acceptodds
Under review as a conference paper at ICLR 2027

Mechanism Parameter Optimization under Behavioral Uncertainty via Sequential Experimental Design

Abstract

In strategic environments, mechanism parameter optimization is important but challenging because the designer must choose parameters, such as auction reserve prices, to maximize a design objective while the optimal choice depends on unknown strategic behavior. The high cost of collecting strategic behavioral data limits the number of available experiments, making it difficult to simultaneously reduce behavioral uncertainty and optimize mechanism parameters. We propose a framework to address this. First, we leverage large language models to generate a pool of candidate parametric behavioral models, without requiring it to contain the true behavioral model. This converts open-ended behavioral uncertainty into a finite pool of behavioral hypotheses. Given this pool, we develop two sequential experimental design solutions to optimize mechanism parameters. The first follows a two-stage approach that selects experiments solely to reduce uncertainty about the behavioral model pool before optimizing the mechanism parameter. The second takes an end-to-end approach that selects experiments to directly reduce uncertainty about the optimal mechanism parameter. To reduce computational complexity, we approximate this experiment selection criterion using a surrogate parameter space. We further provide a theoretical guarantee that the approximation preserves the ranking of experimental designs under suitable conditions. Experiments in two strategic scenarios (static first-price auctions and dynamic pay-per-bid auctions) demonstrate that our methods consistently outperform multiple classes of baselines in mechanism parameter optimization.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Large Language Models, Game Theory, Mechanism Design

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

The paper proposes a framework to optimize the mechanism parameters with limited budget and unknown behavior models. The framework consists of two steps. First, use LLMs to generate many candidate behavior models. Second, use Bayesian updates to update the posterior belief of the behavior models. The optimal mechanism parameters are determined given the posterior belief of the behavior models. The paper also considers two methods in second step: BUFS and MUFS. BUFS is more intuitive, while MUFS is more "end-to-end". Experiments show that BUFS and MUFS outperform other baselines in or out of sequential experimental design literature.

Critical strengths
  1. The proposed framework for sequential experimental design is quite interesting, though not very groundbreaking. I learn something new from the presentation. I would like to give a "weak acceptance" if all my concerns could be resolved.
Critical weaknesses
  1. The paper claims that "using LLMs to generate behavior models" is part of its contribution. My concern is that this does not constitute a real contribution, because it is quite natural and standard. One could also generate behavior models through expert knowledge (e.g., enumerating the behavior models in existing literature). Furthermore, this paper does not validate whether the LLM-generated behavior models are better than other alternatives (e.g., enumerating the behavior models in existing literature), and the advantage of using LLMs is not very clear.

  2. I'm unsure about the marginal contribution of this paper over existing literature. For this reason, I can not recommend acceptance at this stage. Specifically, in "Sequential Experimental Design" paragraph in "Related Works" section, the paper distinguishes itself from existing literature by arguing that existing literature require a ground truth behavior model, while this paper does not. The remedy is to use LLMs to generate several candidate behavior models. I do not see this contribution enough for acceptance (see my weaknesses 1).

  3. I would like to see whether MUFS is a new approach for sequential experimental design. However, the paper does not give an affirmative answer to this question.

  4. It's unclear whether the current baselines are appropriate. It seems to be that current baselines are weak. As the paper mentions that there have been many works in sequential experimental design, and regards proposing BUFS/MUFS as part of their contributions, it is important to compare BUFS/MUFS with some of them.

Questions for authors

Not really questions, but I do not know where to put them, and these points do not affect my rating of the paper.

  1. In equation (3), the paper introduces a metric to measure the quality of mechanism parameter zz, but do not evaluate it in experiments in the main text. I had expected to see this in the main text.
  2. The motivation for introducing Z~\tilde{Z} as a surrogate of ZZ is confusing. The paper claims "Since the model-parameter space is high-dimensional, this results in a large number of optimization solves, making the overall computation particularly challenging", however, replacing ZZ with Z~\tilde{Z} still requires a large number of optimization solves.
  3. The introduction of Proposition 1 is confusing as well. On one hand, the assumption seems quite strong to me; on other hands, the connection between Proposition 1 and the other parts of the paper is unclear. Furthermore, the statement "then the strict pairwise rankings of experiments induced by z~∗\tilde{z}^* are preserved under z∗z^*" is not a rigorous statement, and one should avoid this.
  4. The experiment configurations are not clear enough in the main text. For example, the authors do not introduce the value distributions, the number of bidders, the behavior models generated by LLMs, the value of K,E,ZK,\mathcal{E},\mathcal{Z}, the prior set to the behavior models and the parameters, that are important to uniquely define the experiments. In addition, in all experiments, ee and zz seem to refer the same thing, so it is confusing why the paper introduces two concepts e,ze,z in the main text.
  5. The "Problem Formulation" section is too abstract, I believe that a concrete example can bring a better understanding to the readers.
  6. The main text up to section 5 is a bit wordy, e.g., repeating the same meaning many times. I believe that this can be improved.
8 comments
AuthorsOpenReview ↗

