acceptodds
Under review as a conference paper at ICLR 2027

Information Architecture and Decision Learnability in Strategic Populations

Abstract

A coordinator may value groups' outcomes differently but observe only aggregate responses. We study how communication affects its ability to learn an optimal decision in repeated information design with strategic agents. Signals change equilibrium behavior and thereby determine the observation kernels available for learning. We give conditions under which public feedback depends only on pooled composition, while targeted signals provide local identification. Sublinear regret at every population in an observational equivalence class requires a common optimum. In a canonical game, every public learner has linear regret at some member of a fixed pair, whereas one population-independent private policy has at most logarithmic regret at both. Each architecture is compared with its own oracle. A Kullback–Leibler (KL) allocation lower bound accounts for information from optimal play. We compute its canonical value and construct a matching policy, identifying regions with positive and zero logarithmic coefficients. Alternatives approaching a decision boundary instead have local minimax regret of order . Nonlinear experiments test the induced information geometry: equally costly private probes can differ in whether they resolve a decision, and Gaussian information calculations predict the resulting errors. Adaptive comparisons examine when additional diagnostic sampling improves on greedy decisions.

Primary area:
learning theory
Keywords:
Information design, Bayesian persuasion, Online learning, Partial monitoring, Identifiability, Regret analysis~

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ZYCdOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies how signaling architecture affects learning in strategic populations. Public signaling can make populations observationally indistinguishable despite requiring different optimal decisions, while targeted private signaling can reveal the relevant differences. The paper establishes regret lower bounds, gives a canonical construction with matching guarantees, and presents nonlinear synthetic experiments.

Critical strengths
  1. This paper has an interesting mechanism and motivation. The paper clearly connects strategic aggregation, information architecture, and decision learnability. Its distinction between identifying the population and learning an optimal decision is useful.

  2. The paper has substantial theoretical analysis. The central arguments withstand close checking, including the observational-equivalence obstruction, KL lower bound, and canonical matching policy. Assumptions and scope limitations are generally stated carefully. (Theorems 1–5.)

Critical weaknesses
  1. The incremental novelty needs sharper justification. The canonical private problem reduces to a two-arm Gaussian bandit, as acknowledged in Appendix C.6. Much of the regret analysis then follows familiar bandit principles. The paper should clarify which contributions extend beyond the strategic construction and its architecture separation.

  2. The main acquisition comparison is confounded. Decision + LR and D-Optimal differ in both probe selection and their rules for switching to exploitation. Their performance gap therefore cannot be attributed solely to decision-focused acquisition. A comparison using a shared exploitation rule is needed. (Appendix F.4.)

Questions for authors

as above

1 comment
AuthorsOpenReview ↗

Contribution and the Matched Acquisition Comparison

Thank you for checking the analysis closely. We address the contribution and the requested acquisition comparison below; these also answer the questions referring to those concerns.

W1. Contribution beyond the Gaussian-bandit reduction.

We agree that the private canonical problem reduces to a two-arm Gaussian bandit. The common-optimum obstruction and KL allocation also use established tools. These connections are now explicit in the main text and contribution statements.

The model-specific results derive the observation laws from signaling and equilibrium, establish the public/private learning separation, and calculate exact canonical exploration coefficients with a matching policy. We have focused the contribution claims on these results rather than claiming a new general bandit technique.

Appendix C.7 adds a separation where public experiments are themselves the unique full-library optima. Public observations cannot distinguish the two populations, and learning requires costly private probes. Public-only policies have summed regret at least T/30T/30, whereas one full-library policy achieves O(log⁡T)O(\log T) regret at both fixed populations. This extension addresses the original construction's private dominance, but does not assert matching exploration constants.

W2. Confounded acquisition comparison.

You are right: the original comparison cannot isolate acquisition because the exploitation rules also differ. We ran the requested ablation with a shared exploitation rule and report it in Appendix F.5.

Both methods use the same four initialization probes, estimator, refit schedule, population constraints, likelihood searches, and LR threshold. Both exploit when certified and otherwise use their respective acquisition scores over the same 15 experiments. The extra free-information bypass is disabled in both. The rule is the same function of history; certification times can differ because acquisitions produce different histories.

At T=3000T=3000, mean regrets are:

SettingDecision acquisitionD-optimal acquisition
30 random populations, 10 paired seeds each1.66154.115
Positive-cost P+P^+, 100 paired seeds3.36111.412
Positive-cost P−P^-, 100 paired seeds0.3623.493

The paired 95% bootstrap intervals for Decision minus D-optimal regret are [−70.266,−35.734][-70.266,-35.734], [−10.627,−6.016][-10.627,-6.016], and [−3.603,−2.676][-3.603,-2.676], respectively. Resampling uses population clusters for the random suite and paired seeds for each fixed endpoint. Both methods make the correct terminal decision in all 100 runs at each fixed endpoint.

