Pith. sign in

REVIEW 2 major objections 4 minor 63 references

This paper proves that a principal gains exactly the proxy's within-range regret by misreporting, so honest self-description is optimal precisely when the proxy already plays the best action it can reach.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 02:18 UTC pith:A3MTHLAG

load-bearing objection A clean, honest paper whose central identity is true by construction and whose headline trilemma needs the uniqueness qualifier it only prints in the body. the 2 major comments →

arxiv 2607.14357 v1 pith:A3MTHLAG submitted 2026-07-15 cs.GT cs.MAecon.TH

When Is Delegated Play Truthful? Within-Range Regret and the Trilemma of Aligned Delegation

Taksch Dube This is my paper
classification cs.GT cs.MAecon.TH MSC 91B0391A2668Q17
keywords delegated mechanism designwithin-range regretincentive compatibilityrevelation principleautobiddinglanguage-model agentsguardrailsalignment
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper studies what happens when people delegate decisions to automated proxies—autobidders that place ads, or language-model agents that act for a user—and asks when it is in the principal's interest to describe their true preferences to the proxy. The central result is an identity: the most a principal can gain by misreporting equals the proxy's within-range regret, the gap between the proxy's honest-report outcome and the best outcome reachable through some report. Consequently, honest reporting is optimal exactly when the proxy already plays the best action it can reach, which the paper calls being loyal over its range. From this identity follows a trilemma: no guardrail placed on a proxy—a bid cap, a safety filter, an alignment layer—can simultaneously be binding, truthful, and capability-preserving; any two of the three preclude the third. If the account is right, incentive compatibility becomes a measurable property of any deployed proxy rather than an assumption, and safety constraints that leave the best output reachable create exactly the incentive behind prompt-engineering and jailbreaking.

Core claim

The paper's core claim is Theorem 1: for every principal, type, and proxy, the manipulation gain G_i(θ_i;π)—the most the principal gains by reporting a false type—equals the proxy's within-range regret W_i(θ_i;π), the shortfall of the proxy's honest-report action against the best utility over its reachable set. It follows that truthful reporting is interim optimal at a type if and only if the proxy is range-optimal there, i.e., its honest action maximizes the principal's interim utility over the actions any report can steer it to. From this, Theorem 2 shows that no guardrail can be at once binding (displacing the truthful action from the honest-optimal reachable action), truthful (honest rep

What carries the argument

The central object is the proxy's within-range regret W_i(θ_i;π), defined as the shortfall of the proxy's honest-report action against the best action in its reachable set Reach_i = π_i(Θ_i). The load-bearing identity is G_i = W_i (Theorem 1), which equates the principal's manipulation gain—the maximum utility increase from misreporting—with this regret. The reachable set is the mechanism: the report language determines what a principal can steer the proxy to do, and the wrapped mechanism g∘π coarsens the base game to exactly those messages. A guardrail is a composition c∘π_i that only narrows reach, so by the identity it either leaves the honest action optimal or creates a positive gain fro

Load-bearing premise

The trilemma's 'any two preclude the third' statement assumes the honest-optimal reachable action is unique; if two actions tie, a guardrail can be binding, truthful, and capability-preserving at once, as the paper's own proof makes clear.

