PaperlayerView exact revision on arXiv

Source-linked AI summary

Should I stay or should I show? Learning to selectively disclose information

Carlotta Giacchetta, Alessando Bogani, Cesare Barbera, Giovanni De Toni, Michele Caprio, Andrea Pugnana, Andrea Passerini

arXiv:2609.39818v1cs.LG

TL;DR

Support information can be costly and may not improve human decisions, motivating a way to choose when to disclose it. The paper learns budget-constrained disclosure policies by estimating human decision risks and thresholding their Value of Information. Experiments and user studies report benefits from selective disclosure, while identifying limits on what the studies establish.

  • Problem

    Support information can improve, fail to improve, or harm human decisions, while acquiring it may be costly.

    Source: §1 · p. 1

  • Method

    The paper learns a budget-constrained disclosure policy by estimating regime-specific human risks and thresholding their Value of Information.

    Source: §1 · pp. 1–2

  • Results

    Experiments show the learned policy identifies instances where disclosure improves human decisions, and user studies find it either outperforms or performs on par with human-selected disclosure.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The optimal policy discloses cases with the largest positive Value of Information, targeting disclosure where expected human decision-risk reduction is greatest.

    Source: §1 · p. 1

    The framework treats support information as an indivisible block, and the user studies neither identify which task characteristics reduce advice adherence nor establish a causal link between adherence and policy effectiveness.

    Source: §6 · p. 10

Abstract

from arXiv · show

In many high-stakes settings, human decision-makers can acquire support information before making a decision. However, acquiring information is costly, and disclosure may fail to improve human decisions or may even impair them. We tackle this problem by studying selective disclosure, i.e., the problem of learning when to reveal support information to a human decision-maker under a budget constraint. We first show that the optimal policy is a threshold rule on the Value of Information (VoI), i.e., the expected reduction in human decision risk induced by disclosure. Since VoI is unknown in practice, we estimate the regime-specific human risks and bound the possible degradation of the resulting plug-in policy relative to lack of disclosure, as well as its regret relative to the optimal policy. Experiments on benchmark datasets show that selective disclosure outperforms both no disclosure and full disclosure, regardless of whether the support information is beneficial or harmful. Two user studies show that human-AI team performance can improve when disclosure is led by our learned policy and not human-selected, although this advantage varies across tasks. A counterfactual benchmark, which replaces participants' predictions with a machine-learning prediction when disclosure occurs, suggests that these differences might depend on lower adherence to advice when the information is automatically provided rather than self-requested.

1 INTRODUCTION

Support information can improve, fail to improve, or harm decisions, while acquiring it may be costly. The paper introduces budget-constrained selective disclosure based on expected human decision-risk reduction and evaluates it theoretically and in experiments and user studies.

1 INTRODUCTION

  • LSD learns when to disclose support information under a budget, using its expected effect on human decision risk rather than informativeness alone.The paper also frames the effect of disclosure on decision loss as a causal inference problem.Source: §1 · p. 1
  • The optimal policy discloses cases with the largest positive Value of Information (VoI), while the paper establishes conditions for estimating VoI from data.The paper derives guarantees for plug-in policies relative to both no disclosure and the optimal policy.Source: §1 · pp. 1–2
  • Experiments on synthetic and real-world data show that the learned policy identifies instances where disclosure improves human decisions.Two user studies find that policy-selected disclosure either outperforms or performs on par with human-selected disclosure; a counterfactual benchmark suggests lower adherence to automatically provided advice may partly account for results.Source: §1 · p. 2
  • For classification under 0-1 loss, the paper introduces a class-wise estimator of regime-specific risks and evaluates it on synthetic and real-world datasets.The contributions also include two user studies comparing policy-selected and human-selected disclosure in classification tasks.Source: §1 · p. 2

2 PRELIMINARIES

The setup treats each case as baseline information, a target, and support information that may be disclosed or withheld. The learning objective concerns a decision-maker drawn from a population, using observations from both disclosure regimes.