The lower regret does not simply come from fewer acquisition rounds. At P−P^-, Decision acquisition uses more such rounds (70.03 versus 47.53), but far fewer of those choices are suboptimal (1.43 versus 43.04). This supports the benefit of the selected probes under the common rule. The experiment reuses the existing evaluation populations and does not establish superiority over greedy or an anytime guarantee for the LR threshold.

Reviewer_Zfp5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper considers an information design problem with one sender and multiple receivers organized into classes. The groups are composed of different types, and the type distribution of the different classes is known to the receivers but unknown to the sender. Sender value is heterogeneous across classes, and the sender only observes aggregate information each round. Receiver value is dependent on type, the signal, and the decisions of other receivers.

In this setting, the authors study the learning and decision problem of the sender. They characterize the learnability and the resulting regret bounds in terms of the decision resolution (i.e., how distinguishable different populations are in equilibrium). They then fully work out a canonical example with two classes and two types. Finally, the results are validated numerically, first in the canonical setting, then in a more complex setting (4 classes, 3 types, 3 services) with two candidate populations, and lastly with the complex setting and know known candidates.

Critical strengths
  1. The framework of decision resolution is a nice way of understanding the learnability of the sender's problem.
  2. The analysis seems to be correct.
Critical weaknesses
  1. The paper is very hard to parse. Wording is extremely dense and jargon-filled (e.g. "public signaling pools class-specific compositions, whereas targeted contrasts can provide local identification"). There is also scarce intuition to guide the reader and explain the intuition or value behind results.
  2. The technical contribution is unclear (perhaps in part due to weakness 1 above). The theorems mostly seem to be applications of known arguments (e.g. from Graves & Lai 1997) with slight reframing.
  3. The model and its assumptions are not sufficiently motivated, and their ramifications on the analysis are not explained. Some choices which seem central to the results are stated in one sentence and never unpacked (e.g. receivers know pp but the sender does not). This leaves the reader asking why the problem matters and what real world decisions the results inform.
Questions for authors
  1. It is not clear to me how the results depend on the equilibrium. Doesn't the pooling come from the class-neutral utility and aggregate sender observations? Wouldn't those be true even without load externality? If that is the case, the focus on "strategic population" seems out of place.
  2. Can you provide a real-world motivating example for the model you investigate? What is your model capturing that others don't, and why is it important? I think more of this discussion in the paper would greatly improve its readability and value.
  3. In Section 6 you state that the chosen payoff offsets "align the public and private decision boundaries and make public experiments strictly dominated". Wouldn't a more interesting regime be one where there is a real tradeoff between value and learnability when considering public vs. private signals?
  4. Why does it make sense that receivers know pp, but the sender does not?
1 comment
AuthorsOpenReview ↗

Model Motivation and Public–Private Tradeoffs

Thank you for the careful reading and for finding decision resolution useful. We address the writing, contribution, and modeling concerns below.

W1. Dense wording and insufficient intuition.

We revised the introduction to explain the problem before introducing the terminology: two populations can produce the same total usage but different usage across groups. If those groups have different value to the sender, identical observations can conceal different best disclosure decisions.

We introduce this example before observational equivalence, explain the question behind each theorem, and split public pooling and local identification into separate results. The Gaussian-bandit reduction is also now explained in the main text.

W2. Technical contribution versus known arguments.

We agree that the common-optimum obstruction and KL lower bound use established arguments, including Graves–Lai. We have made that attribution clearer.

The model-specific results derive how signaling and equilibrium determine aggregate feedback, establish the public/private regret separation, and compute exact canonical exploration costs with a matching policy. The factor-of-four difference between the positive coefficients comes from the distance to the decision boundary along the directions each optimal experiment leaves unobserved. We have focused the contribution statements on these results.

W3. Motivation and unexplained assumptions.

We now explain the assumptions alongside their consequences. Class-neutral receiver utility supports pooling. Different sender weights allow identical aggregate observations to conceal different optimal decisions. Informed receivers, fresh cohorts, and fixed equilibrium selection give the stationary response model. Q2 explains the application motivation, and Q4 addresses the information asymmetry.

We answer your four questions in order.

Q1. Does pooling require congestion?

No. You are correct: pooling can hold at zero externality. Class-neutral responses, aggregate observation, and the stated selection convention produce the invariance. Congestion affects the equilibrium response and its sensitivity to composition, but is not necessary for pooling.

The strategic part is that receivers maximize their own payoff, taking the aggregate load created by others as given. Their actions and the resulting load must be consistent in equilibrium. We now distinguish this equilibrium model from the assumptions responsible for pooling.

Q2. What real-world decision does the model capture?

Traffic management offers a documented motivation for aggregate feedback: Jalota et al. (AISTATS 2023) study a controller using road flows without observing travellers' destinations or values of time.

