acceptodds
Under review as a conference paper at ICLR 2027

Marginal Optimization Value of Evidence for Constrained Allocation

Abstract

Allocating limited resources under incomplete information requires choosing which evidence to acquire before choosing an allocation. Yet reducing uncertainty need not resolve which feasible allocation is preferable. We introduce Marginal Optimization Value of Evidence (MOVE), an acquisition rule that targets uncertainty in utility differences between feasible allocations. MOVE pairs predicted covariance reduction with the sensitivity of an entropy-smoothed allocation value, so that constraints shape which observations are useful. We establish an exact allocation-contrast interpretation and connect the implemented score to reoptimization value through explicit virtual-update and prediction-error terms, without assuming that the learned model is a coherent posterior. A masked-evidence model supplies query estimates, and an exact discrete solver returns the final allocation. Across 96 held-out allocation instances on public county data, spanning 32 county subsets and three objectives, MOVE achieves higher overall utility than all evaluated same-budget baselines, with component and objective-wise analyses supporting allocation-sensitive acquisition.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Allocation-sensitive evidence acquisition

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

The paper proposes an evidence-acquisition method that combines predicted utility-covariance reduction with the Hessian of an entropy-smoothed allocation value. Its aim is to acquire information that distinguishes competitive feasible allocations, rather than merely reducing prediction uncertainty. This is an interesting idea and a sensible objective to me, I also find the allocation-contrast interpretation is useful. The authors evaluated their approach using the publicly available dataset.

Critical strengths

I think the problem studied by the authors is important and I appreciated the focus on collecting information to make better decisions rather than simply reducing uncertainty.

Critical weaknesses

Despite solving an interesting problem I still have a few concerns. First, I find the clarity of the paper can be substantially improved. In particular, Section 2 compresses the entire sequential decision problem into one paragraph. Evidence blocks, masks, legal queries, and policy/source descriptions appear without a worked example connecting these objects. More importantly, it does not explicitly specify the joint probabilistic model of utilities and observations, the source of randomness in the expected objective, or the admissible adaptive query policy. Some definitions arrive later, but the reader must reconstruct the setup across multiple sections. The paper should first give a self-contained formulation and perhaps a concrete example. Additionally, the reproducibility of the work is also concerning. For instance, Appendix C.1 does not specify the training loss for the utility mean and covariance, as well as the reproducible model architecture and training configuration. It also says that temperature is computed from predicted utilities without specifying how exactly it is computed. The rank-four covariance parameterization also needs justification and sensitivity analysis. Without these details the evaluated system is not reproducible. The evaluation section can be significantly improved. For instance the baselines are not properly explained, for example, “Uncertainty” which is the strongest external comparator is particularly underspecified. More importantly, it is not clear whether Table 2 compares methods with the same predictors and, therefore, I am not sure if the reported results cleanly attribute improvements to MOVE’s acquisition score. I would also suggest to the authors to better connect their papers to the existing work. For example, the discussion omits particularly relevant work on bandit-based approaches such as Chen et. al. “Combinatorial Pure Exploration of Multi-Armed Bandits” (Neurips, 2014) or Fiez et. al. “Sequential Experimental Design for Transductive Linear Bandits” (Neurips 2019)

Overall, the paper contains a potentially useful acquisition heuristic, but needs much more clarity in explanations, positioning in the related work and having more reproducible implementation details.

Questions for authors

(see all my questions above)

3 comments
AuthorsOpenReview ↗

Response to Reviewer biad — Part 1 (Q1–Q4)

We are glad you found the focus on better decisions valuable. Below, we clarify the method, report the requested sensitivity analysis, and provide additional ablations with the predictor parameters fixed.

Q1: “First, I find the clarity of the paper can be substantially improved. In particular, Section 2 compresses the entire sequential decision problem into one paragraph. Evidence blocks, masks, legal queries, and policy/source descriptions appear without a worked example connecting these objects.” “Some definitions arrive later, but the reader must reconstruct the setup across multiple sections. The paper should first give a self-contained formulation and perhaps a concrete example.”

Answer: A worked example would help here. Consider selecting four of 15 counties within a budget, including at least two high-vulnerability counties. The policy description specifies the equity objective. Suppose two valid selections differ only in whether they include County A or County B. County A’s surveillance block is missing; its source description says it contains recent case counts and trends. The mask marks this block as unobserved, so querying it is legal. Revealing the values updates the utility predictions and may change which selection is preferred. After four queries, the solver returns the valid selection with the highest predicted total utility—the slate. We will place this example before the formal definitions in Section 2 (Appendix B.3, Table 7; Appendix B.4, Eq. (26)).

Q2: “More importantly, it does not explicitly specify the joint probabilistic model of utilities and observations, the source of randomness in the expected objective, or the admissible adaptive query policy.”