What would settle it

Construct a setting with two distinct reachable actions tied for the honest optimum, and a guardrail that maps the honest action to the other tied action without removing either from reach. If honest reporting still has zero within-range regret, the guardrail is binding, truthful, and capability-preserving simultaneously, refuting the trilemma as an unqualified statement.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Truthfulness of delegated play is equivalent to the proxy being range-optimal; the identity unifies first- and second-price autobidding results as a sign change in within-range regret.
  • Every guardrail that changes behavior while keeping the best outcome reachable makes honest reporting strictly suboptimal, so capability-preserving safety constraints create an incentive to inflate or rephrase reports.
  • Exact certification of incentive compatibility is #P-hard, but sampling estimates with finite-sample concentration bounds can certify a wrapped mechanism, and re-certification cost scales with proxy drift, not update frequency.
  • A capability-preserving guardrail can destroy welfare without ever binding: in a coordination game, deleting an off-equilibrium message can eliminate efficient equilibria even though every agent's unique best reply remains reachable.
  • In dominant-strategy mechanisms, a capability-preserving guardrail is outcome-neutral: it changes only the report that realizes the dominant action, typically inflating it.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If within-range regret is the right quantity, a testable design rule follows: safety layers that compress or filter a model's output should be applied inside the model's optimization (constraint-aware decoding) rather than as a post-filter, if honest prompting is a goal; the paper's placement ablation already points this way.
  • The identity suggests a 'reveal only as much as needed' policy: a user should disclose enough that their optimum lies in the proxy's reach, but no more, since extra reachable actions enlarge the surface a guardrail can bind—a consequence the paper states as an open design problem.
  • Measuring within-range regret from a model's refusal behavior, rather than from an imposed cap, would test whether deployed alignment layers actually create report-inflation incentives in the wild; the paper explicitly leaves that measurement open.
  • The trilemma's uniqueness caveat implies that when two actions tie for best, a guardrail that merely breaks the tie can be safe, honest, and capability-preserving at once—an empirical question about whether real models ever present exact ties.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper studies delegated play in a "wrapped" direct mechanism Γ^π = (Θ, g∘π), in which each principal reports a type to a fixed proxy π_i that maps the report to a message. The central object is the proxy's within-range regret W_i(θ_i;π) = max_{m∈Reach_i} U_i(m,π_{-i}|θ_i) − U_i(π_i(θ_i),π_{-i}|θ_i), where Reach_i = π_i(Θ_i). Theorem 1 asserts that the principal's manipulation gain G_i(θ_i;π) equals W_i(θ_i;π), so honest reporting is interim optimal exactly when the proxy is range-optimal. From this the paper derives a guardrail trilemma (Theorem 2): at a type with a unique honest-optimal reachable action, no guardrail is simultaneously binding, truthful, and capability-preserving. It then gives a quantitative version (Prop. 2), placement results (Prop. 3), equilibrium consequences (Props. 4–5), hardness and sampling-certificate results (Props. 6–7), and an empirical study on 25 production language models under an imposed soft bid cap, reporting positive within-range regret and report inflation on every model.

Significance. If read with its stated qualifications, the paper offers a clean unifying formalism: the first-/second-price autobidding reversal becomes a sign change in W_i, and the trilemma gives useful language for thinking about guardrails. The paper is also genuinely strong on execution: the proofs are short, the reproducibility archive is explicit, the estimation section engages with one-sided error and drift, and the empirical section clearly disclaims that the cap is imposed rather than measured. The conceptual contribution is real, although Theorem 1 is very close to a definitional identity. However, the headline trilemma as advertised in the abstract and Figure 1(b) is false without the uniqueness assumption, and the abstract must be corrected before the claims as stated are supportable.

major comments (2)
  1. [Abstract and §5, Theorem 2] The trilemma is advertised unconditionally in the abstract and Figure 1(b): "No guardrail can be at once binding, truthful, and capability-preserving." The formal statement in §5 is qualified: a⋆ is "assumed unique," and the proof of Theorem 2 uses uniqueness at the step "since a⋆ is the unique maximizer over Reach_i ⊇ Reach'_i, the two coincide." The qualifier is load-bearing. If Reach_i contains two equally good actions a and b, the honest action is a, and c swaps a and b, then c is binding (c(a)=b≠a⋆), truthful (b is optimal in Reach'_i), and capability-preserving (a∈Reach'_i), so all three properties hold. The manuscript must either carry the uniqueness assumption through the abstract and Figure 1, or redefine Binding with respect to the whole argmax set and restate the trilemma accordingly. As printed, the headline claim is not supported.
  2. [§4.2, Definition 1 and Theorem 1] Theorem 1(i), G_i = W_i, is true by construction: since Reach_i = π_i(Θ_i), the maxima defining G_i and W_i are taken over the same set, and the equality is immediate from Definition 1. The paper should not present this as a substantive discovery or as the thing that "shows" the answer turns on within-range regret. The real content lies in the interpretation, the corollaries, and the later results. Presenting the identity as a theorem without acknowledging its definitional character invites the circularity objection. I recommend restating it as a definitional equivalence and focusing the novelty claim on the consequences, which are substantive.
minor comments (4)
  1. [Figure 1(b) and abstract] The caption and the abstract should carry the uniqueness qualifier once Theorem 2 is restated; otherwise the figure will continue to overstate the result.
  2. [§7, empirical framing] The paper is appropriately careful that the cap is "alignment-style" and imposed by the authors. The abstract's final sentence could nonetheless be misread; consider saying "under a cap we impose to mimic an alignment layer" there as well, not only in the body.
  3. [§7, Table 2] The placement ablation compares the cap's compressed range with a four-point menu, so the allowed sets differ. The paper justifies this by noting that v/2 lies in both sets for v<300, which is true and makes the comparison meaningful. Still, the text should state more explicitly that the equivalence of the argmax over the two sets, not the sets themselves, is what isolates placement.
  4. [§6, Proposition 7] The drift guarantee depends on the per-update drift Δ being measurable and on L being known. For LLM proxies this is not operational. A short remark noting that the empirical study does not instantiate the drift re-certification would help calibrate expectations.