Our introduction uses a service-group example to explain the additional information-design question. A provider chooses which groups receive availability information, but observes total usage while assigning different value to different groups. The question is whether targeting lets it learn the best disclosure decision. This example illustrates the model; we do not claim a validated deployment.

Q3. Can public and private signals have a real value–learning tradeoff?

Yes. We added this case in Appendix C.7. Adding a targeting cost of 0.430.43 makes FF and NN the respective unique full-library optima at the indistinguishable pair. Receiver behavior and feedback are unchanged. Both private probes are informative but suboptimal at both populations, so learning requires sacrificing value.

Public-only policies retain the summed-regret lower bound T/30T/30, while one full-library policy achieves O(log⁡T)O(\log T) regret at both fixed populations. Each experiment is uniquely optimal at some interior population, and these orderings persist over an open interval of targeting costs. The original lower bound uses the public oracle, so the common public cost shift cancels from regret.

Q4. How can receivers know pp while the sender does not?

Receivers may bring experience from earlier participation that the sender cannot access. Their local histories can give them expectations about congestion or competition, while the sender must learn from aggregate responses. Before receiving messages, they do not observe the current state, which the sender knows. Its messages therefore remain useful.

Our assumption idealizes this experience by giving receivers exact knowledge of pp and correct expectations under each available experiment. Local experience alone need not identify the full composition. We study sender learning once receiver expectations are established; each fresh cohort is assumed informed beforehand. How those expectations are acquired, or how inaccurate beliefs affect the results, is outside our current guarantees.

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

The paper studies repeated information design in which a sender signals a state to a population of receivers. Receivers are partitioned into groups (e.g., subscription plans), and the sender can only send the same message to everyone in a group. Each receiver has a private type, and the type of composition of each group is unknown to the sender. Receivers respond in equilibrium under a congestion externality. The sender's utility depends on the actions of receivers in each group, with groups weighted differently (e.g., usage by premium subscribers may be worth more). However, the sender observes only a noisy aggregate summary pooled across all groups (total usage of each service), so it cannot directly observe its own utility. The sender uses this aggregate feedback to learn which experiment, from a fixed finite library (full reveal or none), to use, with the goal of minimizing regret.

Contributions:

  • They compare between public and private signaling scheme and show that the public one is not rich enough to know the composition and thus can incur linear regrets
  • They provide no regret learning algorithms with lower bounds.
Critical strengths

The model seems interesting. When the receivers' response is independent of the group, public message can only provide information about pooled composition.

Critical weaknesses

My main concern is that the problem is less about information design than about estimation and learning from partial (aggregate) observations, and the learning results largely reduce to known ones.

  1. Reduction to standard bandits. The private canonical game is effectively a two-armed Gaussian bandit with observable rewards. The other learning results (Theorems 1, 3, and 5) are likewise specializations of known partial-monitoring and structured-bandit results.

  2. Thin information-design layer. The design space is a fixed finite library of full-or-no disclosure rules, whereas information design usually optimizes over general signaling schemes. Fixing the library seems to reduce the problem to the finite-armed bandit literature.

  3. Presentation. The paper is hard to follow. For example, Theorem 2 combines several mixed results. Splitting it would make the actual claims much clearer.

Questions for authors

The paper consider a fixed finite set of signaling scheme. Can the framework handle a continuous space of signaling schemes (e.g., all reveal-probability vectors, or all obedient recommendation schemes)? Which parts would break, and what regret rate should one expect relative to the optimal scheme?

Receivers know the composition pp while the sender does not. When is this asymmetry natural?

Theorem 2 assumes that, under multiple equilibria, selection depends only on the pooled fixed-point problem and not on group labels. This is different from the standard conventions (sender-preferred or adversarial (worst-case) selection). Does the public pooling result, and hence the linear-regret separation, survive under sender-preferred or adversarial selection?

1 comment
AuthorsOpenReview ↗

Signaling Scope and Learning Guarantees

Thank you for the careful review. We address the three weaknesses first, followed by your questions.

W1. Reduction to standard bandits.

We agree that the private canonical problem reduces to a two-arm Gaussian bandit. Theorems 1, 3, and 5 also use established learning arguments. We now state these connections in the main text.

Our contribution is the equilibrium analysis that produces the feedback laws, the public/private learning separation, and the exact exploration coefficients with a matching policy in the canonical model. Public observations depend on pooled type frequencies, while the sender values groups differently. We construct populations with identical public observations but different best public decisions. Conflicting decisions, rather than unidentified composition alone, give the linear lower bound. Each architecture is compared against its own oracle.

Appendix C.7 adds a setting where public experiments are the unique full-library optima at the two populations. Private probes sacrifice value to distinguish them. Public-only policies have summed regret at least T/30T/30; one full-library policy achieves O(log⁡T)O(\log T) regret at both fixed populations. Identical feedback under a public experiment then corresponds to different values, so its own observations cannot provide an unbiased value estimate at both populations. This extension does not assert matching exploration constants.

