acceptodds
Under review as a conference paper at ICLR 2027

Incentive Alignment Reshapes Coordination Over Unreliable Communication

Abstract

When communication is unreliable, agents with the same objective may need different participation rules. We study Bayesian coordination games in which utilities interpolate between private and group-average payoffs. For full-quorum relay lines with heterogeneous links and an informed root that vetoes unsafe operation, we characterize all coordinating observation-monotone pure equilibria: each consists of a message-responsive prefix and an unconditionally acting suffix, determined by three posterior-odds inequalities. Permuting the same link reliabilities can change the equilibrium assignment while preserving end-to-end reliability, so no scalar summary of the channel determines the local policies. Under a uniform veto condition, equilibrium existence and the best certified welfare are monotone in alignment, although a canonical equilibrium selection need not be. The characterization yields an exact compiler that separates equilibrium certification, welfare-based selection, and local execution. Experiments with open-weight language models distinguish the formal certificate used for audit from the representation given to the executor. On fresh relay games, executing the full audit certificate directly loses welfare relative to not coordinating at all, whereas the same certificate compiled into a local decision-table execution card is executed almost exactly and recovers the welfare of the selected policy; violations of the declared fallback remain when the certificate's assumptions fail. The results connect an exact coordination structure to a local execution interface while separating policy fidelity from validity enforcement.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Incentive alignment, unreliable communication, Bayesian coordination, policy certificates, LLM agents

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_ybSROpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies Bayesian coordination games on unreliable relay lines, with utilities interpolating between private and group-average payoffs. Under full-quorum, root-veto, and observation-monotonicity assumptions, it characterizes coordinating pure equilibria as a message-responsive prefix followed by an unconditionally acting suffix. Three posterior-odds inequalities determine equilibrium validity. The paper also establishes alignment comparative statics and develops an exact compiler that separates equilibrium certification, welfare-based policy selection, and local execution.

Validation includes reported agreement with direct payoff enumeration on 71,640 cutoff checks, computational studies of channel ordering and equilibrium selection, and language-model experiments on equilibrium reconstruction and policy execution. The fresh-game format study finds that decision-table instructions preserve substantially more welfare than full audit certificates, although failures to follow the prescribed fallback remain when assumptions are violated.

Critical strengths

An interpretable structural characterization with computational consequences. The responsive-prefix/blind-suffix theorem reduces the equilibrium problem within the specified class to three transparent incentive constraints. This is more informative than merely enumerating equilibria: it explains which observations are strategically decisive and supports efficient computation of exact validity intervals. The result is a substantive reason to consider acceptance because it connects an understandable coordination mechanism to a tractable solution procedure.

The evaluation distinguishes policy quality from execution fidelity. The paper does not equate finding an equilibrium with achieving useful outcomes, or correct instruction-following with reliable handling of invalid assumptions. Its fresh-game ablation compares representations of the same selected policy and evaluates welfare as well as action agreement. Table 2 on page 8 reports normalized welfare of 0.164/0.1470.164/0.147 for decision tables versus −0.206/−0.072-0.206/-0.072 for full certificates on the two models. This provides concrete evidence that the execution interface matters, while the separately reported fallback failures appropriately limit the conclusion.

Critical weaknesses

The significance of the restricted equilibrium class is insufficiently established. The characterization requires a relay line, full participation for success, pure observation-monotone strategies, and a root veto. These restrictions exclude actual equilibria: the paper’s own structural search finds 46 of 229 coordinating pure equilibria outside the cutoff family. Welfare optimality is consequently only relative to the certified class, and the paper does not establish the welfare cost of imposing that restriction. This does not invalidate the theorem, but it leaves unclear whether the tractable family captures the important coordination opportunities beyond the selected model. That limitation materially weakens the case for a broadly consequential contribution.