Circularity Check

1 steps flagged

Theorem 1 is a definitional identity: G_i and W_i maximize the same utility over the same set (Reach_i = π_i(Θ_i)) and subtract the same baseline, so the central 'truth-to-proxy' claim is true by construction; the trilemma, hardness, and measurement contributions are independent, making the circularity partial.

specific steps
  1. self definitional [Section 3 (Reach_i = π_i(Θ_i)); Section 4.2 Definition 1 and Theorem 1; Appendix B proof of Theorem 1]
    "as the report t_i ranges over Θ_i, the message π_i(t_i) ranges over exactly π_i(Θ_i)=Reach_i and nothing else, so choosing a report is choosing a reachable action. (i) Maximizing utility over reports equals maximizing over Reach_i ... subtracting the honest payoff ... turns the two sides into G_i and W_i, so G_i = W_i."

    Definition 1 defines W_i as the max over Reach_i of U_i minus the honest payoff, and G_i as the max over reports of the same U_i minus the same honest payoff. Since Reach_i is defined as π_i(Θ_i), the two maxima are over the same set; the equality is a restatement of the definitions, not a derived prediction. Theorem 1(iii) ('truthful iff range-optimal') likewise just restates W_i=0. The identity is load-bearing for the abstract's claim that within-range regret 'decides' honest reporting, but it carries no content beyond the definitions; the paper is transparent in calling it an identity.

full rationale

The only genuinely circular step is the central identity: G_i = W_i is true by construction because both quantities are the same optimization over the same set minus the same baseline. The paper itself concedes in Related Work that the identity 'generalizes and renames their reversal, and we claim no more for it,' so the unification claim is transparently a renaming rather than an independent derivation. I do not count Theorem 2's abstract-vs-statement mismatch as circularity: Section 5 explicitly assumes a unique honest-optimal reachable action ('assumed unique'), and the unqualified trilemma in the abstract/Figure 1(b) is a scope/correctness issue, not a reduction to inputs; a permutation guardrail on two tied optima would violate the advertised headline but not the proved theorem. Section 7's empirical arms are transparently illustrative ('the cap is one we impose ... illustrates the mechanism the identity predicts rather than measuring a deployed alignment layer'), and no parameter is fitted and then renamed a prediction. The author's own code/data citations [29,30] are reproducibility support, not load-bearing premises; the Huang et al. discussion is external. The trilemma, conservation law, #P-hardness, and drift re-certification results are derived from stated assumptions and do not presuppose the identity's content, which is why the score is 6 rather than 8-10.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The central theorems are analytic consequences of definitions; no fitted parameters enter the theory. The empirical magnitudes depend on hand-chosen cap parameters, the report grid, and the competitor distribution. The uniqueness of a⋆ is the critical assumption for the headline trilemma; the paper flags it in the formal theorem but drops it in the abstract.

free parameters (4)
  • soft-cap threshold = 40
    Chosen by hand in Section 7 as the point where the alignment-style cap begins; sets the binding region of the empirical demonstration.
  • soft-cap slope = 0.4
    Chosen by hand in Section 7; the compression factor above the threshold; controls W_i magnitudes.
  • report grid = 25 evenly spaced reports 50..650
    Chosen in Section 7; defines Reach_i empirically as the set of model bids on this grid; W_i is a grid quantity.
  • competitor distribution = Uniform[0,300] headline; [0,200] robustness
    Scenario choice in Section 7; determines payoff and the exact W_i values; not part of the theory.
axioms (6)
  • domain assumption The platform commits to running the published wrapped mechanism g∘π.
    Section 1: 'One assumption holds throughout: the platform commits to running the published g∘π.' Without this, the input game is not well-defined; standard mechanism design commitment.
  • domain assumption Types and messages are finite; extension to continua wherever maxima are attained.
    Section 3 Preliminaries; needed for max to be well-defined and for hardness reductions.
  • domain assumption Common prior F and private values u_i(x,θ_i)∈[0,H].
    Section 3; the Bayesian game and all expectations are defined under this.
  • domain assumption For Theorem 2 and Proposition 3, the honest-optimal reachable action a⋆ is unique.
    Section 5: 'Fix a type θ_i and let a⋆=arg max ... assumed unique.' If not unique, a guardrail can permute among optimal actions and be binding/truthful/capability-preserving, so the unqualified trilemma fails.
  • domain assumption For Proposition 7, utility is L-Lipschitz in the message profile with metrics d_j and a total drift budget.
    Section 6 Lemma 3; the online re-certification bound requires this.
  • standard math P≠NP and FP≠#P for the hardness statement.
    Proposition 6's 'no polynomial-time exact certifier' is conditional on standard complexity separations.

