acceptodds
Under review as a conference paper at ICLR 2027

Healthcare Mechanisms from Policy-as-Code Search under Strategic Provider Response

Abstract

How can AI help design effective hospital incentive and resource-allocation policies when providers respond strategically? Yet conventional research that evaluates these policies against fixed provider programs or a single performance measure can miss distortions that shift across behaviors or emerge through further adaptation. We introduce , a closed-loop multi-agent simulator that integrates hospital operations with five dimensions of strategic provider response: billing classification (coding), patient acceptance, treatment deferral, treatment effort, and bed requests. Administrator and provider policies are typed, inspectable programs, which an LLM edits using rollout feedback. We first map behavior under fixed response rules, then test whether gains from joint program design survive unilateral adaptation and responses to retained policy histories. Stronger audits suppress coding inflation while worsening access for complex patients; contract changes can restore effort while coding distortion persists; and further provider search can erode recovered health. We characterize finite-policy convergence under exact response oracles and separate search and payoff-estimation errors in an approximate stopping bound. These results identify pressure migration across response channels and adaptation stages, suggesting that AI-designed mechanisms should be evaluated through the behaviors they induce and the further adaptations they invite.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI for Healthcare, Pressure Migration, LLM, Code-as-Policy

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_T9jHOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper develops a hospital operations simulator to study how healthcare providers respond to changes in payment, performance incentives, and resource allocation. The paper integrates coding, patient acceptance, treatment deferral, treatment effort, and bed requests into a shared environment to examine how distortions shift across channels. It also represents administrator and provider policies as executable code that LLMs can modify, while retaining historical policies to test resilience to continued adaptation. Its main contribution is a unified experimental environment and evaluation approach. The simulator follows an “Identify–Produce–Settle” cycle. Administrators set financial and quality incentives, capacity, and bonuses. Providers choose actions based on returns, patient complexity, and workload. Between simulation episodes, an LLM modifies policy code. The framework first jointly searches for a high-performing policy pair, then allows each party to pursue its own interests, and finally evaluates interactions with historical policies to identify remaining profitable deviations and changes in health and access.

Critical strengths
  • This paper introduces a hospital operations simulator, which is a quite innovative and interesting study. The paper evaluates policies under continued provider adaptation and reveals cross-channel effects, such as improved billing accuracy accompanied by reduced patient access.
  • Provides interpretable policies and a comprehensive analysis. Executable policies make behavioral changes inspectable, while experiments jointly examine returns, patient health, and strategic stability rather than relying solely on aggregate rewards.
Critical weaknesses
  • The simulator’s real-world validity is insufficiently established. Patient arrivals, provider incentives, treatment costs, and health outcomes are prescribed rather than calibrated against hospital data, and the paper reports no structured independent expert validation. Literature consistency checks cannot establish that the combined environment or discovered strategies are realistic. Consequently, the findings demonstrate behavior within the specified simulator, but their relevance to actual hospital policy remains unsubstantiated. Experimental sugessions for the authors: Calibrate key components using hospital data and validate predictions against independent observations. If such data are unavailable, at minimum obtain structured assessments from (expert) clinicians, or hospital administrators, including blinded reviews of generated strategies. Revise implausible assumptions (particularly payment despite zero treatment effort).
  • The methodological contribution and added value of LLM-based program search are not sufficiently isolated. The framework primarily integrates existing evolutionary search and game-theoretic tools. Current comparisons do not clearly establish whether gains arise from meaningful LLM-generated changes to policy logic, additional search opportunities, or the underlying optimization machinery. Authors should compare against non-LLM evolutionary search, random program mutations, and parameter-only optimization under matched simulation budgets. Ablate the LLM editing component while retaining the remaining framework, conduct multiple independent searches, and report policy outcomes alongside runtime and token costs. These controls are necessary to demonstrate what the LLM contributes beyond the integration of existing methods.
Questions for authors
  • Although the paper reports rejection and waiting alongside treated-patient health, could the authors evaluate health benefits across a fixed cohort of all arriving patients, explicitly accounting for rejected patients and those still waiting at the evaluation horizon? Do the main policy rankings remain unchanged under this assessment?
  • The paper shows that soft penalties and reduced profit incentives do not eliminate all harmful responses. Could the authors test clinically and operationally grounded safeguards (such as minimum-service requirements, payments linked to delivered care, or access constraints) and assess whether they continue to protect patients under further provider adaptation?
5 comments
AuthorsOpenReview ↗

Response to Reviewer T9jH (Part 1/4)

We thank the reviewer for the positive assessment and the concrete suggestions, which guided the additional evaluations below.