The experiments establish a narrow instruction-execution result rather than a strong empirical connection to incentive alignment. The language models receive externally computed policies; forwarding is fixed, and agents do not communicate during action selection. The decisive format study uses two models and 24 constructed games with three to five agents. Thus, the experiments primarily test reading facts and applying prescribed rules—not whether agents respond strategically to the modeled incentive changes. Although the paper acknowledges this scope, it leaves the theoretical and empirical contributions only loosely connected: the evidence establishes that decision tables work well in this sandbox, but not how much that finding depends on equilibrium certification rather than ordinary rule presentation. A stronger demonstration of that connection is important to the acceptance case.

Questions for authors

No

2 comments
AuthorsOpenReview ↗

Response by Authors 1

We thank the reviewer for the careful review. We apologize for the slightly late response: it took us time to establish the new theoretical result and to prepare the improved version of the paper. All changes to the paper are described in the general response; line numbers refer to the revised PDF. We answer the two weaknesses in turn: the restricted equilibrium class (W1) and the link between theory and experiments (W2).

W1. What does the restriction to the certified family cost?

No welfare loss, when an unsafe quorum is costly enough. This is the new Theorem 3 (Section 3.1, lines 270-287; see also the general response, item 1). Assume full support (the prior and every link reliability lie strictly between 0 and 1) and H>βαLH \gt \beta \alpha L, which holds in every game of the language experiments of Section 5. Then:

  • every equilibrium, pure or mixed, has a root veto, so the veto follows from the payoffs and is not an assumption;
  • no pure or mixed equilibrium has higher welfare than the selected policy σW\sigma_W (the best certified cutoff, or all-WAIT).

Under the uniform condition H>βLH \gt \beta L, σW\sigma_W at α=1\alpha = 1 is optimal among all decentralized behavioral profiles.

The 46 of 229 equilibria outside the family. The question about them is what led us to the theorem. We looked at what they are:

  • 31 have a responsive prefix, a blind block, and a suffix of agents that act only when no message arrives. They coordinate on the event that the message was lost, and give up the worlds in which it arrived. For each, the cutoff with the same prefix is also an equilibrium and pays every agent weakly more in every world.
  • 15 have a root that acts when the state is unsafe. They occur only when H≤βαLH \le \beta \alpha L.

Why the theorem holds. On a line, the agents that act only without a message must form the end of the line. Making them blind gives a cutoff that is again an equilibrium and completes the quorum in the worlds they gave up. In a mixed equilibrium, every type that acts with positive probability has a success probability of at least ρ(α)=(1−βα)L/(U+L)\rho(\alpha) = (1 - \beta \alpha) L / (U + L), which is at least ρ(1)\rho(1); at or above ρ(1)\rho(1) acting does not lower total welfare. So raising these types to ACT loses nothing, and upward best responses then reach a pure equilibrium, again without loss.

The line and pure strategies. The review also names these two restrictions. Neither is needed for this conclusion. The argument covers mixed equilibria, and under H>βαLH \gt \beta \alpha L Appendix M proves the same result on any acyclic forwarding network (Theorem 6, line 1674), with a closed form on rooted trees (Theorem 5, line 1582); see the general response, item 2. The exact three-inequality characterization of Theorem 1 remains specific to the line.

Checks and scope. Enumeration in exact arithmetic, exhaustive over pure profiles and partial for mixed equilibria, found no failure on 20,724 line games and on 2,300 trees and 1,600 acyclic graphs (Table 21; Appendix M.6, line 1711). The results need a full quorum, a single informed root, and automatic forwarding with receipt-only observations, and the theorems take the parameters as known (Section 7, lines 528-537). If H≤βαLH \le \beta \alpha L welfare sufficiency can fail; Appendix M.7 (line 1725) reports how often.

AuthorsOpenReview ↗

Response by Authors 2

W2. How are the experiments connected to incentive alignment?

In the paper the payoff-bearing agents are the strategic players, and the language models execute their local rules. The incentive weight α\alpha therefore acts through the compiler: it changes which policy is certified and selected, and the card carries that policy to the model. We agree that this link deserves direct tests. Both tests named in the review are now in the paper as preregistered studies on fresh games, together with a study on scale (Section 5.5, lines 459-489; Appendix N; general response, items 3 and 4). All predictions registered for the two primary models held.