Authors' Answers to Question 1, Question 2, and Question 3

Q1. In equation (3), the paper introduces a metric to measure the quality of mechanism parameter, but do not evaluate it in experiments in the main text. I had expected to see this in the main text.

We thank the reviewer for this question! The metric in Eq. (3) cannot be evaluated exactly because the true agent behavioral model is unavailable. However, since all methods are evaluated against the same agents whose behavioral models are assumed to remain fixed throughout the experiments, the first term in Eq. (3) (i.e., the maximum achievable expected objective value) is constant across methods. Thus, comparing the second term (i.e., expected objective value under each method's optimized mechanism parameters) is sufficient to compare their relative gaps, with higher values indicating smaller gaps. In practice, we estimate this second term using empirical revenue averaged over 10 independent runs, each comprising 30 independent auctions.

We have added this discussion to Section 5 in the revised paper.

Q2. The motivation for introducing $\tilde Z $ as a surrogate of $Z $ is confusing. The paper claims "Since the model-parameter space is high-dimensional, this results in a large number of optimization solves, making the overall computation particularly challenging", however, replacing ZZ with Z~\tilde Z still requires a large number of optimization solves.

We thank the reviewer for this question! We acknowledge that the original explanation did not make this statement sufficiently clear. The key benefit of the surrogate space is not to reduce the number of model-parameter pairs considered, but to reduce the computational cost for each pair. Specifically, for each sampled model-parameter pair, computing $z^* $ requires solving an optimization problem over the full, fine-grained parameter space of $\mathcal Z $. In contrast, finding $\tilde z^* $ requires only evaluating and comparing the objective values over a small discrete space $\tilde{\mathcal Z} $. For example, in first-price auctions, we have $|\tilde{\mathcal Z}|=11 $. Thus, $\tilde z^* $ can be obtained by evaluating the revenue at these 11 points and selecting the one with the highest value.

We have added this discussion to Section 4.3 and Appendix D.2 in the revised paper.

Q3. The introduction of Proposition 1 is confusing as well. On one hand, the assumption seems quite strong to me; on other hands, the connection between Proposition 1 and the other parts of the paper is unclear. Furthermore, the statement "then the strict pairwise rankings of experiments induced by z~\tilde{z}" is not a rigorous statement, and one should avoid this.

We thank the reviewer for these questions about Proposition 1! We agree that the original presentation of Proposition 1 was not sufficiently clear or precise, and now we address your questions separately, as follows.

Regarding the assumptions: First, as discussed in Appendix A.3, we acknowledge that Proposition 1 is intended as theoretical motivation for the surrogate approximation rather than a strict implementation guarantee. Second, assumptions such as Lipschitz continuity are sufficient rather than necessary conditions. Moreover, to reduce the risk of substantially violating the assumptions, we can increase the density of representative points in the surrogate space. A denser set makes it more likely that z∗z^* lies close to at least one representative point, thereby improving the quality of the surrogate approximation.

Regarding its connection to MUFS: Proposition 1 justifies using the surrogate space for experiment selection in MUFS. MUFS ranks candidate experiments according to their criterion values, and selects the next experiment with the highest value. The original space Z\mathcal Z and the surrogate space Z~\tilde{\mathcal Z} each induce a ranking of candidate experiments based on their respective criterion values. Since evaluation in the original space is expensive, MUFS uses the surrogate space instead. Therefore, it is important that the surrogate-space ranking agrees with the original-space ranking. Proposition 1 establishes precisely this property under the stated conditions, ensuring that the same optimal experiment is selected when it is unique.

Regarding the statement: We agree that the wording is imprecise. We have replaced it with the following statement: Under the stated conditions, for any two candidate experiments e1,e2∈Ee_1,e_2\in\mathcal E, S(e1)<S(e2)\mathcal S(e_1)<\mathcal S(e_2) implies S~(e1)<S~(e2)\tilde{\mathcal S}(e_1)<\tilde{\mathcal S}(e_2). Thus, the surrogate space preserves all strict pairwise rankings induced by the original space, ensuring that the same optimal experiment is selected when the optimum is unique.