2 PRELIMINARIES

  • A case comprises baseline information X, target Y, and support information S, which may be an ML prediction and is disclosed in full or withheld.The human predicts the target label, and the action space is set to the target space.Source: §2 · p. 2
  • The dataset records decisions under both disclosure regimes for each case, but not both potential actions for the same human-case episode.At least two distinct decision-makers are assigned to each case, one per regime.Source: §2 · p. 3
  • The objective is to learn a disclosure policy for a decision-maker drawn from population π, rather than for a fixed individual.Actions may differ across decision-makers even for the same case.Source: §2 · p. 3

3 LEARNING TO SELECTIVELY DISCLOSE

The paper formulates disclosure as a budget-constrained decision problem, characterizes the optimal VoI threshold, and develops causal estimation, policy guarantees, and a class-wise risk estimator for 0-1 loss.

3 LEARNING TO SELECTIVELY DISCLOSE

  • The LSD objective minimizes expected decision loss by choosing which cases receive support information, subject to a budget limiting the expected disclosure rate.No disclosure and full disclosure are the policies that never and always reveal support information.Source: §3 · p. 3

3.1 OPTIMAL DISCLOSURE POLICY

  • The optimal policy discloses cases whose VoI reaches a nonnegative, budget-dependent threshold, revealing support only when expected risk reduction is sufficiently high.The threshold is based on the (1 − B)-quantile of VoI, with λ* set to the maximum of zero and that quantile.Source: §3.1 · p. 3

3.2 A CAUSAL INFERENCE INTERPRETATION

  • VoI is the negative population-average causal effect of disclosure on human decision loss, and its identification requires assumptions including unconfoundedness, positivity, consistency, SUTVA, and human-decision-maker exchangeability.Regime-specific conditional risks are identifiable under these assumptions, reducing VoI estimation to a standard conditional average treatment effect problem.Source: §3.2 · p. 4

3.3 GUARANTEES FOR THE PLUG-IN POLICY

  • Plug-in policies can perform worse than no disclosure, but the degradation is bounded by VoI estimation error and the region where the policy intervenes.The oracle-regret bound scales with estimation error and the mass of the disagreement region; errors where policies agree do not affect relative performance.Source: §3.3 · p. 5

3.4 ESTIMATING THE VALUE OF INFORMATION FOR 0-1 LOSS

  • For 0-1 loss, ClassWise estimates class-specific error patterns using a probabilistic classifier and regime-specific multi-head classifiers, then combines the estimated risks to obtain VoI.The class-wise decomposition addresses how a single scalar error probability can blur class-specific effects on VoI.Source: §3.4 · pp. 5–6

4 EXPERIMENTAL EVALUATION

The evaluation tests whether ClassWise learns effective disclosure policies, improves human-AI team performance, and changes how participants interact with support information. Across benchmark datasets and two user studies, results vary with the task and with whether participants follow advice.

4 EXPERIMENTAL EVALUATION

  • The benchmark evaluation compares ClassWise with Confidence, Random, FullDisc, and NoDisc using test-set accuracy across synthetic and real datasets.The user studies also compare learned disclosure with participants’ self-selection and measure advice adherence.Source: §4.1 · p. 6

4.1 EXPERIMENTAL SETTINGS

  • The two user studies included 545 participants, each classifying 20 items; support was selected by participants or ClassWise, under low or high budgets.Budgets allowed support on at most 6 or 14 items; the studies covered Email and ImageNet-16H.Source: §4.1 · p. 7