Does unaided execution track incentive changes? No; the response comes through the compiled card. We vary only α\alpha within each game.

  • Given the game, the objective with its α\alpha, and the selection convention, but no policy, both models play the same heuristic at every α\alpha. A request to reason step by step does not change this (exploratory).
  • With a card compiled at each α\alpha, behavior follows the exact policy: the tracking index, where 1 means switching exactly where the exact solution does, rises from 0.00 to 1.00 and 0.98, and the welfare gap to the exact policy nearly closes.
  • Example (Table 4): in the game of the introduction the exact policy switches from all-WAIT to the hybrid as α\alpha grows. Without a card, agent 3 never acts on a missing message on either side of the switch; with the card it never does below the switch and always does above it (both models).

Does the result depend on certification or on rule presentation? Execution depends on presentation; what certification adds is stability. We keep the decision table fixed and change only the policy on it: the certified policy, the policy of the reversed link order, the scalar-reduction policy, or a naive rule.

  • Fidelity is above 0.99 for all four. We agree with the review on this point: execution does not depend on certification.
  • The three uncertified policies are often not equilibria. The scalar policy even prescribes more welfare than σW\sigma_W (0.26 against 0.22). But under exact best-response dynamics of the payoff-bearing agents it falls to 0.20, for every revision order we tried, and no prescription ends above σW\sigma_W. This is a solver computation on the game, not model behavior; it is Theorem 3 at work.

Scale: two models, 24 games, three to five agents.

  • The three new studies use 68 games, including lines with up to 8 agents and trees. On the longer lines and trees the decision table still loses about 0.01 nUnU or less for the two primary models, against 0.57 and 0.17 for the full certificate.
  • Three replication executors, added post hoc, follow the table less exactly but still do better with it than with the full certificate. The paper claims near-exact welfare for the two primary models (Table 22).

Forwarding and communication. Forwarding is automatic and agents do not communicate during action selection. This is a property of the model and remains a stated limitation.

Reviewer_KKm3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper looks at a simple relay-communication game: a root agent learns whether an operation is safe, and that information has to pass down a chain of agents before they can all act together. Because messages can get lost along the way, each agent has to decide whether to act only after hearing from upstream, or act regardless. The authors prove that the only equilibria of this kind take a simple ``cutoff'' form, such as early agents wait for confirmation, later agents act unconditionally.

Critical strengths

Theorem 1 gives a clean characterization of a nontrivial equilibrium class, and the order-dependence finding is a sharp, well-illustrated structural insight.

Critical weaknesses

It is unclear how far the central theoretical contribution extends beyond the specific relay topology studied. The LLM experiments show substantial and unexplained model-to-model inconsistency (e.g., the action-first format gives Qwen welfare 0.161, Mistral welfare 0.003 on the same games).

Questions for authors
  1. Can the authors say more about why the action-first reordering helps one model and hurts the other?
  2. What principle should guide selection in settings where welfare cannot be directly computed at deployment time?
1 comment
AuthorsOpenReview ↗

Response by Authors

We thank the reviewer for the constructive review. We apologize for the slightly late response: it took us time to establish the new theoretical result and to prepare the improved version of the paper. All changes to the paper are described in the general response; line numbers refer to the revised PDF. We answer the three points in turn.

W1. How far does the theory extend beyond the relay line?

One part is specific to the line; the rest holds on any acyclic forwarding network with a full quorum.

  • Line-specific: the three-inequality cutoff characterization of Theorem 1. The line is what makes the solution a cutoff in position.
  • General: the mechanism behind it. Once the root does not act when unsafe, ACT is a best response for a type iff the others complete a safe quorum with probability at least ρ(α)=(1−βα)L/(U+L)\rho(\alpha) = (1 - \beta \alpha) L / (U + L), on any acyclic forwarding network. Actions are then strategic complements.

