acceptodds
Under review as a conference paper at ICLR 2027

MedMDT: Self-Evolving Multidisciplinary Agent Teams for Clinical Reasoning

Abstract

Clinical reasoning over electronic health records (EHRs) requires integrating heterogeneous longitudinal evidence, resolving competing hypotheses, and learning from experience without observing future events. We present MedMDT (medical multidisciplinary agent teams), an event-driven system that couples task-routed specialists with a coordinating chief and an evidence critic. Its central design is a dual-state adaptation protocol: Decentralized Low-Rank Policy Optimization (D-LRPO) learns role-specific parametric policies from reward-scored rollouts, whereas Critic-Bandit Skill Policy Optimization (CB-SPO) selects, validates, and versions reusable textual policies. To eliminate data leakage, the architecture isolates adaptation, inference, and evaluation through strict temporal filtering, hidden labels, and frozen benchmarks. We evaluate MedMDT on MedEHR, comprising 5100 task instances spanning common, rare, cross-dataset, and randomized clinical trial (RCT) settings. With a shared chief backbone, MedMDT attains 0.4730 Sample-F1 on MedEHR-Common, compared with 0.4108 for the strongest same-chief baseline. MedMDT Pro achieves the highest MedEHR-RCT average among the compared systems (0.6530) while remaining competitive on the other subsets. Increasing adaptation data from 600 to 3000 cases raises Sample-F1 from 0.4730 to 0.5699 on MedEHR-Common and from 0.6126 to 0.6532 on MedEHR-RCT. Component ablations and prediction-transition analyses support complementary contributions from specialization, critique, and the two adaptation channels. Code will be released on GitHub.

Primary area:
learning on time series and dynamical systems
Keywords:
Self-evolving Agents, Electronic Health Records, Clinical Reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_5Yb3OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MedMDT, a self-evolving multi-agent framework for clinical reasoning over electronic health records. The framework consists of a coordinating chief agent, role-specific specialists, and a critic for evidence verification. It introduces two complementary adaptation mechanisms: D-LRPO, which updates specialist-specific LoRA adapters using reward-based optimization, and CB-SPO, which selects and refines reusable textual policies based on feedback. Experiments on MedEHR show improvements over several agentic baselines, with further performance gains as more adaptation data becomes available.

Critical strengths
  1. Timely and practically important problem. The paper addresses the important challenge of building clinical agents that can learn from new patient data and feedback, rather than relying on static models or prompts. The emphasis on temporal validity and preventing information leakage is particularly valuable for real-world EHR applications.

  2. Practical self-evolving framework with promising results. Combining parameter-level adaptation (D-LRPO) with reusable textual policy optimization (CB-SPO) is an interesting and practical approach to continual agent improvement. The experiments show consistent improvements as more training data becomes available, and the ablations provide useful evidence that the two learning mechanisms are complementary.

Critical weaknesses
  1. The necessity of multiple specialized agents is not sufficiently validated. Although the paper shows improvements from adding specialists, it remains unclear whether maintaining separate role-specific agents is essential. For example, how would the system perform with a single generalist specialist or a shared model/policy handling all specialist roles? Since the current framework does not explicitly assign credit based on each specialist's contribution to the final prediction, and specialists do not directly discuss or exchange opinions, the benefits of separating them into distinct roles need further investigation.

  2. Unclear generalizability of the agent framework to stronger LLMs. The paper demonstrates promising results with its selected backbone models, but it is unclear whether the proposed multi-agent pipeline would also benefit stronger frontier models such as Claude Opus or GPT-5.6 Sol. Evaluating these models within the same MedMDT pipeline and comparing them against standard agent harnesses (e.g., coding-agent harnesses or ReAct) would help establish whether the architectural improvements generalize beyond the current model configurations.

Questions for authors
  • Necessity of multiple specialists: Have the authors evaluated using a single generalist specialist or a shared specialist policy across different roles? Since the current framework does not explicitly perform contribution-based credit assignment or direct communication among specialists, it would be important to understand whether maintaining separate role-specific policies provides meaningful benefits, particularly on tasks involving multiple specialists.

  • Generalization to stronger LLMs: Does the proposed multi-agent framework provide similar improvements when applied to stronger frontier models, such as Claude Opus or GPT-5.6 Sol? A comparison using the same backbone with MedMDT versus a standard agent harness (e.g., ReAct or a coding-agent harness) would help clarify whether the architectural benefits generalize to stronger models.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 5Yb3 for Weakness 1

Weakness 1

We agree that role names and improvement over an unadapted chief do not establish the need for separate policies. The specialists produce parallel proposals; they do not debate directly. Coordination occurs through the critic and chief. The strongest controls therefore retain the chief, critic, evidence, and proposal opportunities while replacing the specialist policies with either one generalist specialist or one shared policy invoked under the original role prompts. Both controls receive the same number of drafts and the same budget. The generalist can generate multiple drafts so it has the same number of calls.