W1. Real-world validity: public-data alignment, external comparisons and LLM mechanism assessment. Reviewer CvHR shares this concern; the completed comparisons are reported in CvHR (W1, parts a–d). Code, records and verification scripts for all additional evaluations are in the anonymous repository under rebuttal/. In brief:

  • (a) Public-data alignment. We aligned service load and relative DRG weights using 2024 Chinese public-hospital statistics and Yantai's DRG payment schedule. A separate SPARCS pilot uses 54,865 adult Medicaid discharges to estimate and check clinical input distributions, with severity and admission urgency separated from payment weights. Both scenarios preserve mean truthful payment.
  • (b) External comparisons. We compare the submitted simulator with evidence on upcoding prevalence under audit (12.91%; Zhou et al., Soc. Sci. Med. 2025) and coding responses to financial incentives (Jürges & Köberlein, J. Health Econ. 2015).
  • (c) Reassessment under the aligned parameters. Both scenarios retain profitable fixed-contract responses and health declines with provider utility gains in 5/5 E0 → ABR-A comparisons. S2 keeps health near its starting level while up-coding remains at 76% (China) and 94% (SPARCS). The audit-induced increase in the rejection gap appears in the SPARCS scenario but not the Chinese one.
  • (d) LLM mechanism assessment. Five models provided 100 responses. They agreed that initial performance does not establish performance after adaptation and that provider returns, hospital scores and patient care can diverge; on whether audits shift behaviour to patient selection, 30 expected an increase and 54 judged the information insufficient. These LLM-based conceptual checks supplement the empirical analyses and are not presented as independent clinical expert validation; details are in CvHR (W1d).

W1 (continued). Implausible assumption: payment despite zero treatment effort. We agree that this settlement rule should be revised; our S2 experiment below tests payment linked to delivered care.

  • Where it matters. Under the submitted rule, a completed coded case is paid regardless of delivered effort, and zero effort yields zero health. This boundary is most visible in revised-prompt DeepSeek 200×3, where 99.94% of the doctor marginal is on effort = 0.0 and H < 0.001 (App. L.6). Sec. 4.3 and App. M identify this as a boundary case of the settlement rule.
  • Erosion also occurs at positive effort. In the reference paths, the ABR-A endpoints keep effort at 0.257 (DeepSeek) and 0.172 (GLM), while health falls from 0.984 to 0.791 and from 0.965 to 0.631. The matched-contract and terminal-response checks (App. L.8) also use positive-effort programs. In the submitted L2 audit sweep, mean effort remains about 1.7–2.4.
  • Revision. We have implemented safeguard S2, which pays only for cases reaching a minimum-service level, and tested further adaptation under a fixed contract (Q2 below). We will apply the revised settlement rule to the zero-effort endpoint and report the updated result in the revision.
AuthorsOpenReview ↗

Response to Reviewer T9jH (Part 2/4)

W2. The contribution of LLM-based program search is not isolated from the evolutionary and game-theoretic machinery. Compare with non-LLM evolutionary search, random program mutations and parameter-only optimization under matched budgets; ablate the LLM editor; run multiple independent searches; report runtime and token costs.

We add a direct-generation baseline (B0) against the paper's E0 joint-design result. B0 uses DeepSeek to generate 200 complete administrator–provider programs independently from the task description and DSL schema, without a supplied program or rollout feedback. Following the paper's selection rule, the first valid sample and the five highest-scoring training candidates are compared on validation seeds. We then assess the selected program and the archived E0 program on the same 20 held-out seeds.

MethodInitialization and searchHeld-out UAU_A (submitted parameters)
Paper E0 (DeepSeek, original prompt, high thinking)17 supplied programs + 200 feedback-guided proposals15.45
B0, first sampleOne independently generated program5.68
B0, best of 200200 independent generations, followed by validation selection14.74

The paper's E0 exceeds B0 by 0.71 in held-out UAU_A. Under the Chinese parameters, the same frozen E0 and B0 programs score 15.36 and 14.13, respectively; in the SPARCS scenario, on 20 fresh common seeds, they score 14.34 and 12.70. This is a complete-pipeline comparison: E0 uses the supplied starts, while B0 uses the current DeepSeek endpoint following a provider update. The paper also reports three independent response searches per task and ABR versus CSRO–MGCCE comparisons to examine search variability and history retention (App. L.8–L.9).

Cost. B0 used 200 calls, all successful on the first attempt, totaling 1.07M prompt and 2.04M completion tokens (API-reported). Re-assessing the archived E0 program required no model calls. All 200 generations with token usage, the selection and the held-out assessment are in rebuttal/direct_generation.

AuthorsOpenReview ↗

Response to Reviewer T9jH (Part 3/4)

Q1. Evaluate health over a fixed cohort of all arriving patients, counting rejected patients and those still waiting at the horizon. Do the main rankings survive?

We agree that the treated-patient mean HH conditions on access, so a rule can raise it by treating fewer or easier patients. Using the submitted parameters, we re-assessed every endpoint on the cohort of all arrivals. Each arriving patient is followed to exactly one state at the horizon T=200T=200: treated, rejected, left the queue untreated, or still waiting; the tracker verifies that these four counts sum to arrivals in every rollout. Cohort health is Hcoh(h0)=1Narr[∑i∈treatedhi+h0 (Nrej+Nleft+Nwait)],H_{\rm coh}(h_0)=\frac{1}{N_{\rm arr}}\Big[\sum_{i\in{\rm treated}}h_i+h_0\ (N_{\rm rej}+N_{\rm left}+N_{\rm wait})\Big], where h0h_0 is the health credited to an arrival who received no treatment. Our headline uses h0=0h_0=0 (health actually delivered per arrival); we then vary h0h_0.