We have refined the content in Section 4, Appendix A.2, and Appendix A.3 in the revised paper.

AuthorsOpenReview ↗

Authors' Answers to Question 4, Question 5, and Question 6

Q4. The experiment configurations are not clear enough in the main text. For example, the authors do not introduce the value distributions, the number of bidders, the behavior models generated by LLMs, the value of K, E\mathcal{E}, ZZ, the prior set to the behavior models and the parameters, that are important to uniquely define the experiments. In addition, in all experiments, e and z seem to refer the same thing, so it is confusing why the paper introduces two concepts in the main text.

We thank the reviewer for these questions!

Regarding configurations: We agree that the experimental configurations should be presented more clearly. We had moved the key settings from Appendix C to the main text. We set K=2K=2 in Sections 5.1 and 5.2, while Appendix B.3 examines pool sizes ranging from K=2K=2 to K=4K=4. For conciseness, below we summarize required experimental settings except LLM-generated behavioral models which are specified in Appendix C.4 in the revised paper.

Config.First-price AuctionPay-per-bid Auction
Value distributioni.i.d. vi∼U600,…,1200v_i\sim U \\{600,\ldots,1200\\}.Fixed v=30v=30
Number of biddersI=10I=10I=5I=5
Mechanism parameter space (Z\mathcal Z)1,2,…,2000\\{1,2,\ldots,2000\\}.(b,d)∈0.1,0.2,…,9.02(b,d)\in\\{0.1,0.2,\ldots,9.0\\}^2.
Design space (E\mathcal{E})1,2,…,2000\\{1,2,\ldots,2000\\}.(b,d)∈0.1,0.2,…,9.02(b,d)\in\\{0.1,0.2,\ldots,9.0\\}^2.
Surrogate parameter space (Z~\tilde{\mathcal Z})1,200,400,…,2000\\{1,200,400,\ldots,2000\\}.0.1,1,2,…,92\\{0.1,1,2,\ldots,9\\}^2 .
Behavioral model priorUniform: p(φ(k))=1/Kp(\varphi^{(k)})=1/K.Uniform: p(φ(k))=1/Kp(\varphi^{(k)})=1/K.
Behavioral parameter priorIndependent uniform over θu∈[0.01,1.00]\theta_u\in[0.01,1.00] and δs∈[0.01,1.00]\delta_s\in[0.01,1.00].Independent uniform over θr∈[−0.5,0.5]\theta_r\in[-0.5,0.5] and β∈[0.01,0.5]\beta\in[0.01,0.5].

Regarding ee and zz: Although both are expressed using the same auction parameters (i.e., the reserve price in first-price auctions and the bid fee and bid increment in pay-per-bid auctions), they serve different roles in our framework. Specifically, the experiment ete_t denotes the mechanism parameter tested in round tt to collect behavioral observations and update posterior beliefs. It is selected for the purpose of learning and information collection. In contrast, z^T\hat z_T denotes the final mechanism parameter deployed after TT experimental rounds to maximize expected revenue. Therefore, the most informative experiment need not be the optimal final mechanism.

We have added relevant content to Section 5.1, Section 5.2, and Appendix D.2 in the revised paper.

Q5. The "Problem Formulation" section is too abstract, I believe that a concrete example can bring a better understanding to the readers.

We thank the reviewer for this question! We agree that a concrete example would help readers better understand the problem formulation. We have added the following example to Section 3 in the revised paper.

Consider a first-price auction where the designer aims to maximize revenue by choosing an appropriate reserve price. When the bidders' behavioral model is unknown, and the number of experimental rounds is limited, the designer sequentially conducts auctions with informative reserve prices ete_t, observes the resulting outcomes ytobsy_t^{\mathrm{obs}}, and updates beliefs over the LLM-generated candidate behavioral models and their parameters. After TT experiments, the designer uses the updated beliefs to deploy the final reserve price z^T\hat z_T that maximizes expected revenue.

Q6. The main text up to section 5 is a bit wordy, e.g., repeating the same meaning many times. I believe that this can be improved.