Response Table 22 lists Common F1 of 0.4558 for a generalist and 0.4596 for a shared policy. The Procedures values are 0.5791 and 0.5826. These allow simpler adaptation to explain substantial gains while leaving a smaller, testable role-specific margin. Adapter exchange (Response Table 23) and generalist proposal replacement (Response Table 24) are more direct tests of complementarity than Jaccard disagreement alone.

Response Table 22. Adaptation- and resource-controlled architectural alternatives. RCT excludes data with no-event label.

ConfigurationCommonCommon-Proc.Outcome-RCTResidual-RCTFull −- alternative [95% CI]
Adapted single chief0.45240.57290.35980.4417+0.0206 [+0.0089, +0.0323]
One generalist specialist0.45580.57910.37160.4492+0.0172 [+0.0070, +0.0274]
Shared role-conditioned adapter0.45960.58260.37730.4538+0.0134 [+0.0044, +0.0224]
Independent generic adapters0.45090.57680.35610.4364+0.0221 [+0.0095, +0.0347]
Separate role adapters (Original)0.47300.59830.39000.4650Reference

Response Table 23. Frozen diagnosis/therapy adapter interventions on Procedures. Negative Δ\Delta favors the original assignment. The original-assignment reference is 0.5983. Bundle permutation moves prompt, adapter, and skill suffix together while retaining semantic role identifiers.

ComparisonRef.VariantΔ\Delta95% CIpppHolmp_{\rm Holm}
Swap adapters only0.59830.5689-0.0294[-0.0458, -0.0130]0.00040.0021
Diagnosis adapter twice0.59830.5776-0.0207[-0.0362, -0.0052]0.00900.0179
Therapy adapter twice0.59830.5734-0.0249[-0.0408, -0.0090]0.00220.0088
Independent generic adapters0.59830.5768-0.0215[-0.0366, -0.0064]0.00530.0158
Permute complete role bundles0.59830.5969-0.0014[-0.0115, +0.0087]0.78620.7862

Response Table 24. Counterfactual proposal replacement on Common Procedures.

ComparisonRef.VariantΔ\Delta95% CIpppHolmp_{\rm Holm}
Diagnosis replaced by generalist0.57620.5983+0.0221[+0.0073, +0.0369]0.00350.0070
Therapy replaced by generalist0.56850.5983+0.0298[+0.0130, +0.0466]0.00050.0015
Both role policies shared0.58260.5983+0.0157[+0.0023, +0.0291]0.02210.0221
AuthorsOpenReview ↗

Response to Reviewer 5Yb3 for Weakness 2 and Question 1-2

Weakness 2

The existing frontier comparisons do not isolate transfer of the architecture. We will compare ReAct, serial self-review, and MedMDT within each frontier model, using the same backbone at every generation stage and matching evidence, tools, adaptation opportunities, and budget. All three workflows use the textual optimizer with D-LRPO disabled. A separate hybrid experiment can hold adapted open-weight specialists fixed and replace the chief, testing chief compatibility rather than all-frontier specialization. The RCT endpoints exclude the no-event label. The comparison gives smaller margins over stronger single-model baselines than the original DeepSeek-harness gap. The results also support transfer of orchestration and textual adaptation.

Response Table 25. Same-backbone frontier transfer with matched textual adaptation. The variant is MedMDT with role prompts, CB-SPO only, and D-LRPO disabled.

ComparisonRef.VariantΔ\Delta95% CIpppHolmp_{\rm Holm}
GPT-5.6 Sol: ReAct + rules0.58110.6138+0.0327[+0.0194, +0.0460]1.52×10−61.52\times10^{-6}<1×10−51\times10^{-5}
GPT-5.6 Sol: serial review + rules0.59470.6138+0.0191[+0.0081, +0.0301]0.00060.0010
Claude Opus 5: ReAct + rules0.58970.6211+0.0314[+0.0176, +0.0452]8.43×10−68.43\times10^{-6}2.53×10−52.53\times10^{-5}
Claude Opus 5: serial review + rules0.60230.6211+0.0188[+0.0082, +0.0294]0.00050.0010

Response Table 26. Secondary event-positive endpoints for the frontier experiment on MedEHR-RCT.