pith-pipeline@v1.3.0-alltime-deepseek · 23316 in / 18792 out tokens · 177252 ms · 2026-08-02T02:18:54.142105+00:00 · methodology

0 comments
read the original abstract

Advertisers delegate bidding to autobidders; users delegate tasks to language-model agents. A person describes what they want to an automated proxy that acts in a mechanism on their behalf. This is the revelation principle in production, and it forces a question classical theory assumes away: when is it optimal to describe yourself honestly to your own proxy? We show the answer turns on one quantity, the proxy's within-range regret. The most a principal can gain by misreporting equals the regret of the proxy's honest-report action against those the principal could have steered it to take. Honest self-description is optimal exactly when the proxy already plays the best action it can reach, that is, when it is loyal (Theorem 1). The identity unifies auction-specific autobidding results and pins down when the faithful-communication assumption behind language-model elicitation proxies (Huang et al.) holds. The identity constrains guardrails placed on proxies, from bid caps to a model's alignment layer. No guardrail can be at once binding (it displaces the truthful action from the proxy's best reachable outcome), truthful (honest reporting stays optimal), and capability-preserving (that outcome stays reachable through some report); any two preclude the third (Theorem 2). A safety constraint that alters what a model does while leaving its best output reachable makes honest description of intent suboptimal, so a sharper report can gain. This is the incentive behind prompt-engineering and jailbreaking. Because within-range regret is #P-hard to compute exactly, we estimate it from samples and maintain it as a model is updated, at a cost set by how far the model drifts, not how often it changes. Running it on production language models from five providers under an alignment-style cap, we find honest reporting leaves surplus unclaimed on every model, recovered by inflating the report.

Figures

Figures reproduced from arXiv: 2607.14357 by Taksch Dube.

Figure 1
Figure 1. Figure 1: (a) The delegation stack: each principal reports its type to a proxy, which messages the mechanism g, so principals face the wrapped mechanism g ◦ π. (b) A guardrail composed onto a proxy cannot be binding, truthful, and capability-preserving at once (Theorem 2); which pair a design keeps depends on its placement and its allowed set D (Proposition 3). Constraint-aware decoding stays truthful; an output-cli… view at source ↗
Figure 2
Figure 2. Figure 2: Within-range regret is what governs misreporting. The proxy reaches only Reachi ⊆ Mi, so a principal’s gain from misreporting equals the gap Wi to the best reachable action, not the unrestricted regret Ri to the global optimum a ⋆ M (Theorem 1). and the advertiser strictly gains by reporting b ⋆ in place of v. Honest self￾description is optimal exactly in the format whose optimal bid is the honestly report… view at source ↗
Figure 3
Figure 3. Figure 3: Within-range regret Wi for 25 production language models across five providers, each an autobidding proxy under the alignment-style soft cap (mean over true values v ∈ {100, . . . , 300}; exact values labeled). Honest reporting is suboptimal on every model: the capable ones bid the rational v/2 and cluster at Wi = 5.79, just under the analytic optimal-play regret Ri = 5.82 (dashed; a mean over v, where the… view at source ↗
Figure 4
Figure 4. Figure 4: The mechanism behind [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: (a) The Ri − Wi gap across 60,000 random games and (b) Wi against Ri with the W = R diagonal. The proxy’s range removes real regret in most games, so Wi, not Ri, governs misreporting. 2γ 1 2 1 1.5 2 true value v report inflation t ⋆/v (a) report inflation 2γ 1 2 0 0.1 0.2 true value v within-range regret W (b) within-range regret [PITH_FULL_IMAGE:figures/full_fig_p026_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: (a) Report inflation and (b) within-range regret against the true value v for the soft-cap proxy, both flat until the cap binds at v = 2γ (dashed) and rising past it; within-range regret grows as W = 0.09 (v − 2γ) 2 in this example, the quadratic being the (ε, µ 2 ε 2 )-sharpness of Proposition 2 with binding distance ε ∝ v − 2γ. Within-range regret is genuinely smaller than unrestricted regret ( [PITH_FU… view at source ↗
Figure 7
Figure 7. Figure 7: (a) The sampling certificate’s error against N, decaying at the 1/ √ N rate. (b) The re-certification count against the drift budget BT , staying under the Proposition 7 bound and scaling with BT , not the number of updates. report inflation both rise with v ( [PITH_FULL_IMAGE:figures/full_fig_p027_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