Editor configurationEndpointTreated HHTreatedRejectedLeft untreatedWaiting at TTHcoh(0)H_{\rm coh}(0)Urgent Hcoh(0)H_{\rm coh}(0)
DeepSeek, original, offE00.94662.7%4.1%32.5%0.73%0.5930.595
ABR-A endpoint0.67766.4%0.0%32.9%0.74%0.4490.173
DeepSeek, original, highE00.98497.3%0.0%2.5%0.26%0.9570.946
ABR-A endpoint0.79199.6%0.0%0.3%0.10%0.7870.664
DeepSeek, revised, offE00.99561.2%35.4%3.1%0.25%0.6090.963
ABR-A endpoint0.81663.4%17.8%18.2%0.55%0.5170.014
DeepSeek, revised, highE01.00999.1%0.5%0.4%0.00%1.0000.920
ABR-A endpoint0.14899.7%0.0%0.3%0.00%0.1470.122
GLM-5.3E00.96588.7%0.0%10.8%0.46%0.8570.935
ABR-A endpoint0.63199.0%0.1%0.7%0.27%0.6250.534

Counts and cohort health are pooled over the paper's 20 held-out environment seeds (≈48,000 arrivals per endpoint); treated H uses the paper's original aggregation. Every replay reproduces its released record, with no model calls. The full table with ABR-D and both MGCCE endpoints, the replay records and the scripts are in rebuttal/fixed_cohort.

  • The main erosion result holds. Erosion of jointly designed pair under provider adaptation (E0 → ABR-A endpoint) holds on full cohort in 5/5 configurations, including GLM. Across five main endpoints of each configuration (E0, ABR-A, ABR-D, MGCCE 200×3, MGCCE 50×12), 47 of 50 pairwise orderings agree between treated HH and Hcoh(0)H_{\rm coh}(0), and none of the reversals involves the E0 → ABR-A comparison. Two are E0 versus ABR-D endpoint, where treated HH differs by less than 0.01. The third is E0 versus MGCCE 200×3 in original/off, where the mixture treats more arrivals (66.2% vs 62.7%).
  • The cohort view strengthens the harm finding for urgent patients. Health delivered per urgent arrival falls from 0.595 to 0.173 and from 0.963 to 0.014 in the two thinking-off configurations, a loss that the treated mean (0.946 → 0.677; 0.995 → 0.816) understates.
  • Robust to the credit given to untreated patients. For h0∈{0.25,0.5,0.75}h_0\in\lbrace 0.25, 0.5, 0.75\rbrace, E0 remains above the ABR-A endpoint in 5/5 configurations, and at most 3 of 50 within-configuration orderings change relative to h0=0h_0=0.
  • Few patients remain waiting. Patients still waiting at the horizon account for at most 0.78% of arrivals at any endpoint.
  • The cohort measure reveals differences in access. For two E0 pairs, treated HH is 0.946 and 0.995, while Hcoh(0)H_{\rm coh}(0) is 0.593 and 0.609. Across configurations, 48 of 300 pairwise orderings reverse. We will make HcohH_{\rm coh} the primary health measure in the revision, with treated HH as a secondary column.

On the paper's five selection-validation seeds of the four DeepSeek configurations, the conclusions are the same: erosion holds in 4/4, and 32 of 36 orderings agree.

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

This paper develops a simulator of healthcare provider behavior (a "closed-loop multi-agent simulator") and use this simulator to explore how policy interventions or changes in incentives impact simulated hospital behavior.

Critical strengths

It is not entirely clear to me how realistic the simulator is, but I appreciate that behavior is defined in Appendix C in part using models of utility functions, treatment production and cost, etc. It is nice to see a simulator that isn't entirely based on LLMs acting out behavior.

The paper appears to pull from a variety of work in economics / health policy to inform the simulator design...see Appendix C.1.

Table C.1 connects each component to its economic or operational motivation and records the specific abstraction used in MEDI-SIM. The cited work motivates the modeled relationships; the functional forms and parameter settings define the controlled environment evaluated here.

Critical weaknesses

First of all, there is a substantial lack of clarity in the exposition, and a related lack of clarity in motivation/significance of the work.

Related to that lack of clarity is the concern that I do not understand the purpose of the exercise undertaken in this paper. What is the purpose of this simulator? I see a few possibilities, from least to most compelling:

  1. To demonstrate the generic insight that policies can have unintended consequences under rational behavior from strategic actors? That's not quite a novel observation.
  2. To demonstrate that this health-specific simulator can recover existing knowledge of trade-offs in incentives in health policy? I'm not sure this paper achieves that goal, given that most existing knowledge they cite is already "baked into" the simulator.
  3. To provide a real tool to anticipate the effect of real health policies? I do not believe the authors would argue that this simulator is anything more than a proof of concept.