We thank the reviewer for this question! We agree that some parts of the main text are repetitive. We have refined relevant paragraphs in Section 3 and Section 5 to improve conciseness and clarity in the revised paper. For example, in Section 3 in the revised paper, we have replaced "Since the true behavioral model and its parameters are unknown, reducing uncertainty about them will help in reducing uncertainty about the optimal mechanism parameter" with "Since the true model-parameter pair is unknown, reducing uncertainty about it helps identify the optimal mechanism parameter".

AuthorsOpenReview ↗

Authors' Reply to Weakness 1

W1. The paper claims that "using LLMs to generate behavior models" is part of its contribution. My concern is that this does not constitute a real contribution, because it is quite natural and standard. One could also generate behavior models through expert knowledge (e.g., enumerating the behavior models in existing literature). Furthermore, this paper does not validate whether the LLM-generated behavior models are better than other alternatives (e.g., enumerating the behavior models in existing literature), and the advantage of using LLMs is not very clear.

We thank the reviewer for raising this concern! We agree that we should be more precise in positioning. Generating behavioral models with LLMs is not, by itself, a standalone technical contribution, and our key technical contribution is MUFS.

First, our conceptual contribution is to introduce the use of LLMs to generate candidate behavioral models for mechanism optimization under unknown strategic behavior and limited experimental budgets. To the best of our knowledge, we are the first to explore this idea in this problem setting and empirically demonstrate its feasibility. Specifically, LLMs offer a practical advantage: they can rapidly convert open-ended behavioral uncertainty into a diverse pool of plausible behavioral models without requiring extensive manual enumeration.

Second, our key technical contribution is MUFS. We develop sequential experimental design strategies, BUFS and MUFS, for mechanism optimization under unknown strategic behavior and limited experimental budgets. Given this candidate pool, BUFS focuses on distinguishing among them, while MUFS directly targets the optimal mechanism parameter. Please see our Reply to Weaknesses 2 & 3 for further details.

Third, generating candidate models with LLMs alone is insufficient for mechanism parameter optimization. This is further supported empirically: our methods and the sequential design baselines (i.e., Seq-XXX) use the same LLM-generated candidate model pool, yet achieve different results. Thus, the empirical performance improvements cannot be attributed solely to LLM-based model generation, but also reflect the effectiveness of our sequential experimental design strategies.

We have refined the positioning in Section 1 of the revised paper.

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

This paper proposes a sequential experimental design for optimally selecting parameters of a mechanism (such as the reserve price of an auction) when facing participants with behavioral uncertainty. It uses LLMs to generate a set of participant behavior models and then develops two experimental design methods to select experiments, observations from which will inform the final selection of mechanism parameters. The paper combines Bayesian experimental design with behavioral uncertainty for the application of mechanism design.

Critical strengths
  • The paper combines explicit behavioral hypotheses/models with sequential experimental design for mechanism optimization. This approach is interesting.

  • The paper’s experiments provide useful evidence that experimental selection matters (at least for LLM agents participating in the mechanisms).

Critical weaknesses
  • My main reservation about the paper is that the contribution is somewhat marginal. The main method used is existing Bayesian experimental design. The paper combined it with LLM-generated behavioral hypotheses. This defines the set being considered by the Bayesian experimental design, but arguably is not a substantial methodological improvement.

  • There is a gap between what this paper’s methods try to optimize and the true objective of a mechanism designer. For example, MUFS maximizes the expected information gained about which mechanism parameter is optimal according to the current behavioral hypothesis. But the mechanism designer cares about maximizing posterior expected revenue. The latter depends on the revenue magnitude under each behavioral model. So, it seems that the proposed methods may not necessarily optimizing expected revenue.

Questions for authors

Would it be possible to compare your methods with an alternative method that uses the same behavioral model pool but selects the experiment that is expected to produce the largest improvement in expected revenue?

2 comments
AuthorsOpenReview ↗

Authors' Answer to Question

Q. Would it be possible to compare your methods with an alternative method that uses the same behavioral model pool but selects the experiment that is expected to produce the largest improvement in expected revenue?

We thank the reviewer for this insightful question! As suggested, we implement Seq-ER, which uses the same behavioral model pool but selects the next experiment that is expected to produce the largest improvement in expected revenue [1,2]. Thus, Seq-ER provides a direct comparison to MUFS, as both focus on the task of mechanism optimization rather than model-parameter identification. For conciseness, Tables 1 and 2 report MUFS, Seq-ER and Seq-Random over 300 first-price auctions and 300 pay-per-bid auctions, respectively. For the displayed methods, the best value is shown in bold, and the second-best is shown in italics.