Answer: We agree that these three elements should be stated together. We will make them explicit in Section 2:

  • Joint model: P(u⋆,X∣Z)P(u^\star,X\mid Z) describes the joint distribution of candidate utilities u⋆u^\star and evidence blocks XX, given the objective, constraints and public metadata ZZ. MOVE learns to predict utilities and missing evidence from the observed inputs.
  • Source of randomness: The expectation is over the unknown utilities and evidence given the current observations, plus any randomization in query selection. In a fixed benchmark instance, a query reveals a stored measurement block.
  • Allowed query policy: Each query can depend on the observations collected so far and must select a still-hidden block. After KK queries, the method returns a selection that satisfies the allocation constraints.

Q3: “Additionally, the reproducibility of the work is also concerning. For instance, Appendix C.1 does not specify the training loss for the utility mean and covariance, as well as the reproducible model architecture and training configuration.” “Without these details the evaluated system is not reproducible.”

Answer: We appreciate the request for a precise training specification. The utility mean, covariance and missing-evidence predictions are trained jointly from randomly masked inputs; the text encoder remains frozen.

Loss. We minimize the joint negative log-likelihood of the utility vector and hidden measurements, conditioned on the observed blocks and allocation objective:

=-\mathbb E_{t,O}\log q_\theta \left(u_t^\star,x_{t,H}\mid x_{t,O},p_t\right).$$ Here $O$ is the randomly sampled set of observed blocks, $H$ contains the hidden blocks, and $p_t$ specifies the allocation objective. **Architecture.** A frozen **e5-mistral-7b-instruct** encoder embeds objective and source descriptions. Trainable projections combine these embeddings with observed values, masks, missing-block embeddings and public metadata. The output heads predict hidden measurements, utility means $\hat\mu_\theta\in\mathbb R^{15}$ and covariance $C_\theta=D_\theta+L_\theta L_\theta^\top$, where $L_\theta\in\mathbb R^{15\times4}$ and $D_{\theta,ii}=\operatorname{softplus}(\eta_i)+10^{-6}$ (Section 3.1; Appendix C.1). **Training.** We use batches of eight episodes for up to 20 epochs and select the model on the development set. We will add the implemented likelihood factorization, layer dimensions and optimizer settings to Appendix C.1. **Q4:** “It also says that temperature is computed from predicted utilities without specifying how exactly it is computed.” **Answer:** You are right that the numerical rule should be explicit. **We set the temperature using the standard deviation of the predicted utilities:** $$\tau=\operatorname{clip} \left(0.35 \operatorname{sd}(\hat\mu), 0.035, 0.12\right).$$ The standard deviation is computed across candidates. We recompute $\tau$ after each real query and keep it fixed while scoring the possible next queries (Section 3.2; Appendix C.1). We will add the numerical rule to Appendix C.1.
AuthorsOpenReview ↗

Response to Reviewer biad — Part 2 (Q5–Q8)

Q5: “The rank-four covariance parameterization also needs justification and sensitivity analysis.”

Answer: This is a useful sensitivity check. We followed your suggestion by varying the number of covariance factors.

Table R1. Allocation utility with different numbers of covariance factors.

Factor rankUtility ratio ↑95% CI
20.868[0.835, 0.899]
40.878[0.852, 0.902]
80.853[0.824, 0.882]

Four factors achieve the highest mean utility among the tested settings, supporting the default choice. The 95% CIs are obtained by bootstrapping the 32 test county subsets, keeping each subset’s three objectives together and each trained model fixed (Appendix C.3).

Q6: “The evaluation section can be significantly improved. For instance the baselines are not properly explained, for example, “Uncertainty” which is the strongest external comparator is particularly underspecified.”

Answer: Good point. Here is the exact procedure. Uncertainty uses its own ridge-regression utility predictor and a fixed variance-based score for each evidence block. The block score is the mean training-set variance across its available features. At each step, every unrevealed candidate–block pair receives its block’s score, and the highest-scoring legal query is selected with deterministic tie-breaking. Revealed evidence is added to the observed state; after the query budget is exhausted, the predictor estimates utilities and the shared exact solver returns the allocation. We will add these implementation details to Appendix C.2.

Q7: “More importantly, it is not clear whether Table 2 compares methods with the same predictors and, therefore, I am not sure if the reported results cleanly attribute improvements to MOVE’s acquisition score.”

Answer: We agree that the contribution of the acquisition score needs a controlled test. We therefore ran additional ablations using the same trained predictor with all its parameters fixed. The initial observations, four-query budget and final solver are also shared.

Table R2. Ablations with the predictor parameters fixed.