The challenge I see is that the simulator is both (a) quite fiendishly complicated, but (b) unlikely to be realistic. Insofar as the simulation is designed to inform actual policy, this paper seems more appropriate for a health policy or economics journal.

I could see value in trying to demonstrate that the simulator gives us predictive "out-of-sample" insights, e.g., by developing the simulation to reflect a few known things, then showing it predicts other known consequences of policies. But where realism is discussed in Section 4, there is the caveat

Literature consistency and setup. Nine comparisons check the intended incentive and operational mechanisms (Appendix Table I.1). Because the same literature informed the design, agreement in direction serves as a behavioral sanity check rather than independent validation. The table distinguishes literature-supported mechanisms from simulator findings such as cross-channel substitution and adverse KPI-steered allocation

Second, I have concerns as to whether or not the authors have a sufficiently robust understanding of how healthcare systems operate. For instance, consider the first sentence of the introduction:

Consider a hospital administrator trying to improve care with a limited budget. Paying more for complex cases may encourage their treatment, but it also makes reporting the complex diagnosis more profitable.

This sentence seems to conflate the role of hospital administrators (who manage hospitals that provide care) with the role of those who set policies for payment (e.g, private insurance companies, Medicare, NHS). Moreover, this paper has the annoying tendency (which I have seen in many other papers) of jumping to the conclusion that their approach is a natural one:

A natural approach is to optimize hospital rules in simulation and select the highest-return candidate

I don't think this is a natural approach at all! Nor is there any evidence presented that this is actually done to inform policy! Certainly one could imagine micro-economic analysis that tries to game out the impact of specific changes to derive some insights, but the general "we will simulate a complex system in-silico and use that to select the numerically highest-scoring policy" does not seem like an accurate description of how policy-making is done. Nor is it obvious how to even measure the "highest return" in a multi-stakeholder setting like healthcare? Also, as discussed above, I do not believe many of these rules are "hospital rules" in reality.

There are other concerns with clarity throughout (e.g., in Section 4.1, the paper discusses "sweep[ing] financial sensitivity α\alpha and quality sensitivity β\beta" but those terms have not been defined yet), but the above are the most critical weaknesses in my view.

Questions for authors

I can think of no questions at this time whose answers could materially affect my assessment.

5 comments
AuthorsOpenReview ↗

Response to Reviewer fhcw (Part 1/3)

We thank the reviewer for the detailed reading and for appreciating that Medi-Sim's behavior is grounded in economic models rather than LLM role-play. Several central concerns trace back to two things the submission left implicit: its purpose and its institutional setting. We clarify both below and will state them explicitly in the revision.

1. What is the purpose of the exercise? Our goal is to use LLM-based program search as a heuristic method for designing healthcare mechanisms and evaluating how providers adapt to them. LLM program search (FunSearch, AlphaEvolve, OpenEvolve) proposes executable rules from rollout feedback. Medi-Sim pairs joint design with provider-response search and tracks administrator scores, provider utilities and patient care separately. Under the submitted parameters, E0 improves over the best supplied start in validation U_A in all five configurations; both the administrator score and health then fall from E0 to the ABR-A endpoint in all five.

Answering this requires three things that existing environments do not combine: (i) several coupled response channels, so that distortion can move rather than vanish; (ii) policies represented as inspectable programs, so that we can see how a rule earns its return and which channel a response exploits; and (iii) a protocol from joint design to unilateral adaptation to responses against retained histories, with a convergence analysis (App. N).

In this sense, Medi-Sim offers a feasible template for LLMs to participate in social simulation: as heuristic policy searchers whose outputs are executable, and checkable, rather than as free-form role-players. We consider this a natural fit for ICLR's primary area of alignment, safety and societal considerations. The paper does not prescribe health policy, and we state explicitly that the policies are not administrative guidance (Ethics statement).

2. "Unintended consequences under rational behavior is not novel." We agree that the generic statement is not novel, and the paper does not claim it. Our findings are specific, quantified and, we believe, new for this setting:

  • Distortion migrates across channels. In the submitted setting, audits that cut up-coding by >99% (balanced regime: 0.636 → 0.001) simultaneously raise cherry-picking of easier cases in the same regime (0.100 → 0.233). This coupled effect is not encoded in any single response rule (App. I).
  • Distortion migrates across adaptation stages. In all five configurations under the submitted parameters, provider utility rises while health falls from E0 to the ABR-A endpoint. The exploited channel differs by configuration, coding inflation in some and pure effort reduction with 0% up-coding in others (Sec. 4.3, App. L.6).
  • Repair on one channel leaves another open. Removing the profit weight restores effort in 3/3 independent searches, yet up-coding stays at ~86%. Code inspection shows that the recovery comes from a rewritten effort rule (min(1.5 * base_effort, stamina_ratio)) and full acceptance, not from any change that closes the coding channel (App. L.8).
  • Stability is not care quality. MGCCE mixtures with small deviation gains (≤ 0.132) can still deliver low health (0.449), and one mixture concentrates 99.94% of the doctor marginal on effort = 0.0 (Sec. 4.3, App. L.9).