Table 1

MUFSSeq-ERSeq-Random
DeepSeek-V3.1Mean ↑1072.51051.31043.3
DeepSeek-V3.1SD ↓44.2197.798.9
GPT-5.1Mean ↑1084.21060.01042.6
GPT-5.1SD ↓70.2198.6112.1

Table 2

MUFSSeq-ERSeq-Random
DeepSeek-V3.1Mean ↑70.967.662.9
DeepSeek-V3.1SD ↓20.928.426.4
GPT-5.1Mean ↑108.893.549.5
GPT-5.1SD ↓21.539.333.5

Tables 1 and 2 show that MUFS consistently achieves the highest mean revenue and lowest SD across all settings. Although Seq-ER achieves high mean revenue, its SD is consistently even higher than that of Seq-Random. One possible explanation for Seq-ER's large SD is that it prioritizes potential revenue improvements without explicitly penalizing revenue variability. For example, in first-price auctions, higher reserve prices may increase revenue but also raise the risk of no-sale outcomes, potentially leading to greater revenue variability.

We have added these results and the related discussion to Section 5 in the revised paper.

[1] Frazier, Peter I. "Bayesian optimization." Recent Advances in Optimization and Modeling of Contemporary Problems. INFORMS, 2018.

[2] Huan, Xun, Jayanth Jagalur, and Youssef Marzouk. "Optimal experimental design: formulations and computations." Acta Numerica, 2024.

AuthorsOpenReview ↗

Authors' Reply to Weakness 1 and Weakness 2

W1. My main reservation about the paper is that the contribution is somewhat marginal. The main method used is existing Bayesian experimental design. The paper combined it with LLM-generated behavioral hypotheses. This defines the set being considered by the Bayesian experimental design, but arguably is not a substantial methodological improvement.

We thank the reviewer for this comment! Our primary contribution is MUFS, a new sequential experimental design strategy for mechanism optimization under unknown strategic behavior and limited experimental budgets. We clarify the contributions of our work below.

Problem-level: Our work addresses an important but underexplored problem: optimizing mechanism parameters under unknown strategic behavior and limited experimental budgets. Existing Bayesian experimental design methods [3-5] cannot directly address this setting. More broadly, our framework is applicable to a wide range of mechanism design problems in which designers can interact with agents by varying mechanism parameters and use the resulting outcomes to optimize the mechanism.

Method-level: BUFS and MUFS are two complementary instantiations of our proposed framework. While BUFS largely follows standard Bayesian experimental design, our primary technical contribution lies in MUFS. Specifically, once LLMs convert open-ended behavioral uncertainty into a finite pool of behavioral hypotheses, BUFS focuses on distinguishing among these behavioral hypotheses, while MUFS directly targets the optimal mechanism parameter. Since directly evaluating MUFS over the pool is computationally intractable, we introduce a surrogate space that substantially simplifies the computation while theoretically preserving the ranking of candidate experiments under suitable conditions (Proposition 1).

We have added this discussion to Section 1, Section 2 and Section 6 in the revised paper.

[3] Zhong, Shijie, et al. "Goal-oriented bayesian optimal experimental design for nonlinear models using markov chain monte carlo." SIAM/ASA Journal on Uncertainty Quantification 14.1 (2026): 19-47.

[4] Huang, Daolang, et al. "Amortized bayesian experimental design for decision-making." Advances in Neural Information Processing Systems (2024).

[5] Zhang, Zheyu, et al. "Goal-oriented sequential bayesian experimental design for causal learning." arXiv preprint arXiv:2507.07359 (2025).

W2. There is a gap between what this paper’s methods try to optimize and the true objective of a mechanism designer. For example, MUFS maximizes the expected information gained about which mechanism parameter is optimal according to the current behavioral hypothesis. But the mechanism designer cares about maximizing posterior expected revenue. The latter depends on the revenue magnitude under each behavioral model. So, it seems that the proposed methods may not necessarily optimize expected revenue.

We thank the reviewer for this insightful comment! Please refer to Answer to Question for the suggested method and its comparison with our methods. Here, we further explain the rationale for MUFS's learning objective and clarify its distinction from Seq-ER.

Regarding rationale for targeting $z^ $:* Since uncertainty about $z^* $ stems from behavioral uncertainty, we start with a pool of LLM-generated behavioral models to transform the open-ended behavioral uncertainty into a finite hypothesis space. MUFS maintains a belief over the pool and maximizes information gain about induced $z^* $, enabling exploration of designs with lower posterior expected revenue but higher information gain as it helps identify $z^* $ in the end. More importantly, the mechanism is finally optimized by maximizing expected performance under the posterior over the pool, ensuring that the information collected during experiments ultimately serves the goal of maximizing expected objective value.