63 extracted references · 2 canonical work pages

  1. [1]

    In: Proceedings of the 15th Conference on Web and Internet Economics (WINE)

    Aggarwal, G., Badanidiyuru, A., Mehta, A.: Autobidding with constraints. In: Proceedings of the 15th Conference on Web and Internet Economics (WINE). pp. When Is Delegated Play Truthful? 21 17–30 (2019)

    2019

  2. [2]

    Econometrica88(2), 425– 467 (2020)

    Akbarpour, M., Li, S.: Credible auctions: A trilemma. Econometrica88(2), 425– 467 (2020)

    2020

  3. [3]

    In: Proceedings of the 2023 ACM Conference on Economics and Computation (EC) (2023), arXiv:2301.13414

    Alimohammadi, Y., Mehta, A., Perlroth, A.: Incentive compatibility in the auto- bidding world. In: Proceedings of the 2023 ACM Conference on Economics and Computation (EC) (2023), arXiv:2301.13414

    Pith/arXiv arXiv 2023

  4. [4]

    The Review of Economic Studies 75(1), 259–293 (2008)

    Alonso, R., Matouschek, N.: Optimal delegation. The Review of Economic Studies 75(1), 259–293 (2008)

    2008

  5. [5]

    Angelopoulos, A.N., Bates, S.: A gentle introduction to conformal prediction and distribution-free uncertainty quantification (2021), arXiv:2107.07511

    Pith/arXiv arXiv 2021

  6. [6]

    Econometrica 78(1), 213–244 (2010)

    Armstrong, M., Vickers, J.: A model of delegated project choice. Econometrica 78(1), 213–244 (2010)

    2010

  7. [7]

    Games and Economic Behavior67(1), 2–21 (2009)

    Ashlagi, I., Monderer, D., Tennenholtz, M.: Mediators in position auctions. Games and Economic Behavior67(1), 2–21 (2009)

    2009

  8. [8]

    Frontiers of Theoretical Economics1(1) (2002)

    Ausubel, L.M., Milgrom, P.R.: Ascending auctions with package bidding. Frontiers of Theoretical Economics1(1) (2002)

    2002

  9. [9]

    In: Proceedings of the 2019 ACM Conference on Economics and Com- putation (EC)

    Balcan, M.F., Sandholm, T., Vitercik, E.: Estimating approximate incentive com- patibility. In: Proceedings of the 2019 ACM Conference on Economics and Com- putation (EC). p. 867 (2019), full version: arXiv:1902.09413

    Pith/arXiv arXiv 2019

  10. [10]

    Econometrica69(4), 1077–1098 (2001)

    Bester, H., Strausz, R.: Contracting with imperfect commitment and the revelation principle: The single agent case. Econometrica69(4), 1077–1098 (2001)

    2001

  11. [11]

    In: Proceedings of the 2018 ACM Conference on Economics and Computation (EC)

    Braverman,M.,Mao,J.,Schneider,J.,Weinberg,S.M.:Sellingtoano-regretbuyer. In: Proceedings of the 2018 ACM Conference on Economics and Computation (EC). pp. 523–538 (2018)

    2018

  12. [12]

    In: IEEE 61st Annual Symposium on Foundations of Computer Science (FOCS)

    Camara, M.K., Hartline, J.D., Johnsen, A.: Mechanisms for a no-regret agent: Beyond the common prior. In: IEEE 61st Annual Symposium on Foundations of Computer Science (FOCS). pp. 259–270 (2020)

    2020

  13. [13]

    Cambridge Uni- versity Press (2006)

    Cesa-Bianchi, N., Lugosi, G.: Prediction, Learning, and Games. Cambridge Uni- versity Press (2006)

    2006

  14. [14]

    PNAS Nexus5(4), pgag097 (2026)

    Chen, P.Y., Shen, H., Das, P., Chen, T.: Fundamental safety-capability trade- offs in fine-tuning large language models. PNAS Nexus5(4), pgag097 (2026). https://doi.org/10.1093/pnasnexus/pgag097

    work page doi:10.1093/pnasnexus/pgag097 2026

  15. [15]

    Journal of the ACM56(3), 14:1–14:57 (2009)

    Chen, X., Deng, X., Teng, S.H.: Settling the complexity of computing two-player nash equilibria. Journal of the ACM56(3), 14:1–14:57 (2009)