The contribution is to follow mechanism design and subsequent exploitation within one program-search framework, tracing both to executable code.

AuthorsOpenReview ↗

Response to Reviewer fhcw (Part 2/3)

3. Independence between design and validation (knowledge "baked into" the simulator). We agree that agreement with the literature used to build a model is not independent validation, and App. I states this explicitly. To keep the two apart, we separate the sources used to build the model from the evidence we compare it with, and evaluate a frozen model on the latter only.

  • Design set (used to build Medi-Sim): the works in App. C and Table I.1 motivate coding rents (Dafny; Kuhn & Siciliani), case-mix selection (Ellis; Ma), target gaming (Bevan & Hood; Propper et al.) and multitasking (Holmström & Milgrom). Our experiments examine their joint effects through coding, audit, settlement and shared resources.
  • Frozen model: the table below uses the rules, functional forms and parameters of the submitted anonymous repository, with no tuning to the comparison evidence (recomputed in rebuttal/china_parameters).
  • Comparison evidence (not used in design, not cited in the paper): DRG evidence collected after submission, used as a post hoc consistency check on a model it played no part in building.
  • Comparison targets: the magnitude of coding under audit and its response to financial incentives.
TargetExternal evidenceSubmitted Medi-Sim
Magnitude of coding under routine audit12.91% of audited DRG records upcoded, 28 hospitals (Zhou et al., Soc. Sci. Med. 2025)12.1% (quality-driven), 29.9% (balanced), 62.3% (profit-driven) at the baseline audit level (q = 0.10)
Coding under stronger financial incentivesGerman DRG neonatology: upcoding is systematically higher at birth-weight thresholds with larger reimbursement jumps; ≥ 12,000 infants, > €100M (Jürges & Köberlein, J. Health Econ. 2015)Up-coding rises with the financial weight: 0.422 → 0.636 → 0.851 for (α, β) = (0.2, 0.8), (0.5, 0.5), (0.8, 0.2) at q = 0

Under the submitted parameters, the quality-driven regime gives a similar coding rate (12.1% vs 12.91%), while the other regimes are higher. This is a consistency comparison across different populations and measurement protocols.

Two further qualitative observations point the same way. China's DRG regulator flags both upcoding into higher-paying groups and quota-driven forced discharges and transfers; in the submitted Medi-Sim setting, raising audit q from 0 to 0.5 cuts up-coding 0.636 → 0.001 while cherry-picking rises 0.100 → 0.233. Under the US HRRP, readmissions fell (20.0% → 18.4%) while heart-failure mortality rose (7.2% → 8.6%; Gupta et al., JAMA Cardiol. 2018); in Medi-Sim, the KPI–health correlation is −0.659 in the balanced regime.

These comparisons examine outcomes of the coupled system beyond the individual rules motivating its design. In L3, LLM search constructs provider responses, and code inspection shows how they exploit a contract. Both the Chinese parameter reassessment and the SPARCS pilot in CvHR (W1) retain profitable responses that reduce health under fixed contracts. The SPARCS pilot separates clinical severity from payment weights; its audit–selection effect varies with the relative tariffs.

4. "Conflates hospital administrators with payers (insurers, Medicare, NHS)." The submission did not name its institutional setting, and its introductory examples come from the US and England, so we understand this reading. In the US and UK, the payer (Medicare, private insurers, the NHS) and the hospital's internal management are distinct actors. In the Chinese public-hospital system we model, external payment rules and the hospital's internal incentive rules are tightly coupled. The insurer pays the hospital by DRG group. Hospital management then translates these payment signals into internal department- and team-level performance pay, case-mix (CMI) targets, bed allocation, bonus pools and internal coding audits (medical-record front-page quality control). These are exactly the levers our administrator controls: α, β, capacity, bonus pool and tournament sharpness, KPI weights and audit intensity (Sec. 2.1).

NHSA's DRG/DIP 2.0 guidance explicitly addresses this transmission channel by prohibiting the use of group payment standards as staff quotas or performance-pay criteria. In our model, the administrator controls internal hospital incentives, with (α, β) representing reduced-form financial and quality weights. We will clarify this institutional setting in Sec. 1 and Sec. 2.1.

AuthorsOpenReview ↗

Response to Reviewer fhcw (Part 3/3)

5. "A natural approach is to optimize hospital rules in simulation and select the highest-return candidate. I don't think this is natural at all." The sentence refers to a starting point in automated mechanism design, as in the AI Economist and RegretNet: optimize rules in simulation and select the highest-return candidate. Our next step is to evaluate the responses that selection invites. Providers may find a more profitable response after a rule changes, or improve the targeted metric while shifting costs to another part of care. L1–L3 therefore evaluate fixed responses, unilateral adaptation and responses to retained histories, with outcomes reported channel by channel.