Regarding the difference between Seq-ER and MUFS: Seq-ER tends to favor experiments that are expected to produce the largest improvement in expected revenue. However, an experiment that yields high posterior revenue may reveal little about which mechanism parameter is truly optimal. In contrast, MUFS explicitly targets uncertainty about the optimal mechanism parameter, prioritizing experiments that can help distinguish between competing mechanism choices. This distinction is particularly important under a limited experimental budget, when only a few experiments can be conducted. In this setting, prioritizing experiments based on the revenue improvement risks leaving key uncertainties about the optimal mechanism unresolved before the budget is exhausted.

We have added these results and the related discussion to Section 5 in the revised paper.

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

Researchers face a problem when designing auctions: they do not know exactly how bidders will act, and each test auction is expensive.

This paper offers a way to choose good auction settings (such as reserve prices or bid fees) with only a small number of tests.

First, a large language model suggests a short list of possible bidder behavior models. The true model does not have to be on the list (because the method only needs a useful set of approximations that help it identify good auction settings, not a perfect model of bidder behavior). Next, the researchers run a sequence of chosen test auctions.

Two strategies guide the choice of which auctions to run when: one first learns which behavior model is most likely, then picks the best settings (BUFS); the other tries to learn the best settings more directly by discretizing the space of possible auction parameters (MUFS).

In computer experiments with simulated bidders, both strategies found auction settings that earned more money than random testing, standard learning methods, or other theory-based choices.

Critical strengths
  • The overall problem (optimizing auction parameters under unknown strategic behavior with limited costly experiments) is well-motivated. Overall, it is a solid, focused contribution at the intersection of mechanism design and sequential experimental design.

  • The LLM-generated finite model pool (explicitly allowing misspecification) is a pragmatic and relatively new way to turn open-ended behavioral uncertainty into a tractable hypothesis space. That said, I do not know the mechanism design literature that well to evaluate this confidently.

  • The MUFS method, which is motivated by arXiv:2507.07359, is interesting because it focuses on learning the best auction settings directly, instead of first figuring out how bidders behave. It uses a simple grid of nearby settings to keep the calculations fast while still ranking the options correctly.

Critical weaknesses
  • Approach does not introduce a new foundational technique or solve a broad open problem, but rather uses the MUFS objective + surrogate theory + LLM-pools to produce a usable end-to-end pipeline for mechanism-parameter optimization under limited experimental budgets.

  • BUFS is essentially standard sequential Bayesian optimal experimental design, followed by a posterior-conditioned optimization step.

  • The empirical gains, while consistent, are potentially modest in absolute revenue and confined to two synthetic/LLM-agent environments. The experimental settings are fully simulation-based. The approach is not shown to work with real human bidders instead of LLM agents.

  • Both the candidate behavioral models and the strategic agents are LLM-derived. It is unclear how, or whether, the model-intelligence ratio between the strategic agent and behavioral model matters. In general, more could be done to characterize the inter-LLM dynamics within the method and its evaluation.

  • As this is a strategic environment, more could be done to explore adversarial dynamics herein. For example, because the designer’s next experiment is chosen publicly or possibly can be inferred, bidders who understand the EIG criterion can shade their actions during the learning phase to inflate or deflate the apparent value of certain parameter regions, steering the final z-hat in their favor. Generally, what about adversarial or non-stationary bidders?

Questions for authors
  • In Appendix B.3, we vary the size of the LLM-generated candidate behavioral-model pool, but only up to size 3. Is this due to cost considerations (i.e., more expensive to run the approach past that point? Or would additional candidates harm the approach here?
6 comments
AuthorsOpenReview ↗

Authors' Answer to Question

Q. In Appendix B.3, we vary the size of the LLM-generated candidate behavioral-model pool, but only up to size 3. Is this due to cost considerations, i.e., more expensive to run the approach past that point? Or would additional candidates harm the approach here?

We thank the reviewer for this question. Limiting the pool size to 3 in Appendix B.3 was due to evaluation cost. To further address this concern, we have extended the tested pool size to 4, and the extended results in Table 1 show that performance remains strong from size 2 to 3 to 4.