4.2 EXPERIMENTAL RESULTS

  • Across four datasets, ClassWise outperforms simple baselines; whenever the ImageNet-16H budget binds, it remains ahead of Confidence.Figure 2 reports budget-accuracy curves; at budgets of .80 or higher, Confidence matches ClassWise as nearly every instance is disclosed.Source: §4.2 · pp. 7–8, Figure 2 · p. 8
  • On Email, NoDisc outperforms FullDisc (Acc ≈.79 vs Acc ≈.76), while ClassWise peaks at B = .30 (Acc ≈.81).At higher budgets its accuracy declines and plateaus slightly below NoDisc, but remains within NoDisc’s confidence band.Source: §4.2 · p. 7
  • On ImageNet-16H, FullDisc achieves Acc ≈.85 versus NoDisc Acc ≈.77; ClassWise surpasses FullDisc at B = .70 and peaks at ≈.86 at B = .90.ClassWise is always different from Random; Confidence matches it only at B ≥ .80, where nearly every instance is disclosed.Source: §4.2 · pp. 7–8
  • In ImageNet-16H, human accuracy is higher under MP than HP (0.80 ± 0.11 vs 0.77 ± 0.13); Email shows no significant MP–HP difference.For the budget comparison, ImageNet-16H accuracy is higher under HB than LB (0.83 ± 0.10 vs 0.74 ± 0.13), while Email shows no significant difference.Source: §4.2 · p. 8, Figure 3 · p. 8
  • Advice adherence is lower under MP than HP in Email (0.85 ± 0.16 vs 0.96 ± 0.15), but not significantly different in ImageNet-16H or between budgets.Figure 3 reports accuracy and adherence by budget and support condition, with separate panels for the two tasks.Source: §4.2 · p. 8, Figure 3 · p. 8
  • In the counterfactual benchmark, MPCFT accuracy exceeds MP significantly in Email (0.85 ± 0.10 vs 0.82 ± 0.12), but not in ImageNet-16H (0.82 ± 0.10 vs 0.80 ± 0.11).The cross-task pattern suggests lower adherence may partly contribute to the absence of an MP advantage in Email; the authors conclude that deployment should attend to how people receive unrequested information.Source: §4.2 · p. 9

5 RELATED WORK

The paper situates selective disclosure within human-AI decision-making and budget-constrained policy learning. Its distinction is that the human remains the decision-maker while the policy allocates information, with VoI treated as a causal quantity.

Human-AI Decision-Making

  • Unlike learning-to-defer systems, LSD leaves the decision with the human and allocates support information; it frames disclosure as a causal intervention and identifies conditions for learning VoI.The authors describe LSD as, to their knowledge, the first to establish when VoI is identifiable from data collected under both regimes.Source: §5 · p. 9

Policy Learning

  • Theorem 1 parallels optimal policy learning under a capacity constraint, but disclosure acts on the decision-maker and identification requires decision-makers to be exchangeable across regimes.The related policy-learning threshold uses the larger of zero and a budget-dependent quantile.Source: §5 · p. 9

6 CONCLUSION

LSD selects disclosures by estimated Value of Information under a budget, and the conclusion reports gains in two user studies alongside important limits on the framework and evidence. The appendix proves the VoI threshold rule and documents research practices and materials.

6 CONCLUSION

  • ClassWise-selected disclosure improved team accuracy over human-selected disclosure in two user studies, with effectiveness varying across tasks; a counterfactual benchmark suggests adherence may help explain the differences.The benchmark replaced participants’ predictions with machine-learning predictions when disclosure occurred.Source: §6 · pp. 9–10
  • For 0–1 classification, the class-wise VoI estimator proved effective on synthetic and real benchmarks, while plug-in policies received guarantees on degradation and regret.The optimal policy thresholds VoI at a budget-dependent level, disclosing where expected benefit exceeds that threshold.Source: §6 · p. 9
  • The framework treats support information as an indivisible block, and the user studies neither identify task features behind adherence nor establish that adherence causes policy effectiveness.Selecting which parts of support information to reveal and studying these adherence questions are left for future work.Source: §6 · p. 10

AI USE STATEMENT

  • The authors used generative AI for code support and writing feedback, manually reviewed the code, and take responsibility for the final content.ChatGPT Astra Pro was used for proofreading and presentation clarity.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