W2. Finite signaling library.

Full-or-no disclosure is restricted to the canonical example. The general finite library allows Bayes-plausible kernels, and the nonlinear benchmark includes partial disclosure.

The finite-library limitation is real: our oracle is the best scheme in that library. Once equilibrium values and observation laws are derived, learning over the library is a finite-action partial-feedback problem. We have clarified this scope. The paper does not optimize over all persuasion schemes.

W3. Readability of Theorem 2.

We split Theorem 2 as suggested. The public-pooling theorem explains what public signals cannot distinguish. A separate proposition gives conditions under which targeting distinguishes nearby population compositions. We introduce the question and intuition before each statement, so the reader sees the purpose of the result before its notation.

We answer your three questions below.

Q1. Continuous signaling spaces.

Public pooling and equality of adaptive history laws extend to measurable infinite libraries of public schemes under the same response and selection assumptions. The finite allocation program and matching policy are proved for finite libraries.

A larger library may introduce a common optimum or make the infimum of the combined gaps zero, so the existing witness need not yield a linear lower bound against that larger oracle. Appendix A.4 gives the positive-gap condition that does suffice. We have not established a general continuous-space regret rate. A discretization approach would need control of both equilibrium values and feedback across schemes, which is not provided by our current analysis.

Q2. When is the information asymmetry natural?

One interpretation is that receivers bring experience from prior participation, whereas the sender cannot access their individual histories and sees only aggregate responses. Receivers can then have established expectations about congestion or competition. Historical experience does not reveal the current state before messages arrive. The sender observes that state and can still provide useful information.

We idealize this experience by assuming that receivers know pp and form correct equilibrium expectations for every available experiment. Local experience need not identify the full pp; exact knowledge remains a benchmark assumption. Each fresh cohort is assumed informed before interacting with the sender. We do not model the acquisition of that knowledge.

Q3. Sender-preferred or worst-case selection.

The canonical separation survives both conventions because its equilibrium is unique. The general multiple-equilibrium pooling result is conditional on the stated selection assumption.

Equal pooled compositions give the same aggregate fixed-point problem, but the sender's group-weighted values may rank its solutions differently at the two populations. Selection based on sender value therefore need not preserve pooling. Appendix A.4 distinguishes this limitation from the unique-equilibrium construction.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Contribution, Tradeoff, and Controlled Comparison

Thank you for the careful reviews. The shared concerns were the contribution beyond known learning tools, the motivation and readability of the model, and the acquisition comparison. We have revised the manuscript and added two analyses in response.

Contribution and presentation. We now state the Gaussian-bandit reduction and the established origins of the common-optimum, KL, and two-point arguments explicitly. Our contribution claims focus on deriving feedback from signaling and equilibrium, proving the public/private learning separation, and obtaining exact canonical exploration coefficients with a matching policy. The introduction explains the aggregate-feedback ambiguity through a group-usage example before introducing observational equivalence. We also split Theorem 2 into public pooling and a separate local-identification proposition, with intuition before each statement.

A value–learning tradeoff with public optima. Appendix C.7 adds a targeting cost without changing receiver behavior or feedback. At the indistinguishable pair, the respective unique full-library optima are now the public experiments FF and NN. Learning which to choose requires private probes that sacrifice value. Every public-only policy has summed regret at least T/30T/30, whereas one full-library policy achieves O(log⁡T)O(\log T) regret at both fixed populations. The public and full-library oracles coincide at this pair, so the separation also holds without private static-value dominance. This extension does not claim matching exploration constants.

Acquisition under a shared exploitation rule. Following ZYCd's request, Appendix F.5 holds initialization, estimation, refits, and LR certification fixed, and disables the extra free-information bypass in both methods. At T=3000T=3000, Decision versus D-optimal acquisition gives mean regrets of 1.6611.661 versus 54.11554.115 on the random suite, 3.3613.361 versus 11.41211.412 at P+P^+, and .362.362 versus 3.4933.493 at P−P^-. All three paired regret intervals exclude zero. At P−P^-, Decision uses more acquisition rounds but far fewer suboptimal acquisitions, so the gain is not simply fewer diagnostic rounds. These are reused evaluation populations; this ablation does not establish superiority over greedy or an anytime guarantee for the LR test.

The model section now explains prior receiver experience as a motivation for established congestion expectations, while retaining exact knowledge of pp as an idealization. It also distinguishes this knowledge from the current state disclosed by the sender. Appendix A.4 separates the unique-equilibrium canonical guarantee from the conditional multiple-equilibrium pooling result, and explains which arguments extend to infinite public libraries. Regret against a continuous-space oracle remains open.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.