Specifically, we now consider 5 LLM-generated misspecified models. To reduce the influence of any particular pool composition, we exhaustively evaluated all (5n)\binom{5}{n} possible combinations for each tested pool size nn, with 10 independent runs per combination. Thus, increasing the number of candidate models will combinatorially increase the cost of comprehensive evaluation. The reported results in Table 1 are averaged over (52)=10\binom{5}{2}=10 size-two pools, (53)=10\binom{5}{3}=10 size-three pools, and (54)=5\binom{5}{4}=5 size-four pools across 10 independent runs. We do not report size 5 because (55)=1\binom{5}{5}=1, meaning there is only a single possible pool composition at that size. Therefore, it is hard to disentangle the effect of that particular composition from the general effect of pool size. The best value in Table 1 is shown in bold, and the second-best is shown in italics.

Table 1

Pool SizeMethodMean ↑SD ↓Mean Regret ↓Exceedance Rate ↑
2BUFS23.814.965.4992.38%
2MUFS24.424.914.8894.23%
2Seq-Random22.894.666.4188.69%
2Seq-HC23.584.795.7291.58%
2Seq-TS23.865.085.4492.57%
3BUFS24.554.784.7594.63%
3MUFS25.154.534.1595.93%
3Seq-Random22.434.436.8786.21%
3Seq-HC22.484.396.8286.59%
3Seq-TS24.094.895.2193.36%
4BUFS26.213.893.0997.85%
4MUFS26.813.302.4998.68%
4Seq-Random21.994.207.3183.67%
4Seq-HC22.524.186.7886.85%
4Seq-TS23.994.435.3193.00%
-Random Sampling21.932.737.3783.20%
-Heuristic Search20.560.978.7465.33%
-Coarse Search21.561.567.7480.25%
-LinTS25.082.514.2295.83%
-EpochGreedy21.822.447.4882.43%
-SPNE18.761.3610.5422.19%

Table 1 shows that MUFS and BUFS remain effective as the pool size increases from 2 to 4, with both methods consistently improving in mean revenue and regret. MUFS achieves increasingly strong performance as the pool size grows and outperforms all sequential design baselines in terms of mean revenue and mean regret for pool sizes 3 and 4.

We have added these results to Section 5.3 and Appendix B.3 in the revised paper.

AuthorsOpenReview ↗

Authors' Reply to Weakness 1 and Weakness 2

W1. Approach does not introduce a new foundational technique or solve a broad open problem, but rather uses the MUFS objective + surrogate theory + LLM-pools to produce a usable end-to-end pipeline for mechanism-parameter optimization under limited experimental budgets.

We thank the reviewer for this comment! Our work addresses a broad class of mechanism design problems involving mechanism parameter optimization under unknown strategic behavior and limited experimental budgets, and our primary technical contribution lies in MUFS. We clarify the contributions of our work below.

Problem-level: Our work addresses an important but underexplored problem: mechanism parameter optimization under unknown strategic behavior and limited experimental budgets. Existing Bayesian experimental design methods [1-3] cannot directly accommodate this setting. Specifically, [1] relies on a specified agent response model, while [2,3] require sufficient offline design-outcome trajectories. In contrast, our framework operates with limited sequential behavioral observations and does not require the true behavioral model to be included in the candidate pool. More broadly, our framework is applicable to a wide range of mechanism design problems in which designers can interact with agents by varying mechanism parameters and use the resulting outcomes to optimize the mechanism.

Method-level: Beyond the problem setting, our primary technical contribution lies in MUFS. Given a pool of candidate model-parameter pairs, conventional Bayesian experimental design methods focus on distinguishing among them, whereas MUFS directly targets uncertainty about the optimal mechanism parameter. This allows MUFS to prioritize behavioral information that is relevant to the final mechanism optimization, rather than spending the limited experimental budget on behavioral distinctions that may not affect the optimal mechanism. Furthermore, directly evaluating MUFS requires repeatedly solving mechanism optimization problems over numerous model-parameter pairs. We therefore introduce a surrogate-space approximation that substantially reduces computational cost while theoretically preserving the ranking of candidate experiments under suitable conditions (Proposition 1).

We have added this discussion to Section 1, Section 2 and Section 6 in the revised paper.

[1] Zhong, Shijie, et al. "Goal-oriented bayesian optimal experimental design for nonlinear models using markov chain monte carlo." SIAM/ASA Journal on Uncertainty Quantification 14.1 (2026): 19-47.

[2] Huang, Daolang, et al. "Amortized bayesian experimental design for decision-making." Advances in Neural Information Processing Systems (2024).