  • The user-study protocol was judged to pose no risk to participants’ well-being or rights and therefore did not require full ethics review.The stated reviewing body was the institution’s Research Ethics Committee.Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

  • The authors report that proofs, additional experimental results, user-study details, implementation scripts, collected study data, and statistical analysis scripts are available.The materials are provided through an anonymized repository.Source: REPRODUCIBILITY STATEMENT · p. 10

A.1 PROOF OF OPTIMAL DISCLOSURE POLICY

  • Under a continuous VoI distribution, the optimal policy discloses when VoI(x) ≥ λ*, with λ* equal to the larger of zero and the (1−B)-quantile.This is the budget-constrained threshold rule established in the appendix.Source: §A.1 · pp. 15–16

A.2 PROOF OF IDENTIFIABILITY OF THE VALUE OF INFORMATION

The appendix establishes that, under assumptions A1–A5, regime-specific risks and their difference—the conditional Value of Information—are identifiable from observed data. Its proof connects each regime’s observed conditional losses to the corresponding potential risks.

A.2 PROOF OF IDENTIFIABILITY OF THE VALUE OF INFORMATION

  • Under assumptions A1–A5, the regime-specific potential risks are identifiable from observed data, and their difference identifies the conditional Value of Information.The proposition states the identifiability result for the potential risks under disclosure and no disclosure.Source: §A.2 · pp. 16–17
  • The proof uses unconfoundedness, overlap, consistency, and exchangeability of decision-makers to link observed losses within each regime to the corresponding potential losses.Overlap ensures the observed conditional expectations are defined; exchangeability links the assigned decision-maker to the population pool.Source: §A.2 · p. 17
  • Once each regime-specific risk is identified by its observed conditional expectation, subtracting the disclosure risk from the no-disclosure risk identifies VoI(x).The paper writes the difference as r0(x) − r1(x).Source: §A.2 · p. 17

A.3 PROOF OF THE NO-DISCLOSURE DEGRADATION BOUND

The appendix explains the no-disclosure degradation bound for plug-in policies and when it can certify safety. Degradation can arise when the estimator overstates VoI on disclosed cases, especially when the threshold provides no positive margin.

A.3 PROOF OF THE NO-DISCLOSURE DEGRADATION BOUND

  • The bound’s estimation-error cost applies only to disclosed cases, while the policy’s estimated gain offsets some of that cost by at least the threshold per disclosed case.The error contribution vanishes as the disclosure rate shrinks.Source: §A.3 · p. 17
  • The bound certifies no degradation when estimation error is at most the estimated threshold times the square root of the realized disclosure rate.Otherwise, the bound’s right-hand side may be positive and no-degradation is not guaranteed.Source: §A.3 · p. 18
  • A plug-in policy can underperform no disclosure only if it discloses cases with negative true VoI, which requires overestimating VoI by more than the threshold on some disclosed cases.This risk is greatest when harmful disclosure is common and the budget is not binding, leaving the threshold at zero.Source: §A.3 · p. 18

A.4 PROOF OF THE ORACLE REGRET BOUND

This appendix develops the oracle-regret bound, class-wise risk decomposition, and causal interpretation, then documents synthetic and Email experiments.

A.4 PROOF OF THE ORACLE REGRET BOUND

  • Oracle regret depends on disagreement-region mass and VoI estimation error there; errors where both policies disclose or withhold do not contribute.The final bound replaces disagreement-region mass with the upper bound B* + B-hat, which can be loose when the policies largely agree.Source: §A.4 · p. 18

A.5 PROOF OF CLASS-WISE DECOMPOSITION

  • Under 0–1 loss, conditional risk decomposes over labels into class probabilities and class-conditional human error probabilities.The proof obtains this by expanding the error event over possible target labels and applying the probability chain rule.Source: §A.5 · p. 19

B A CAUSAL INFERENCE INTERPRETATION

  • The causal interpretation treats disclosure as a binary treatment and defines VoI as the population-average reduction in human decision loss, subject to identification assumptions.These include conditional unconfoundedness, overlap, consistency, SUTVA, and exchangeability of decision-makers across regimes.Source: B A CAUSAL INFERENCE INTERPRETATION · pp. 19–20

C.1 SYNTHETIC EXPERIMENTS.