To state this precisely we added Appendix M (general response, item 2):

  • Theorem 4 (line 1531; if H>βLH \gt \beta L). A greatest equilibrium exists, bounds every equilibrium, and is nondecreasing in α\alpha. So coordination, once possible, stays possible as α\alpha grows.
  • Theorem 5 (line 1582). On rooted trees the certified profiles have a closed form, which on a line is exactly Theorem 1. On a star the responsive agents are those with the most reliable links.
  • Theorem 6 (line 1674; if H>βαLH \gt \beta \alpha L). No equilibrium has higher welfare than the best certified profile or all-WAIT.

Checks in exact arithmetic (exhaustive for pure profiles, partial for mixed ones) found no failure on 2,300 trees and 1,600 acyclic graphs (Table 21, Appendix M.6). Cards compiled from Theorem 5 are executed on trees with fidelity 0.999 and 0.998 (Table 28). The limits: the results need a full quorum, one informed root and automatic forwarding.

W2 and Q1. Why does action-first help one model and hurt the other?

We can say where the difference arises and that it is systematic, but not why.

Where. The difference is in blind agents that refuse a prescribed ACT. Relative to compact prose, action-first lowers this rate on Qwen3-30B (8.3% to 0.2%) and raises it on Mistral-24B (0.9% to 4.8%). In a full-quorum game each refusal breaks the quorum, which accounts for the welfare gap (Appendix N.1).

Systematic, not noise. To check this we ran three more executors on the same games, as post hoc replications. Action-first minus compact prose, in units of nUnU:

Qwen3-30BQwen3-8BMistral-24Bgpt-oss-20bGranite-3.3-8B
+0.147+0.088-0.063-0.138-0.043

Each difference has a 95% bootstrap interval that excludes zero. The sign is positive for both Qwen models and negative for the other three. We do not identify a mechanism inside the models, and the paper says so (see also the general response, item 4).

What follows. In this study the effect of a prose format depends on the model. The decision table does not have this problem for the two primary models, which follow it almost exactly. For the three post hoc replication executors the best compact format differs by model (Table 33), and the paper claims near-exact welfare for the two primary models (abstract, lines 25-26; Section 7, lines 528-537).

Q2. What should guide selection when welfare cannot be computed at deployment?

Selection does not have to happen at deployment: the compiler selects offline, and the executor only follows its card. The principle (Section 4): deploy only a policy that is certified for every parameter value the compiler considers possible, and otherwise fall back to WAIT. In practice:

  • Payoffs and channel parameters known: select by exact welfare, as in the paper.
  • Channel parameters uncertain: among the policies certified throughout the parameter box and all-WAIT, take the one with the best worst-case welfare. It is an equilibrium for every parameter value in the box; the price is some expected welfare, and it grows with the uncertainty (Appendix N.3, line 1887; general response, item 6).
  • No welfare comparison at all: the two natural rules, the maximal cutoff and the greatest equilibrium of Theorem 4, often select equilibria with negative welfare, so such rules need at least a sign check against WAIT. We do not claim that no welfare-free rule can work.
  • Payoffs or α\alpha unknown: the paper does not cover this case. The equilibrium test itself uses these values, so we leave it for future work.
Reviewer_vz5COpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies how the weighting of individual and group payoffs affects participation and coordination strategies among multiple agents under unreliable communication. Under a full-quorum requirement, relay-line communication, and a root-veto condition, the authors prove that observation-monotone pure-strategy coordinating equilibria have a structure consisting of a message-responsive prefix and an unconditionally acting suffix, and provide the corresponding equilibrium conditions and validity intervals.

The paper further distinguishes how equilibrium existence, achievable welfare, and the selection of a particular equilibrium vary, and turns the theoretical results into a compilation pipeline that produces audit certificates and local execution cards. Through independent solver verification, comparisons of policy selection, and representation ablations with language models, the experiments examine whether formally valid policies yield beneficial and compliant execution in practice.

Critical strengths

• The theoretical results offer a clear structural characterization and are computationally tractable. The authors reduce equilibrium characterization within the restricted policy family to verifiable local conditions and reveal the relationships among link ordering, participation patterns, and incentive weights. These results also support the computation of validity intervals and welfare-based selection, making the theory directly useful for policy generation rather than merely explaining individual examples.