Method or variantUtility ratio ↑Jaccard ↑
MOVE0.8780.486
w/o allocation sensitivity0.8350.441
w/o covariance contraction0.8300.449

With the same frozen predictor, MOVE outperforms both acquisition-score ablations on utility and Jaccard, supporting the contribution of its query-selection rule.

Q8: “I would also suggest to the authors to better connect their papers to the existing work. For example, the discussion omits particularly relevant work on bandit-based approaches such as Chen et. al. “Combinatorial Pure Exploration of Multi-Armed Bandits” (Neurips, 2014) or Fiez et. al. “Sequential Experimental Design for Transductive Linear Bandits” (Neurips 2019)” “Overall, the paper contains a potentially useful acquisition heuristic, but needs much more clarity in explanations, positioning in the related work and having more reproducible implementation details.”

Answer: These are relevant work, and we appreciate the pointers. Both papers study how to acquire information to identify the best decision, but their observation models differ from ours. [Chen et al. (2014)] collect repeated reward samples from individual items. Our queries reveal a missing feature block once, and utility labels cannot be queried (Section 2; Appendix B.3). [Fiez et al. (2019)] assume a specified linear relationship between measurements and decision values. MOVE uses a learned model to predict how revealing a feature block changes candidate utilities. We will cite both papers and explain these connections and differences in Section 5.

Thank you for responding to my review. I will maintain my current score for now.

Reviewer_obDTOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper is written in the context of resource allocation problems under limited information. The paper proposes how to decide which evidence to acquire to reduce uncertainty in utility differences between feasible allocations. The proposed acquisition rule combines predicted covariance reduction with the sensitivity of an entropy-smoothed allocation value. The paper also shows results on a retrospective county-level public-health allocation task and finds that the proposed method achieves a mean utility ratio of 0.878 relative to the oracle, compared with 0.814 for the strongest evaluated same-budget baseline.

Critical strengths
  1. The work is situated at an interesting intersection of constrained resource allocation and active feature acquisition and provides useful theoretical results connecting the acquisition score to expected one-step improvements in allocation value, with explicit bounds accounting for prediction, approximation, and smoothing errors.
  2. The work provides results on a real-world public-health dataset combining county-level CDC surveillance and vaccination records with social vulnerability, chronic-condition, and Census demographic data.
Critical weaknesses
  1. The paper is hard to follow and problem formulation describes multiple terms like slate and masks, hidden blocks without describing what they refer to. It is also unclear what a query means in the context of the empirical experiments. While some of these terms are eventually described, a concrete example earlier in the paper would make the formulation easier to understand.
  2. The empirical evaluation is limited to one retrospectively constructed public-health allocation setting. Although the paper considers three objectives, these share the same data sources and allocation structure. Evaluation in another domain would help establish how broadly the proposed method applies.
Questions for authors
  1. Are there more applications where this problem of allocation and query uncertainty can be applied? Specifically, how would the benefit of this method change across different problem contexts?
  2. All acquisition methods are evaluated with four additional queries. How does MOVE’s advantage change with the query budget?
2 comments
AuthorsOpenReview ↗

Response to Reviewer obDT

We appreciate your positive assessment of the theoretical results and the public-health evaluation. Your suggestions prompted us to add a concrete example, a Bike Sharing evaluation and tests at several query budgets.

Q1: “The paper is hard to follow and problem formulation describes multiple terms like slate and masks, hidden blocks without describing what they refer to. It is also unclear what a query means in the context of the empirical experiments. While some of these terms are eventually described, a concrete example earlier in the paper would make the formulation easier to understand.”

Answer: We agree that a concrete example would make the public-health setting easier to follow. A query reveals one missing group of measurements for one county; the final allocation chooses four counties under resource and coverage constraints. For example, querying a county’s surveillance block reveals its recorded case counts and trends, which can change its predicted utility and whether it belongs in the selected group. The mask records which blocks are already visible, and slate denotes the final selected group. Under the equity objective, at least two selected counties must have high social vulnerability (Appendix B.3, Table 7; Appendix B.4, Eq. (26)). We will introduce these terms through this example before the formal definitions in Section 2.

Q2: “The empirical evaluation is limited to one retrospectively constructed public-health allocation setting. Although the paper considers three objectives, these share the same data sources and allocation structure. Evaluation in another domain would help establish how broadly the proposed method applies.” “Are there more applications where this problem of allocation and query uncertainty can be applied? Specifically, how would the benefit of this method change across different problem contexts?”

Answer: We took up this suggestion and applied MOVE to the UCI Bike Sharing dataset, including baselines from the paper. The task uses hourly rental counts across 361 test days; each instance selects four of 15 hourly slots under a synthetic budget and peak-hour constraint. MOVE achieves the highest utility ratio and Jaccard score among all nine evaluated methods.

Table R1. Allocation quality on the Bike Sharing task.