  • Synthetic experiments use generated binary and four-class data, simulate human decisions with logistic regression, and compare ClassWise, T- and S-Learners, and an uncertainty-only Confidence baseline.The Confidence baseline ranks by predictive uncertainty regardless of whether support information helps.Source: §C.1 · pp. 20–21

C.2 REAL DATA EXPERIMENTS - EM A I L.

  • Email experiments use 1,000 labeled messages and human responses from 300 participants; training augmentation yields up to 14 variants per email.Human responses are summarized by regime for each email, and the training set also uses back-translated paraphrases.Source: §C.2 · p. 21

C.3 REAL DATA EXPERIMENTS - IM A G ENE T-16H.

The ImageNet-16H experiments use noisy images, conformal prediction sets, and human responses from separate disclosure studies, with cross-study exchangeability treated as plausible but unestablished.

C.3 REAL DATA EXPERIMENTS - IM A G ENE T-16H.

  • The dataset contains 1,200 images from 16 categories at phase-noise level 110, with conformal prediction sets derived from a fine-tuned VGG-19 classifier.The highest available noise level is used as the hardest regime.Source: §C.3 · p. 22
  • No-disclosure responses come from one study, while disclosure responses come from another; the latter showed prediction sets as suggestions participants could disregard.For each image and miscoverage level, the analysis retains the response associated with a prediction set of the corresponding size.Source: §C.3 · pp. 22–23
  • Training augmentation composes three randomly sampled geometric operations from a 14-operation pool while preserving image labels and human response distributions.Figure 4 illustrates the available operations on a cat image at noise level 110.Source: §C.3 · p. 23, Figure 4 · p. 23
  • 0.7713 full-set-trial accuracy versus 0.7647 no-disclosure accuracy offers descriptive evidence of baseline similarity, but does not establish exchangeability.The comparison uses the same images, and a non-discriminative prediction set is not identical to withholding support.Source: §C.3 · p. 24

D USER STUDIES

The user studies compare human-selected and policy-led assistance across tasks and budgets, while the extended results describe sampling and analysis procedures rather than reporting outcomes in these passages.

D USER STUDIES

  • Participants were randomly assigned to human- or machine-policy conditions at low or high budgets, then classified 20 items and rated confidence without accuracy feedback.Human-policy participants could request assistance; machine-policy participants received assistance or a no-assistance message according to the ClassWise policy.Source: D USER STUDIES · p. 25
  • Items were sampled to preserve test-set proportions for policy assistance and support usefulness; Email sampling also stratified by ground-truth label.The study omitted a random-disclosure condition because its primary comparison concerned overall human- versus machine-led disclosure.Source: D USER STUDIES · p. 25

E EXTENDED RESULTS

E.1 Q1: DOES OUR APPROACH LEARN EFFECTIVE POLICIES?

ClassWise estimates Value of Information more effectively than standard meta-learners, improving budget–accuracy performance across synthetic and real data. Its advantage is greatest when disclosure is harmful and policies need to withhold unhelpful information.

E.1 Q1: DOES OUR APPROACH LEARN EFFECTIVE POLICIES?