    2009

  16. [16]

    In: Proceedings of the 18th Conference on Uncertainty in Artificial Intelligence (UAI)

    Conitzer, V., Sandholm, T.: Complexity of mechanism design. In: Proceedings of the 18th Conference on Uncertainty in Artificial Intelligence (UAI). pp. 103–110 (2002), arXiv:cs/0205075

    Pith/arXiv arXiv 2002

  17. [17]

    In: Proceedings of the 5th ACM Conference on Electronic Commerce (EC)

    Conitzer, V., Sandholm, T.: Computational criticisms of the revelation principle. In: Proceedings of the 5th ACM Conference on Electronic Commerce (EC). pp. 262–263 (2004)

    2004

  18. [18]

    Econometrica56(6), 1247–1257 (1988)

    Crémer, J., McLean, R.P.: Full extraction of the surplus in bayesian and dominant strategy auctions. Econometrica56(6), 1247–1257 (1988)

    1988

  19. [19]

    In: Advances in Neural Information Processing Systems 33 (NeurIPS) (2020), arXiv:2006.08742

    Curry, M., Chiang, P.Y., Goldstein, T., Dickerson, J.: Certifying strategyproof auction networks. In: Advances in Neural Information Processing Systems 33 (NeurIPS) (2020), arXiv:2006.08742

    Pith/arXiv arXiv 2020

  20. [20]

    The Review of Economic Studies 46(2), 185–216 (1979)

    Dasgupta, P., Hammond, P., Maskin, E.: The implementation of social choice rules: Some general results on incentive compatibility. The Review of Economic Studies 46(2), 185–216 (1979)

    1979

  21. [21]

    SIAM Journal on Computing39(1), 195–259 (2009) 22 T

    Daskalakis,C.,Goldberg,P.W.,Papadimitriou,C.H.:Thecomplexityofcomputing a nash equilibrium. SIAM Journal on Computing39(1), 195–259 (2009) 22 T. Dube

    2009

  22. [22]

    Jour- nal of Public Economics11(1), 25–45 (1979)

    d’Aspremont, C., Gérard-Varet, L.A.: Incentives and incomplete information. Jour- nal of Public Economics11(1), 25–45 (1979)

    1979

  23. [23]

    In: Proceedings of the Web Conference 2020 (WWW)

    Deng, Y., Lahaie, S., Mirrokni, V., Zuo, S.: A data-driven metric of incentive compatibility. In: Proceedings of the Web Conference 2020 (WWW). pp. 1796– 1806 (2020)

    2020

  24. [24]

    In: Proceedings of the ACM Web Conference 2024 (WWW) (2024), arXiv:2311.10679

    Deng, Y., Mao, J., Mirrokni, V., Teng, Y., Zuo, S.: Non-uniform bid-scaling and equilibria for different auctions: An empirical study. In: Proceedings of the ACM Web Conference 2024 (WWW) (2024), arXiv:2311.10679

    Pith/arXiv arXiv 2024

  25. [25]

    Deng, Y., Mao, J., Mirrokni, V., Zhang, H., Zuo, S.: Efficiency of the first-price auction in the autobidding world (2022), arXiv:2208.10650

    Pith/arXiv arXiv 2022

  26. [26]

    In: Proceedings of the Web Conference 2021 (WWW)

    Deng, Y., Mao, J., Mirrokni, V., Zuo, S.: Towards efficient auctions in an auto- bidding world. In: Proceedings of the Web Conference 2021 (WWW). pp. 3965– 3973 (2021)

    2021

  27. [27]

    In: Ad- vances in Neural Information Processing Systems 32 (NeurIPS) (2019)

    Deng, Y., Schneider, J., Sivan, B.: Strategizing against no-regret learners. In: Ad- vances in Neural Information Processing Systems 32 (NeurIPS) (2019)

    2019

  28. [28]

    Econometrica 90(4), 1463–1500 (2022)

    Doval, L., Skreta, V.: Mechanism design with limited commitment. Econometrica 90(4), 1463–1500 (2022)

    2022

  29. [29]

    Zenodo,https://doi.org/ 10.5281/zenodo.21380456(2026)

    Dube, T.: Delegated-play experiments: Code and data. Zenodo,https://doi.org/ 10.5281/zenodo.21380456(2026)

    work page doi:10.5281/zenodo.21380456(2026 2026

  30. [30]

    Dube, T.: A research library for simulating markets populated by LLM and zero- intelligence agents.https://github.com/takschdube/llm-market-sim(2026), version 0.2.0