[3] Zhang, Zheyu, et al. "Goal-oriented sequential bayesian experimental design for causal learning." arXiv preprint arXiv:2507.07359 (2025).

W2. BUFS is essentially standard sequential Bayesian optimal experimental design, followed by a posterior-conditioned optimization step.

We thank the reviewer for this comment! We acknowledge that BUFS follows standard sequential Bayesian optimal experimental design. However, BUFS is one of two complementary instantiations of our proposed framework: once LLMs convert open-ended behavioral uncertainty into a finite pool of behavioral hypotheses, BUFS focuses on distinguishing among them, while MUFS directly targets the optimal mechanism parameter. Our primary technical contribution lies in the other instantiation, MUFS.

We have added this discussion to Section 1 in the revised paper.

AuthorsOpenReview ↗

Authors' Reply to Weakness 3 and Weakness 4

W3. The empirical gains, while consistent, are potentially modest in absolute revenue and confined to two synthetic/LLM-agent environments. The experimental settings are fully simulation-based. The approach is not shown to work with real human bidders instead of LLM agents.

We thank the reviewer for this thoughtful comment!

Regarding evaluation beyond LLM-based agents: To address this concern, we conduct additional experiments where we replace LLM-based agents with agents instantiated from a human-data-calibrated behavioral model. Following [4], we fit this behavioral model to real human behavioral data and treat it as the ground-truth agent interacting with different experimental design methods. Since this model is calibrated on real human bidding data, it captures genuine human behavioral patterns and serves as a principled proxy for human bidder behavior. The results are reported in Table 1 in our Answer to Question. Notably, even when the LLM-generated candidate models are misspecified to human behavioral patterns to varying degrees, MUFS and BUFS still achieve strong performance, confirming that our framework generalizes beyond LLM-based agents.

Regarding the magnitude of the revenue gains: Absolute improvements depend on the revenue scale of each auction setting. Across the four settings with LLM agents, MUFS achieves mean revenue improvements of up to 16.4% over the strongest non-BUFS/MUFS baseline, while consistently exhibiting lower standard deviation.

We have added these results to Section 5.3 and Appendix B.3 in the revised paper.

[4] Brunner, Tobias, et al. "Prospect theory in a dynamic game: Theory and evidence from online pay-per-bid auctions." Journal of Economic Behavior & Organization 164 (2019): 215–234.

W4. Both the candidate behavioral models and the strategic agents are LLM-derived. It is unclear how, or whether, the model-intelligence ratio between the strategic agent and behavioral model matters. In general, more could be done to characterize the inter-LLM dynamics within the method and its evaluation.

We thank the reviewer for this insightful comment! We agree that LLM capability affects both the quality of the generated candidate behavioral models and the strategic decision-making of LLM-based agents. To examine the effect of relative LLM capabilities, we use publicly reported GPQA Diamond scores [5] to qualitatively characterize models as "stronger" or "weaker," and test two contrasting configurations of strategic agents and candidate model generators:

  • WASG (Weaker Agent + Stronger Generators): A relatively weaker strategic agent (DeepSeek-V3.1) paired with stronger candidate model generators (GPT-5.2 Thinking and Claude Opus 4.5).
  • SAWG (Stronger Agent + Weaker Generator): A relatively stronger strategic agent (GPT-5.1) paired with a weaker candidate model generator (ChatGPT-o4-mini).

The results reported in Table 2 are averaged over 300 first-price auctions and 300 pay-per-bid auctions, respectively. For the displayed methods, entries mean revenue ± standard deviation. The best value is shown in bold, and the second-best is shown in italics.

Table 2

Config.AuctionMUFSBUFSSeq-ER
WASGFirst-price1072.5 ± 44.21050.4 ± 48.61051.3 ± 197.7
SAWGPay-per-bid108.8 ± 21.5100.6 ± 24.293.5 ± 39.3

The results show that MUFS achieves the highest mean revenue in both configurations. BUFS remains competitive with Seq-ER, achieving comparable mean revenue with substantially lower standard deviations in WASG and outperforming Seq-ER in both mean revenue and stability in SAWG. These results provide empirical evidence that our methods maintain strong performance under contrasting capability configurations between strategic agents and candidate model generators. We interpret this analysis as empirical evidence of robustness under different inter-LLM capability configurations.

We have added these results to Section 5.3 and Appendix B.5 in the revised paper.

[5] Rein, David, et al. "GPQA: A graduate-level google-proof q&a benchmark." Conference on Language Modeling (2024).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.