  • ClassWise ranks instances better and abstains more selectively than standard meta-learners, with abstention most valuable when disclosure is harmful for many instances.The budget–accuracy curves compare VoI-based policies across synthetic and real datasets.Source: §E.1 · p. 29, Figure 6 · p. 30
  • On SynthBin and SynthMulti, ClassWise outperforms T-Learner and S-Learner at every budget, reaching about .85 versus .83 and .81 on SynthBin.On SynthMulti, ClassWise exceeds full-disclosure accuracy from B ≈.4, while both meta-learners converge to full-disclosure accuracy of ≈.62.Source: §E.1 · p. 29, Figure 6 · p. 30
  • On Email, ClassWise reaches ≈.81 at B = .30, while T-Learner reaches ≈.79 versus ≈.78 for B ≥.70; S-Learner degrades after B = .10.Because S-Learner’s estimated VoI is compressed toward zero, it loses sign information and largely stops abstaining where disclosure is harmful.Source: §E.1 · p. 30, Figure 6 · p. 30
  • On ImageNet-16H, ClassWise peaks at ≈.86 at B = .90, above every other curve and full disclosure at ≈.85.Disclosure helps almost everywhere there, and S-Learner tracks ClassWise closely up to B = .60.Source: §E.1 · p. 30, Figure 6 · p. 30
  • ClassWise’s positive-VoI set plateaus at about 54% of SynthBin instances and 71% of SynthMulti instances, compared with 69% and 83% for T-Learner.ClassWise also attains the highest accuracy at every budget on these datasets, indicating that its smaller positive-VoI set reflects withholding unhelpful information rather than missing useful disclosures.Source: §E.1 · p. 30, Figure 7 · p. 31
  • On Email, T-Learner and ClassWise plateau at ≈66% and ≈53%, respectively, while on ImageNet-16H they disclose to 98% and nearly all instances at B = 1.The Email policy’s greater abstention favors ClassWise at intermediate budgets, whereas T-Learner is slightly more accurate at large budgets; on ImageNet-16H, the estimators mainly differ in instance ordering.Source: §E.1 · pp. 30–31

E.2 Q2: DOES OUR LEARNED POLICY IMPROVE HUMAN-AI TEAM PERFORMANCE?

The user studies compare machine-policy and human-policy disclosure, examining team accuracy, confidence, advice correctness, and request alignment. Outcomes vary by task: the learned policy can improve performance and confidence discrimination, while human requests only partially overlap with policy-selected disclosures.

E.2 Q2: DOES OUR LEARNED POLICY IMPROVE HUMAN-AI TEAM PERFORMANCE?

  • The learned policy matches or outperforms self-selection, with greater confidence discrimination overall, but its advantage varies by task.In Email, confidence discrimination was greater in MP than HP even though selective disclosure did not significantly increase overall accuracy.Source: §E.2 · pp. 31, 33
  • In ImageNet-16H, MP accuracy tends to exceed HP when support is correct and is approximately comparable when support is incorrect; Email shows the opposite descriptive pattern.The authors relate these patterns to differences in advice adherence between the studies.Source: §E.2 · pp. 31–32
  • Across both studies, MP participants tended to show greater confidence discrimination than HP participants, assigning relatively higher confidence to correct than incorrect classifications.The pattern also appeared in Email, where selective disclosure did not significantly increase overall classification accuracy.Source: §E.2 · p. 33
  • Human requests overlap only partly with ClassWise selections: overlap ranges from 0.40 ± 0.30 to 0.87 ± 0.13 in ImageNet-16H and 0.31±0.26 to 0.56±0.20 in Email.Participants requested assistance on fewer trials than the policy provided it, and the authors note that why they request less requires further study.Source: §E.2 · p. 33
  • The studies compare complete disclosure strategies, so they do not identify why participants request assistance relatively infrequently.The authors identify overconfidence and uncertainty about assistance value as possibilities requiring further study.Source: §E.2 · p. 33

E.3 Q3: HOW DO PARTICIPANTS INTERACT WITH OUR POLICY?

  • Trust ratings differed by task: MP and HP did not significantly differ in ImageNet-16H, whereas MP trust was lower in Email, particularly under the lower budget.The Email MP–HP difference was significant under LB (2.95 ± 0.81 versus 3.40 ± 0.70; p = .002), but not under HB.Source: §E.3 · p. 35

E.4 Q4: HOW DOES CHANGING THE INFORMATION CONTENT AFFECT OUR POLICY?