On whether indicator-based selection reflects practice: in China, hospital management is in fact steered by composite indicator systems. The national performance assessment of tertiary public hospitals scores hospitals on standardized indicators, including CMI, and hospitals pass such indicators down to departments. Using administrator-facing indicators to give directional guidance on candidate rules is therefore realistic in our setting. Our contribution is to show that such indicator-driven selection must be stress-tested against the responses it invites before it is trusted.

"How to measure the highest return in a multi-stakeholder setting?" We agree this is the right question. For administrator-side selection, "highest return" refers to the specified objective U_A. The paper deliberately does not collapse all stakeholders' interests into one number: the provider utility U_D and patient outcomes (treated-patient health, rejection and deferral gaps by true complexity, waiting, effort, up-coding) are evaluated and reported separately throughout. Our results show these can diverge: both institutional utilities can rise while patient health falls (App. L.6).

6. Clarity. α and β are the financial and quality incentive weights in Sec. 2.1 and Eq. (1). We will repeat this definition at the start of Sec. 4.1 and use consistent terminology.

7. Venue. The contribution is methodological: how to evaluate rules produced by LLM program search when the governed agents can also search. The health-specific environment is the testbed, not the end product. This is why the submission targets ICLR's alignment, safety and societal track rather than a health-policy journal.

We hope these clarifications address the reviewer's main concerns about purpose, institutional realism and framing. We would be grateful if the reviewer would reconsider the assessment in light of them, and we are happy to answer further questions.

Reviewer_CvHROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates healthcare mechanism design under strategic provider responses, focusing on pressure migration: reducing one distortion may redirect incentives toward another behavior, or allow distortions to reappear through further adaptation. The hospital workflow is modeled through an IPS loop, covering five response channels: coding, patient acceptance, treatment deferral, treatment effort, and resource requests. The study proceeds through three levels. Administrator and provider behavior is evaluated through prescribed weighted utility functions incorporating financial returns, health outcomes, bonuses, workload, and penalties.

Critical strengths
  • The experimental design captures a complex, multidimensional system. Modeling several response channels jointly allows the study to examine interactions and substitutions that evaluations of individual behaviors would miss. The results illustrate why improving one metric is insufficient. Stronger audits reduce coding inflation but can worsen access for complex patients. Contract revisions can restore effort while leaving coding distortion unresolved, and subsequent provider adaptation can erode recovered health.
Critical weaknesses
  • The application is very specific, but its empirical grounding is limited. The simulator uses predefined clinical, behavioral, and economic relationships and does not use any real hospital data. The findings therefore establish behavior within the constructed environment; it's hard to say their relevance to actual healthcare systems.
  • It's hard to say the quality of LLM-generated responses. Also, there is no guarantee that Budget-limited LLM search can find an optimal response or even converging to equilibrium.
Questions for authors
  • How sensitive are the conclusions to the different LLM models and prompts, or initialzations? Do independent searches consistently reveal similar response channels and policy vulnerabilities? Also, how robust are the findings to different weights, and to alternative utility functions.
  • Although health and access enter the existing objectives, the strategic actors are the administrator and providers. How is patient welfare defined independently of their utilities, and what ensures alignment? Can both institutional utilities improve while patient welfare deteriorates?
5 comments
AuthorsOpenReview ↗

Response to Reviewer CvHR (Part 1/5)

We thank the reviewer for the careful summary and for recognizing that jointly modeling several response channels exposes substitutions that single-behavior evaluations miss. We address the questions with the appendix results and additional evaluations below.

W1. Empirical grounding. We have added Chinese public-data parameter alignment, a SPARCS inpatient-data pilot, re-evaluation of the core results, external comparisons and an LLM mechanism assessment. Code, records and verification scripts are in the anonymous repository under rebuttal/.

(a) Public-data parameter alignment. We aligned the service-load scale and relative DRG weights with the NHC's 2024 health statistics bulletin and Yantai's 2024 DRG payment schedule. Public hospitals report 84.8% bed occupancy and an 8-day mean stay (NHC, Table 7). An equivalent 8-day service block at reference capacity 10 gives λarr=10×0.848×8/8=8.48\lambda_{\rm arr}=10\times0.848\times8/8=8.48. This sets a service-load proxy; mean stay is used only for this conversion. For the five macro-DRGs, we take median weights over matched payment groups, using acute myocardial infarction as the ACS proxy.

ParameterSubmittedData-aligned
Arrival load λarr\lambda_{\rm arr}128.48
DRG weights: sepsis, ACS, heart failure, joint replacement, abdominal surgery2.0, 1.6, 1.4, 1.2, 1.00.6682, 0.8495, 0.7690, 4.1433, 2.3771
Common tariff scale106.558

The tariff scale preserves mean truthful payment under the existing case frequencies. Updating the DRG weights also changes the model's CMI proxy and high/low-CMI groups through the original normalization. Case frequencies, cost coefficients, behavioural rules and incentive weights retain their submitted values. Mapping and records: rebuttal/china_parameters.