    2026

  31. [31]

    Dütting, P., Mirrokni, V., Paes Leme, R., Xu, H., Zuo, S.: Mechanism design for largelanguagemodels.In:ProceedingsoftheACMWebConference2024(WWW). pp. 144–155 (2024), arXiv:2310.10826

    Pith/arXiv arXiv 2024

  32. [32]

    Feng, Z., Schrijvers, O., Sodomka, E.: Online learning for measuring incentive compatibility in ad auctions (2019), arXiv:1901.06808

    Pith/arXiv arXiv 2019

  33. [33]

    American Eco- nomic Review77(5), 927–940 (1987)

    Fershtman, C., Judd, K.L.: Equilibrium incentives in oligopoly. American Eco- nomic Review77(5), 927–940 (1987)

    1987

  34. [34]

    Fish, S., Gonczarowski, Y.A., Shorrer, R.I.: Algorithmic collusion by large language models (2024), arXiv:2404.00806

    arXiv 2024

  35. [35]

    Econometrica54(6), 1375– 1385 (1986)

    Forges, F.: An approach to communication equilibria. Econometrica54(6), 1375– 1385 (1986)

    1986

  36. [36]

    American Economic Review104(1), 66–83 (2014)

    Frankel, A.: Aligned delegation. American Economic Review104(1), 66–83 (2014)

    2014

  37. [37]

    Econometrica 41(4), 587–601 (1973)

    Gibbard, A.: Manipulation of voting schemes: A general result. Econometrica 41(4), 587–601 (1973)

    1973

  38. [38]

    In: Proceedings of the 2015 ACM Conference on Economics and Computation (EC) (2015), arXiv:1504.07687

    Gopalan, P., Nisan, N., Roughgarden, T.: Public projects, Boolean functions, and the borders of Border’s theorem. In: Proceedings of the 2015 ACM Conference on Economics and Computation (EC) (2015), arXiv:1504.07687

    Pith/arXiv arXiv 2015

  39. [39]

    In: Proceedings of the 2016 ACM Conference on Innovations in Theoretical Com- puter Science (ITCS)

    Hardt, M., Megiddo, N., Papadimitriou, C., Wootters, M.: Strategic classification. In: Proceedings of the 2016 ACM Conference on Innovations in Theoretical Com- puter Science (ITCS). pp. 111–122 (2016), arXiv:1506.06980

    Pith/arXiv arXiv 2016

  40. [40]

    In: Boyer, M., Kihlstrom, R.E

    Holmström, B.: On the theory of delegation. In: Boyer, M., Kihlstrom, R.E. (eds.) Bayesian Models in Economic Theory, pp. 115–141. North-Holland (1984)

    1984

  41. [41]

    Hoshino, Y., Kitadai, A., Nishino, N.: Do matching mechanisms work with LLM agents? (2026), arXiv:2606.03030

    Pith/arXiv arXiv 2026

  42. [42]

    In: Proceedings of the 21st Conference on Web and Internet Economics (WINE) (2025), arXiv:2501.14625 When Is Delegated Play Truthful? 23

    Huang, D., Marmolejo-Cossío, F., Lock, E., Parkes, D.C.: Accelerated preference elicitation with LLM-based proxies. In: Proceedings of the 21st Conference on Web and Internet Economics (WINE) (2025), arXiv:2501.14625 When Is Delegated Play Truthful? 23

    arXiv 2025

  43. [43]

    In: Proceedings of the 2016 ACM Conference on Economics and Computation (EC)

    Kephart, A., Conitzer, V.: The revelation principle for mechanism design with reporting costs. In: Proceedings of the 2016 ACM Conference on Economics and Computation (EC). pp. 85–102 (2016)

    2016

  44. [44]

    In: Proceedings of the 2018 ACM Conference on Economics and Computation (EC)

    Kleinberg, J., Kleinberg, R.: Delegated search approximates efficient search. In: Proceedings of the 2018 ACM Conference on Economics and Computation (EC). pp. 287–302 (2018)

    2018

  45. [45]

    Kleine Büning, T., Gan, J., Mandal, D., Kwiatkowska, M.: Strategyproof reinforce- ment learning from human feedback (2025), arXiv:2503.09561

    arXiv 2025

  46. [46]

    In: Pro- ceedings of the ACM Web Conference 2022 (WWW)

    Kolumbus, Y., Nisan, N.: Auctions between regret-minimizing agents. In: Pro- ceedings of the ACM Web Conference 2022 (WWW). pp. 100–111 (2022), arXiv:2110.11855