The value of selective disclosure depends on the information content of prediction sets: larger sets are less informative and reduce the gains available from disclosure. ClassWise prioritizes small sets, but at α = 0.01 no policy separates from full disclosure.

E.4 Q4: HOW DOES CHANGING THE INFORMATION CONTENT AFFECT OUR POLICY?

  • Mean set size decreases from 6.05 at α = 0.01 to 1.64 at α = 0.05; full sets account for 387 images at α = 0.01 and 169 at α = 0.02.At α = 0.05, only 2 of 1200 images have sets larger than 10, making this the paper’s most informative operating point.Source: §E.4 · p. 37, Table 11 · p. 37
  • At α = 0.01 and α = 0.02, full disclosure reaches about 80.5%, only about 3.3 points above NoDisc≈77.2%, versus an 8.3-point gap at α = 0.05.As prediction sets grow larger and less discriminative, the benefit of full disclosure shrinks.Source: §E.4 · p. 38, Figure 9 · p. 38
  • At α = 0.02, ClassWise reaches full-disclosure accuracy at B = 0.5, while at α = 0.01 no policy separates from full disclosure within confidence bands.At α = 0.01, ClassWise is ≈79.3% at B = 0.9, about three images below full disclosure.Source: §E.4 · p. 38, Figure 9 · p. 38
  • At α = 0.01, full sets are frequent and uninformative, shrinking the positive-VoI population and making it harder to estimate which disclosures help.Ranking errors can lead the policy to withhold informative sets in favor of uninformative ones, explaining why it can fall slightly below full disclosure at high budgets.Source: §E.4 · pp. 36, 38
  • ClassWise favors small sets at every budget for α = 0.05, but at α = 0.01 it discloses 77 of 85 full sets (91%) at B = 1.At α = 0.02 it discloses all 38 full sets (100%) at B = 1.Source: §E.4 · pp. 38–39, Figure 10 · p. 39

F EXTENDED RELATED WORK

The related work spans delegation, support adaptation, active feature acquisition, policy learning, and treatment-effect estimation. The paper distinguishes its approach by allocating information rather than decision authority and by learning a budget-constrained policy from the causal effect of disclosure on human decisions.

F EXTENDED RELATED WORK

  • Learning to Defer allocates decision authority between model and human, whereas this work keeps the human as decision-maker and allocates information instead.The paper therefore defines risk over human actions under two information regimes rather than over model and human predictions.Source: F EXTENDED RELATED WORK · p. 39
  • Prior support-adaptation work includes instance-level disclosure rules (Noti & Chen, 2023) and policies choosing among support forms (Bhatt et al., 2025; Buc¸inca et al., 2024).LSD differs by imposing a disclosure budget and using causal VoI based on how people use the information.Source: F EXTENDED RELATED WORK · p. 40
  • Active feature acquisition selects missing feature values for a predictive model, while LSD makes one binary disclosure decision about indivisible support under an expected-rate budget.Unlike the classical Bayesian value of information, human-decision VoI can be negative.Source: F EXTENDED RELATED WORK · pp. 40–41
  • Budget-constrained policy learning also thresholds treatment effects, but here disclosure acts on the decision-maker and the outcome is correctness of a case decision.The paper states exchangeability across decision-makers assigned to regimes and SUTVA over decision episodes as relevant assumptions.Source: F EXTENDED RELATED WORK · p. 41
  • VoI estimation connects to heterogeneous treatment-effect methods, but this policy needs both effect ranking for budget allocation and the effect sign to avoid harmful disclosure.The paper notes that accurate effect estimation is neither necessary nor sufficient for good causal decisions.Source: F EXTENDED RELATED WORK · p. 41
Loading 2609.39818v1…