Separate clinical inputs. We also built an independent scenario from SPARCS 2024 inpatient records. Among 54,865 adult Medicaid discharges in the five clinical families, 44,094 from 139 hospitals estimate the joint distribution of case type, APR severity and admission type; 10,771 from 38 other hospitals check the input distributions. Payment weights use a semantic crosswalk to the New York tariff schedule continued in 2024, with mean truthful payment preserved. Clinical severity and urgency remain fixed when payment weights change. For example, joint replacement has higher relative payment but lower mean severity than sepsis (1.44 versus 2.90 on the 1–4 APR scale). Numerical severity and urgency encodings are checked in a sensitivity run; the cost and effort–health functions retain their original values. Inputs and records: rebuttal/sparcs_pilot.

(b) External consistency checks. We compare with two empirical findings that were not used to choose these parameters:

  • Upcoding prevalence under routine audit: 12.91% of 5,157 audited DRG records from 28 hospitals (Zhou et al., Soc. Sci. Med. 2025). Medi-Sim at the baseline audit level (q = 0.10) gives 12.1% (quality-driven), 29.9% (balanced) and 62.3% (profit-driven) as submitted, and 16.8%, 40.6% and 56.2% under the Chinese parameters.
  • Coding under stronger financial incentives: in German DRG neonatology, upcoding is systematically higher where reimbursement jumps are larger (Jürges & Köberlein, J. Health Econ. 2015). In the submitted Medi-Sim results, upcoding rises 0.422 → 0.636 → 0.851 as (α, β) changes from (0.2, 0.8) to (0.5, 0.5) to (0.8, 0.2); under the Chinese parameters it rises 0.528 → 0.553 → 0.609.

These comparisons concern the scale and direction of coding responses; the empirical samples and interventions differ from the simulator's. The Medi-Sim values are recomputed by rebuttal/china_parameters/analyze.py.


AuthorsOpenReview ↗

Response to Reviewer CvHR (Part 2/5)

W1 (continued).

(c) Core results under the aligned parameters. We reran L2 (30 seeds per cell) and replayed the saved L3 programs under both parameter settings on 20 common seeds per comparison; the data mapping, programs and comparisons were fixed in advance. Appendix results quoted elsewhere keep their original seeds.

ComparisonSubmitted parametersChinese parameters
L2: audit q 0 → 0.5, balanced regimeUp-coding 0.636 → 0.001; rejection gap 0.100 → 0.233Up-coding 0.553 → 0.062; rejection gap −0.1764 → −0.1760
L3: first doctor response, fixed contractH 1.009 → 0.773 (DeepSeek), 0.999 → 0.603 (GLM); UDU_D risesH 1.008 → 0.769 (DeepSeek), 0.988 → 0.564 (GLM); ΔUD=+1.587,+9.043\Delta U_D=+1.587,+9.043
L3: E0 → ABR-A, five configurationsTreated and all-arrival H fall, UDU_D rises: 5/5Same: 5/5
L3: three saved searches each at α = 0.5 / α = 00/3 vs 3/3 recover effort; up-coding ≈ 86%0/3 vs 3/3; at α = 0, H 0.417 → 0.971–0.986, up-coding 63–67%
L3: three terminal-contract responsesΔUD=+0.38\Delta U_D=+0.38 to +0.46+0.46; H fallsΔUD=+0.26\Delta U_D=+0.26 to +0.32+0.32; H falls in 3/3
L3: safeguards, starting → adapted doctorS1: H 0.995 → 0.962; S2: 0.993 → 0.993S1: H 0.987 → 0.948; S2: 0.986 → 0.985, 76% up-coding

Fixed-contract responses remain profitable and reduce health. S2 maintains health for the replayed responses (all-arrival H 0.955 → 0.955; treated-H difference −0.0003, paired 95% CI [−0.0017, +0.0011]). Audits still suppress up-coding but no longer widen the high/low-CMI rejection gap: changing load alone keeps the increase (0.154 → 0.254), changing DRG weights alone does not (−0.273 → −0.298), so this effect depends on the weight-based CMI definition.

SPARCS pilot (same seed counts, fresh seeds). From E0 to ABR-A, both health measures fall and UDU_D rises in 5/5 configurations. Under fixed contracts, the first DeepSeek and GLM responses gain ΔUD=+8.963\Delta U_D=+8.963 and +23.304+23.304 (intervals exclude zero) while H falls 0.983→0.743 and 0.983→0.277; all three terminal-contract responses are also profitable and reduce health. The endpoint and first-response results hold under an alternative severity/urgency encoding.

Audits cut up-coding 0.611→0.045 while the rejection gap between APR severity 3–4 and 1–2 rises 0.284→0.411; under the original relative tariffs it instead falls 0.147→0.062, so the selection effect depends on payment structure. S2 keeps H at 0.954 (paired difference +0.0002, 95% CI [−0.0007, +0.0010]) with 94% up-coding. At α=0, all three saved responses restore effort and health; two still raise UDU_D.