    Pith/arXiv arXiv 2022

  47. [47]

    In: Advances in Neural Information Processing Systems 35 (NeurIPS) (2022), arXiv:2112.07640

    Kolumbus, Y., Nisan, N.: How and why to manipulate your own agent: On the in- centives of users of learning agents. In: Advances in Neural Information Processing Systems 35 (NeurIPS) (2022), arXiv:2112.07640

    Pith/arXiv arXiv 2022

  48. [48]

    In: Proceedings of the 33rd International Joint Conference on Artificial Intelligence (IJCAI)

    Li, J., Tang, P.: Vulnerabilities of single-round incentive compatibility in auto- bidding: Theory and evidence from ROI-constrained online advertising markets. In: Proceedings of the 33rd International Joint Conference on Artificial Intelligence (IJCAI). pp. 2886–2894 (2024), arXiv:2210.06107

    Pith/arXiv arXiv 2024

  49. [49]

    Liaw, C., Zhu, W.: Risk-averse and optimistic advertiser incentive compatibility in auto-bidding (2025), arXiv:2508.16823

    Pith/arXiv arXiv 2025

  50. [50]

    In: Proceedings of the 22nd Conference on Learning Theory (COLT) (2009)

    Maurer, A., Pontil, M.: Empirical Bernstein bounds and sample variance penaliza- tion. In: Proceedings of the 22nd Conference on Learning Theory (COLT) (2009)

    2009

  51. [51]

    Econometrica 47(1), 61–73 (1979)

    Myerson, R.B.: Incentive compatibility and the bargaining problem. Econometrica 47(1), 61–73 (1979)

    1979

  52. [52]

    Mathematics of Operations Research6(1), 58–73 (1981)

    Myerson, R.B.: Optimal auction design. Mathematics of Operations Research6(1), 58–73 (1981)

    1981

  53. [53]

    Econometrica54(2), 323– 358 (1986)

    Myerson, R.B.: Multistage games with communication. Econometrica54(2), 323– 358 (1986)

    1986

  54. [54]

    Journal of Economic Theory129(1), 192–224 (2006)

    Nisan, N., Segal, I.: The communication requirements of efficient allocations and supporting prices. Journal of Economic Theory129(1), 192–224 (2006)

    2006

  55. [55]

    Sci- ence349(6245), 267–272 (2015)

    Parkes, D.C., Wellman, M.P.: Economic reasoning and artificial intelligence. Sci- ence349(6245), 267–272 (2015)

    2015

  56. [56]

    Pieroth, F.R., Sandholm, T.: Verifying approximate equilibrium in auctions (2024), arXiv:2408.11445

    Pith/arXiv arXiv 2024

  57. [57]

    In: Proceedings of the 57th IEEE Symposium on Foundations of Com- puter Science (FOCS)

    Rubinstein, A.: Settling the complexity of computing approximate two-player nash equilibria. In: Proceedings of the 57th IEEE Symposium on Foundations of Com- puter Science (FOCS). pp. 258–265 (2016)

    2016

  58. [58]

    Shah, A., Zhu, K., Jiang, Y., Wang, J.G., Dayi, A.K., Horton, J.J., Parkes, D.C.: Learning from synthetic labs: Language models as auction participants (2025), arXiv:2507.09083

    Pith/arXiv arXiv 2025

  59. [59]

    Review of Economic Studies88(3), 1503–1540 (2021)

    Sugaya, T., Wolitzky, A.: The revelation principle in multistage games. Review of Economic Studies88(3), 1503–1540 (2021)

    2021

  60. [60]

    Games and Economic Behavior49(2), 363– 373 (2004)

    Tennenholtz, M.: Program equilibrium. Games and Economic Behavior49(2), 363– 373 (2004)

    2004

  61. [61]

    Wei, A., Haghtalab, N., Steinhardt, J.: Jailbroken: How does LLM safety training fail? In: Advances in Neural Information Processing Systems 36 (NeurIPS) (2023)

    2023

  62. [62]

    You, S., Zhuang, Z., Wang, K., Wang, Z.: Bridging the gap between estimated and true regret towards reliable regret estimation in deep learning based mechanism design (2026), arXiv:2601.13489 24 T. Dube

    arXiv 2026

  63. [63]

    Zou, A., Wang, Z., Carlini, N., Nasr, M., Kolter, J.Z., Fredrikson, M.: Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043 (2023) When Is Delegated Play Truthful? 25 A Additional empirical detail This appendix collects the supporting figures for the empirical study of Section 7: the within-range gap ...

    Pith/arXiv arXiv 2023