• The distinction among certification, selection, and execution has practical research value. The paper makes clear that equilibrium validity alone does not guarantee deployment benefits, and faithful execution cannot compensate for an unsuitable policy choice. By comparing representations of the same target policy and using independent game verification, the authors connect local instruction compliance to team outcomes while also documenting execution problems when the assumptions fail.

Critical weaknesses

• The incremental contribution relative to the closest theoretical work is not sufficiently delineated. Awaya and Krishna have studied participation rules based on message arrival and equilibrium coordination under information diffusion, warranting a more direct comparison with this line of work. Although Section 6 identifies differences such as a participating veto source, heterogeneous links, and incentive weights, it gives only a brief account of which conclusions build on existing mechanisms and which structural results require new technical arguments, making the extent of theoretical novelty harder to assess.

• Attributing execution-interface improvements to specific mechanisms remains difficult. The compact-text, content-reordering, and conditional-error analyses support the representation design, but the execution-card and deterministic-validation pipelines use different contexts for fact extraction, so payoffs remain jointly affected by information recognition and rule application. The authors acknowledge this limitation; it primarily constrains interpretation of the execution method’s independent effect rather than negating the empirical value of the overall pipeline (Section 5 and Appendices F and H).

Questions for authors

• Theoretical contribution: Relative to existing coordination models with information diffusion, what are the key new structural results and the associated proof challenges? A brief comparison clarifying the respective roles of heterogeneous links, a participating veto source, and incentive weights would help assess the substantive contribution.

• Experimental attribution: Could the existing outputs from the execution-card condition be used to compare model actions with deterministic rule-based actions given the same extracted facts? This analysis would help distinguish the contributions of fact recognition and rule application to final payoffs and complement the existing conditional-error analysis.

• Policy restrictions: Can the existing theory or solver results clarify whether any welfare-dominance relationship holds between the certified policy family and other pure-strategy equilibria satisfying the root-veto condition? This would help assess how the observation-monotonicity restriction affects deployment choices without requiring a full characterization of all equilibria.

1 comment
AuthorsOpenReview ↗

Response by Authors

We thank the reviewer for the detailed review. We apologize for the slightly late response: it took us time to establish the new theoretical result and to prepare the improved version of the paper. All changes to the paper are described in the general response; line numbers refer to the revised PDF. We answer the three questions in turn, starting with the one that led to the main new result.

Q3. Is there a welfare-dominance relationship with other root-veto equilibria?

Yes. On a line, every coordinating pure equilibrium with a root veto outside the cutoff family has one shape: a responsive prefix, a blind block, and a suffix of agents that act only when no message arrives. Each is dominated: the cutoff with the same prefix is also an equilibrium and pays every agent weakly more in every world. Enumeration on 20,724 games found 2,671 such equilibria, all dominated (Table 21).

The question led us further. Under full support and H>βαLH \gt \beta \alpha L, no equilibrium, pure or mixed, has higher welfare than the selected policy. So restricting to observation-monotone equilibria (all-WAIT included) loses no welfare relative to any other equilibrium. Both statements are in the new Theorem 3 (Section 3.1, lines 270-287; general response, item 1).

W1 and Q1. What is new relative to Awaya and Krishna?

We compared the two analyses step by step; the answer is below. So that the novelty can be judged from the paper itself, Section 6 now states the difference in one sentence, and Appendix O (line 2037) gives the comparison (general response, item 5).

What is shared. In both models agents act either on a message or without one. With identical links on a line, our model gives the same three qualitative regimes as theirs, including all-WAIT (Corollary 1).

What each ingredient adds.