(d) LLM mechanism assessment. Five models (DeepSeek V4.1 Flash, GPT-5.4 mini, Gemini 3.8 Flash, Qwen 3.8 Flash, GLM-5.3 Flash) answered our questionnaire under five professional role prompts, four repeats each, giving 100 responses (rebuttal/llm_questionnaire: instrument and English-translated answers). These conceptual checks supplement the empirical analyses and are not clinical expert validation. All responses agreed that initial performance does not establish performance after providers adapt, and that provider returns, hospital scores and patient care can diverge; this supports the motivation, not specific results. The rule-specific answers are more informative:

  • Stronger coding audits. All 100 judged that audits make upcoding less attractive; on a shift to patient selection, 30 expected an increase and 54 judged the information insufficient. Those expecting a shift tied it to low-cost substitutes such as admission discretion, which mandatory admission or constraints on other channels would block.
  • Minimum delivered-care rule. All 100 judged that it blocks payment below the minimum and none that it directly constrains upcoding: 75 chose "no direct constraint" and 25 "may indirectly affect", citing changes in revenue structure or threshold pressure. The two answers are compatible.

Both premises and the no-direct-constraint judgment also hold in all 60 responses from GPT, Gemini and Qwen, model families not used in the search.

AuthorsOpenReview ↗

Response to Reviewer CvHR (Part 3/5)

W2. Quality of LLM-generated programs; optimality and convergence.

  • Quality. Before execution, each proposal must pass the typed, assignment-only DSL checks, and the host projects its outputs onto admissible actions. In the DeepSeek administrative response blocks, 1180–1196 of 1200 proposals per configuration are valid (App. L.10). The released programs also explain the outcomes. For example, when the profit weight is removed (α = 0) from a low-effort starting doctor, the selected program sets effort to min(1.5 * base_effort, stamina_ratio) and accepts all cases subject to resource constraints. Health recovers from 0.411 to 0.999, while up-coding stays at ~86%, linking the recovery to the effort rule (App. L.8).
  • Optimality. "Best response" denotes the program selected by budget-limited search (Sec. 3). Our central conclusions are existence claims about profitable responses found by that search, as formalized in (iii).

Convergence and response certificates. We use the notation of App. N. Retained libraries are LA⊆PA\mathcal L_A\subseteq\mathcal P_A and LD⊆PD\mathcal L_D\subseteq\mathcal P_D, and ρ\rho is a joint distribution with marginals qA,qDq_A,q_D. Let vr(ρ)=E(A,D)∼ρMr(A,D)v_r(\rho)=\mathbb E_{(A,D)\sim\rho}M_r(A,D) and br(p;q−r)=Ep′∼q−rMr(p,p′)b_r(p;q_{-r})=\mathbb E_{p'\sim q_{-r}}M_r(p,p'). The deviation gap is Er(ρ)=[sup⁡p∈Prbr(p;q−r)−vr(ρ)]+,\mathcal E_r(\rho)=\Big[\sup_{p\in\mathcal P_r}b_r(p;q_{-r})-v_r(\rho)\Big]_+ , and ρ\rho is a CCE iff EA(ρ)=ED(ρ)=0\mathcal E_A(\rho)=\mathcal E_D(\rho)=0.

(i) Why ABR needs a separate stability check. A finite game has the finite improvement property iff it admits a generalized ordinal potential (Monderer & Shapley, GEB 1996, Lemma 2.5); we do not assume this structure here. Our trajectories show why a recovered endpoint still needs testing. Along DeepSeek ABR-D, the contract moves to α=0\alpha=0 and the doctor restores effort. At the terminal contract (A3: α=0.1,β=1\alpha=0.1,\beta=1), all three independent doctor searches again cut effort and resume up-coding (87–89%; App. L.8). We therefore use ABR as a stress test and base equilibrium statements on history-retaining CSRO–MGCCE.

(ii) Exact termination over a finite program class (Proposition 2). A finite universe can be defined by bounding normalized DSL source length by LL over a finite alphabet V\mathcal V: ∣Pr∣≤∑ℓ=1L∣V∣ℓ<∞.|\mathcal P_r|\le\sum_{\ell=1}^{L}|\mathcal V|^{\ell}<\infty . With exact response oracles and the same exact payoffs for admission and the CCE solver, each admission adds a novel program and libraries never shrink. At most (∣PA∣−∣LA0∣)+(∣PD∣−∣LD0∣)(|\mathcal P_A|-|\mathcal L_A^0|)+(|\mathcal P_D|-|\mathcal L_D^0|) rounds admit a program. The first round with no admission returns an exact CCE over that universe, because max⁡p∈Prbr(p;q−r) ≤ max⁡p∈Lrbr(p;q−r) ≤ vr(ρ).\max_{p\in\mathcal P_r}b_r(p;q_{-r})\ \le\ \max_{p\in\mathcal L_r}b_r(p;q_{-r})\ \le\ v_r(\rho). This establishes finite termination under the stated assumptions; the experiments use fixed proposal budgets.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.