MethodUtility ratio ↑ [95% CI]Jaccard ↑
MOVE0.925 [0.913, 0.937]0.604
Covariance-trace0.918 [0.912, 0.935]0.598
Random0.913 [0.902, 0.923]0.569
Ridge0.815 [0.804, 0.824]0.356
DFL0.708 [0.696, 0.722]0.235
Uncertainty0.864 [0.851, 0.877]0.448
EDDI0.879 [0.867, 0.890]0.463
Dynamic MI0.889 [0.874, 0.904]0.490
ACO0.816 [0.806, 0.826]0.356

Intervals use seven-day block bootstrap resampling. MOVE, Covariance-trace and Random use the same trained predictor and four-query budget. Covariance-trace selects queries by total variance reduction.

Across applications, the benefit depends on whether missing evidence can change which feasible selection is best (Proposition 1; Table 1). For illustration, possible applications include selecting machine inspections before allocating a repair budget, or choosing pilot measurements before deciding which projects to fund.

Q3: “All acquisition methods are evaluated with four additional queries. How does MOVE’s advantage change with the query budget?”

Answer: Good point. We extended the comparison with Uncertainty, EDDI and Dynamic MI to one, two, four and eight queries. MOVE achieves the highest allocation utility at every tested budget.

Table R2. Allocation utility at different query budgets (higher is better).

Method1 query2 queries4 queries8 queries
Uncertainty0.7920.7900.8140.843
EDDI0.7940.7990.7970.814
Dynamic MI0.8030.8030.8090.820
MOVE0.8520.8650.8780.890

Even with one query, MOVE achieves higher mean utility than all three baselines with eight. These results show that its advantage is not specific to the four-query setting.

Thanks to the authors for the rebuttal! It has adequately addressed my questions. Thus, I remain with the positive score.

Reviewer_MwT8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces MOVE, which combines uncertainty reduction with allocation sensitivity, to choose more useful evidence under constraints. The authors connect the score to decision value and report higher overall utility than same-budget baselines.

Critical strengths

The exact allocation-contrast interpretation gives the score a clear theoretical basis. With the same query budget, MOVE achieves a utility ratio of 0.878 versus the best external baseline’s 0.814, supporting its practical value.

Critical weaknesses

Baselines use different predictors, so gains may come from prediction rather than acquisition. Experiments use only 15 candidates and enumerate feasible allocations, leaving larger-scale practicality unproven.

Questions for authors

Do the ablations keep the predictor and its weights fixed? Do virtual updates use a single-point approximation or multiple samples?

1 comment

Thank you for the careful review and for highlighting MOVE’s theoretical basis and practical value. We have added frozen-predictor ablations and benchmarks with up to 50 candidates, and clarify the virtual updates below.

Q1: “Baselines use different predictors, so gains may come from prediction rather than acquisition.” “Do the ablations keep the predictor and its weights fixed?”

Answer: Good suggestion. The MOVE-family ablations in the submitted paper use the same predictor architecture (Table 3). Following your suggestion, we have now repeated the ablations using the same trained predictor with all its parameters fixed. The initial observations, four-query budget and final solver are also shared.

Table R1. Ablations with the predictor parameters fixed.

Method or variantUtility ratio ↑Jaccard ↑
MOVE0.8780.486
w/o allocation sensitivity0.8350.441
w/o covariance contraction0.8300.449

MOVE achieves the highest utility and Jaccard under this control. In particular, removing either allocation sensitivity or covariance contraction lowers both metrics, directly supporting the contribution of the acquisition score (Eq. (5)).

Q2: “Experiments use only 15 candidates and enumerate feasible allocations, leaving larger-scale practicality unproven.”

Answer: We agree that the computational cost at larger candidate counts is important. For a fixed number of selected candidates, MOVE’s allocation computations grow polynomially with the candidate count, and feasible allocations are cached once and reused across queries (Section 3.4). Our additional benchmarks confirm that these computations remain practical up to 50 candidates:

Table R2. Allocation computations with supplied predictions.

CandidatesFeasible allocationsBuild feasible set (ms)Score all queries + final solve (ms)
157020.68.5
3015,51611.938.1
5080,275139.9174.2

The benchmark selects four candidates and measures one query-scoring step plus the final exact allocation. These component timings use supplied predictions and exclude neural evaluations of hypothetical responses. At the original 15-candidate size, a complete four-query episode takes a median of 9.2 s.

Q3: “Do virtual updates use a single-point approximation or multiple samples?”

Answer: Good question. We used a single-point approximation at the model’s predicted response mean in the reported experiments. For each possible query, we insert that response into a temporary copy of the observed input and rerun the same model to estimate the updated covariance (Section 3.3; Appendix C.1).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.