  • Heterogeneous links. Their model has one common message-loss probability, and their proof passes dominance along a path. With a reliability per link the odds ηj\eta_j are not ordered along the path, the argument stops, and hybrid equilibria appear (Theorem 1). Their best adoption rate is the same on every acyclic network; our role assignment depends on the link order.
  • Participating root. Their informed source is a planner outside the game; ours is a player with a veto. If H>βαLH \gt \beta \alpha L, the veto holds in every equilibrium and the remaining game has strategic complementarities.
  • Incentive weight. One threshold ρ(α)=(1−βα)L/(U+L)\rho(\alpha) = (1 - \beta \alpha) L / (U + L) gives the validity intervals and, for H>βLH \gt \beta L, a greatest equilibrium that grows with α\alpha and the team optimum at α=1\alpha = 1.

New proof steps (Appendix M). Making anti-responsive agents blind raises welfare when welfare is positive (Lemma 6, line 1644; a Harris-inequality argument). A marginal-welfare argument covers mixed equilibria (Lemma 4, line 1629).

W2 and Q2. Model action versus deterministic rule on the same extracted facts

Yes, the existing outputs allow this comparison, and it needs no new queries: every single-call response already contains the reported facts and the action. We score the model's action and the card applied to the facts of the same response. This separates reading the facts from applying the rule, which is what the question asks for (new Appendix N.1, line 1784; general response, item 4). Welfare loss in units of nUnU:

Representation, modelFact reportRule application
Full certificate, Qwen3-30B0.0190.352
Full certificate, Mistral-24B0.0650.171
Decision table, Qwen3-30B0.0000.000
Decision table, Mistral-24B0.068-0.051
  • With the full certificate, most of the loss comes from applying the rule, not from reporting the facts. With the decision table this loss is gone: the component is 0.000 for Qwen3-30B and -0.051 for Mistral-24B.
  • The negative entry means that Mistral's action is better than the rule applied to the facts it reported. So a validator on reported facts is not an upper bound, and the paper now says so (Section 5.3, lines 428-430).
  • Three more executors, run as post hoc replications, follow the decision table less exactly, so the paper claims near-exact welfare for the two primary models (Table 22).
Reviewer_AAhsOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

It is an interesting work but its contributions are unclear at this point. I would need to spend more time working on its reviewing.

I hope to hold it on for more clarifications.

Critical strengths

There are some potential insight for ICLR: 1. Correct equilibrium prescriptions do not ensure faithful agent execution. 2. Compilation into local rules is a transferable way to improve execution. 3. Policy fidelity and validity enforcement are distinct capabilities.

Critical weaknesses

The issue is not whether it solves equilibrium implementation in general. Instead: does the paper deliver enough evidence and understanding to justify its place at ICLR. I would like to see this kind of work represented at ICLR, but need to be explained what the learning community gains from it.

Questions for authors

I still need time to check into details to justify my decision. I am a little slow checking into the details.

1 comment
AuthorsOpenReview ↗

Response by Authors

We thank the reviewer for raising the question of the paper's relevance to ICLR. We apologize for the slightly late response: it took us time to establish the new theoretical result and to prepare the improved version of the paper. All changes to the paper are described in the general response; line numbers refer to the revised PDF.

Our view. In many agent systems a planner or solver computes a joint policy, and language models carry out its local parts from partial information. In our setting the right answer can be computed exactly, so three questions that are easily conflated can be measured separately:

  • Is the policy an equilibrium for the agents who bear the payoffs?
  • Is it worth deploying?
  • Does a language model that executes it follow it?

Each can fail on its own. An equilibrium can have negative welfare. A correct and complete certificate, shown to a model, can yield less welfare than not coordinating at all. On 20 new games with longer lines and trees, the same policy written as a local decision table does better than the full certificate for all five models we tested, and for the two primary models it nearly attains the welfare of exact execution. In a controlled sweep, model behavior did not follow an incentive weight that was only stated; it did once the weight was compiled into the local rule.

The design principle we draw from this: certify globally, select explicitly, and give the executor only a compiled local rule. The three points listed in the review say much of this, and we used them to state it in the paper.

