)=m(B+ncloc)−gs+m⌈log⁡2(g+1)⌉J(g)=m(B+nc_{loc})-gs+m\lceil\log_2(g+1)\rceil with s=(m/G−1)B−(m/G)n(csh−cloc)s=(m/G-1)B-(m/G)n(c_{sh}-c_{loc}); the optimum is g=Gg=G if s>ms>m, g=0g=0 if s≤0s\le0, otherwise a minimizer that can be partial, since the routing width jumps at g=1,2,4,8g=1,2,4,8. Exact rational enumeration (a grid check, not a proof) agrees everywhere: on 93,422 parameter points the original hypotheses held but the structure failed at 2,349 (all with ε>0\varepsilon>0), the corrected window at none; on 155,550 cost-grid points the corrected rule matches the exhaustive optimum, with partial sharing strictly optimal at 1,202.

Holm and multiplicity. Your suggestion led to the largest improvement we found. The exact binomial bound with the same Bonferroni correction, together with serving a cell without a certified local predictor by a certified shared action or referring it voluntarily, lowers cost within CSDS by 21–35% (Covertype) and 37–41% (Electricity) at referral prices of 4–16 bits, in all 30 seeds of every setting, with one held-out violation among about 16,500 served pairs (test error 0.401 vs τ=0.40\tau=0.40, consistent with δ=0.05\delta=0.05). Holm at the same δ\delta adds up to 0.03 bits. These compare CSDS variants, as the baselines use the original certificate. On pool size, a 4× larger family (2 → 11.4 candidates) raises the radius by only 7% while Covertype W1 cost falls from 5.14 to 2.34.

Main results. On Covertype, CSDS beats the tuned router in all 30 seeds of each setting, by 2.68–3.70 bits; over all five settings the gain is +2.708 (145/0/5), with no held-out violation under the headline certificate. Electricity, with the least calibration data per cell (about 130 samples), is the closest comparison and the one where the tighter certificate helps most (37–41% within CSDS). After recomputing the main tables from the cached calibration artifacts, CSDS uses about 56% fewer bits than all-local deployment there (all 30 seeds) and about 15% fewer than the tuned router (+0.708 bits, 25 of 30 seeds), and is level with the certified MoE router (0.090 bits, 18/30).

We hope these changes resolve the points you raised on the theory, the novelty of the certification result and the fixed-K comparison, and we would be glad to discuss anything that remains unclear. Thank you again for the detailed and constructive review.

Thank you for the detailed response and additional analyses. The new greedy and fixed-ranking controls provide meaningful support for the empirical “not search” claim. The revised Theorem 3 also addresses the missing risk slack and routing costs, including cases where partial sharing is optimal.

My remaining concern is the strength of the theoretical contribution. Proposition 1 follows from standard simultaneous bounds and cell separability, as acknowledged in the response. The corrected sharing analysis also replaces the original full-sharing versus all-local boundary with a more qualified result. I therefore view the main contribution as the deployment formulation, the structural observations, and the empirical evidence. The “not search” conclusion should remain tied to the evaluated settings.

In short, the corrections and additional experimental controls improve my assessment, and I will raise my rating from 2 to 3. Thank you again for the careful and constructive response.

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

This paper proposes Certified Shared Decoder Selection (CSDS), a deployment-time framework that certifies cell–action pairs under hard per-cell risk constraints and then selects a cost-efficient shared predictor portfolio. The key idea is that cell-separable contracts allow certification of only (m(K+1)) primitive cell–action pairs rather than exponentially many deployment plans, while experiments show substantial cost reductions over local and clustered baselines with no observed violations in the main settings.

Admittedly, I am not familiar with the paper's research area and had informed the AC I will not be able to provide a meaningful review at the beginning of the review process.

Critical strengths

I like the clean decomposition between candidate generation, certification, and portfolio selection. Moreover, the controlled experiments show that the main gains come from certified candidate coverage rather than combinatorial search.

Critical weaknesses

A weakness is that the empirical study is still relatively narrow—mostly tabular CART settings plus one four-domain LoRA experiment—so it is not yet clear how broadly the conclusions extend to larger or more heterogeneous real-world deployment problems.

Questions for authors

None

1 comment
AuthorsOpenReview ↗

Response to Reviewer zeVP: direct controls for the coverage finding, and scope

Dear Reviewer zeVP,

Thank you for the review. We are glad that the split into candidate generation, certification and selection came across clearly, and that you picked up the main finding: the gains come from certified candidate coverage, not from combinatorial search over subsets.

We have tested that finding more directly since. Policies that do not enumerate subsets keep 99.6% (greedy, about 3 ms vs 0.1 s) and 94.4% (a ranking fixed on the tune split) of the exact solver's gain over a tuned router. On Covertype, growing the pool of certified candidates from 2 to about 11 lowers cost from 5.14 to 2.34 bits per sample.

On scope, we agree the evidence is limited to a class of problems: cell-separable contracts with per-cell calibration data. It covers five tabular contracts (four from Covertype, one from Electricity), deeper CART and MLP portfolios, and one four-domain LoRA adapter deployment. Larger and more heterogeneous deployments are the natural next test. Within this class, all 150 paired runs beat per-cell deployment, 145 of 150 beat the tuned router (all 120 Covertype runs, by 2.7–3.7 bits), and none of the 15,567 deployed pairs exceeded its cap on held-out data.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.