In the paper. The Introduction now states this before the contributions (lines 93-99) and says who is strategic and who only executes (lines 58-60); the caption of Figure 1 maps the three steps to their failure modes. Section 7 (lines 528-537) states the remaining limits. See also the general response, items 3 and 7.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response by Authors: Paper Update

We thank the reviewers. This comment describes what changed in the paper and why. The model, Theorems 1 and 2, the compiler, and all submitted experimental results are unchanged. The main text is 10 pages; line numbers refer to the revised PDF. Items 1 to 3 are the main additions.

1. Welfare sufficiency: new Theorem 3 (Section 3.1, lines 270-287). The paper selects the best policy among the certified cutoffs and all-WAIT. The new theorem shows what this selection is worth among all equilibria: under full support and H>βαLH \gt \beta \alpha L, no pure or mixed equilibrium has higher welfare than the selected policy σW\sigma_W. In that case the coordinating pure equilibria outside the family all contain agents that act only when no message arrives, and each is dominated, agent by agent, by a cutoff equilibrium. Under the uniform condition H>βLH \gt \beta L, σW\sigma_W at α=1\alpha = 1 is optimal among all decentralized behavioral profiles. So when an unsafe quorum is costly enough, the cutoff family with all-WAIT is enough for selection: more complex equilibria cannot improve welfare.

2. Beyond the relay line: new Appendix M. It separates what depends on the line from what does not. The three-inequality cutoff characterization is line-specific. The threshold and welfare arguments hold on any acyclic forwarding network with a full quorum: a greatest equilibrium that grows with α\alpha (Theorem 4, if H>βLH \gt \beta L), a closed form on rooted trees (Theorem 5), and welfare sufficiency (Theorem 6, if H>βαLH \gt \beta \alpha L). Exact checks (exhaustive for pure profiles, partial for mixed ones) found no failure (Table 21). So the mechanism is not tied to the line.

3. Incentives and certification in the experiments: new Section 5.5 and Appendix N. The earlier experiments fix α\alpha in each game and always supply a certified policy. Three preregistered studies on fresh games vary these two factors, to show how the incentive weight reaches the model and what certification adds. All predictions registered for the two primary models held.

  • Policy source (lines 462-467): with the interface fixed, certified and uncertified policies are executed equally faithfully. What certification adds is stability: an uncertified policy can prescribe more welfare than σW\sigma_W, but exact best-response dynamics of the payoff-bearing agents (a solver computation, not model behavior) bring it below σW\sigma_W.
  • Alignment sweep (lines 469-476): in this sweep the incentive weight reaches the model only through the compiled card. With a card compiled at each α\alpha, the tracking index (1 means following the exact policy) rises from 0.00 to 1.00 and 0.98; Table 4 shows one game.
  • Longer lines and trees (5 to 8 agents): the interface result carries over. For the two primary models the decision table still loses about 0.01 nUnU or less, against 0.57 and 0.17 for the full certificate.

4. Attribution and generality of the interface result (Appendix N.1; Tables 22 and 23). A same-response decomposition shows where the full certificate loses welfare: for the two primary models, mostly in applying the rule, less in reporting facts. With the decision table this loss disappears, which explains its benefit. On the 20 new games the decision table beats the full certificate for all five models. Three post hoc replication executors follow it less exactly, so the paper claims near-exact welfare for the two primary models.

5. Relation to Awaya and Krishna (Section 6, new Appendix O). A step-by-step comparison shows which conclusions share their mechanism and which need new arguments.

6. Selection under uncertainty (new Appendix N.3). For uncertain channel parameters, we add a robust rule that remains an equilibrium throughout the parameter box, at some cost in expected welfare, so the compiler stays usable when the channel is known only approximately. The two welfare-free rules that we tested often select equilibria with negative welfare, which supports selecting by welfare.

7. Presentation for an ML audience. The Introduction says who is strategic and who executes and states the takeaway before the updated contributions; the caption of Figure 1 maps certification, selection and execution to their failure modes. Details of Sections 4, 5, 5.1 and 5.4 moved to Appendices C to F and I; "audit object" is now "certificate"; Section 7 states the limits in one place.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.