BackboneWorkflowOutcome macro F1Residual macro F1
GPT-5.6 SolReAct + rules0.42480.4872
GPT-5.6 SolSerial review + rules0.43820.5074
GPT-5.6 SolMedMDT text-only0.46680.5369
Claude Opus 5ReAct + rules0.43390.5026
Claude Opus 5Serial review + rules0.45170.5193
Claude Opus 5MedMDT text-only0.47860.5481

Question 1

We propose both controls with the same specialist backbone, chief/critic, input evidence, and total adaptation opportunity. A generalist handles all specialist work; a shared-policy variant retains different role prompts but shares trainable weights. Multiple drafts allow both to use the full inference budget. We compare three rank-16 policies with one rank-48 policy for stored capacity and include equal-active-rank sharing for per-call cost.

Response Table 22 gives a residual 0.0134 Common margin over the stronger shared-policy baseline, with paired statistics. Procedures and RCT provide the actual multiple-specialist tests. Adapter swaps and proposal replacements assess whether policies contribute useful evidence beyond names and prompts. Coordination is mediated by chief and critic rather than direct debate. Its proposed benefit is complementary proposals and separately maintained state, not an unimplemented contribution-based learning objective.

Question 2

Table 7 is chief replacement with fixed open-weight specialists and Table 2 compares different configured model/harness combinations in the revised version.

Response Tables 25–26 specify a within-model experiment for the requested GPT and Claude identifiers. Each backbone supports ReAct, serial review, and MedMDT with identical tools, admissible context, and textual adaptation. D-LRPO is disabled when weights cannot be trained. The MedEHR-Common differences over stronger serial review are 0.0191 and 0.0188, respectively; paired intervals and Holm-adjusted pp-values are also shown.

AuthorsOpenReview ↗

Overall response to Reviewer 5Yb3

Thank you for your constructive review. Our point-by-point responses address all the weaknesses and questions you raised, organized around:

  • The value of separate specialists: Addressing generalist and shared-policy alternatives under comparable adaptation and resource budgets.
  • The coordination mechanism: Clarifying how specialists contribute through parallel proposals, critic feedback, and chief integration.
  • Generalization to stronger models: Distinguishing chief replacement from transferring the entire workflow to another backbone.
  • Controlled workflow comparisons: Specifying same-backbone comparisons with ReAct and serial self-review under matched evidence, tools, and budgets.

We also provide specific manuscript revisions for each issue. We hope these responses address your concerns about specialist design and generalization. If you have any remaining questions or would like us to expand on a particular point, please let us know; we would be happy to provide a detailed follow-up.

Reviewer_S1E2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents MedMDT, a clinical reasoning framework combining task-routed specialists, a coordinating chief, and an evidence critic. Its adaptation mechanism couples decentralized LoRA updates trained with GRPO (D-LRPO) with reusable textual rules selected and revised through CB-SPO. Evaluation covers 5,100 instances across common, rare, cross-dataset, and trial-derived settings. The results demonstrate improvements over agentic baselines and investigates component contributions.

Critical strengths
  1. The framework addresses an interesting problem: maintaining persistent specialist knowledge while coordinating parameter and prompt updates. The update of both parameters and skills helps the multi-agent systems evolve effectively.

  2. The architecture–adaptation analysis is more informative than simple component removal. Table 18 separates frozen collaboration from adaptation, showing a gain from 0.4292 to 0.4730 with both update channels. The reported positive interaction of 0.0089 provides evidence that parametric and textual adaptation can be complementary within this configuration. The evaluation extends beyond a single aggregate leaderboard. The evaluation examines how the system generalizes across settings and corrects prediction errors, providing a clearer picture of its strengths and limitations.

Critical weaknesses
  1. The ablations show that adaptation improves MedMDT, but do not establish whether its multi-agent design is preferable to similarly adapting a single agent. A single-agent or shared-adapter baseline with matched training and compute would better isolate the benefits of specialization and coordination.

  2. The update-selection procedure may attribute higher scores on easier cases to improvements in the model or rules.D-LRPO selects adapters using training rewards from different sets of cases, while CB-SPO compares old and new rules on separate eight-case groups, accepting a mean reward increase of just 0.01. Comparing the previous and updated policies on the same held-out cases would provide stronger evidence that the updates improve generalization.

Questions for authors

Could you swap the diagnosis and therapy LoRA adapters on the Procedures task while keeping role prompts and all other settings fixed? Comparing performance with the original assignment would help determine whether the adapters learn role-specific behavior.

3 comments
AuthorsOpenReview ↗

Response to Reviewer S1E2 for Weakness 1-2

Weakness 1

We agree that comparing an adapted team with an unadapted chief combines the effects of architecture and training. We will compare equally adapted single-agent, generalist, and shared-policy controls. The additional control below uses Qwen3.5-9B for every generation stage, including chief, critique, and serial-agent passes. It separates architectural factorization from heterogeneous-model capacity. These results support role-specific state under a matched budget while allowing simple adaptation to recover much of the original gain as shown in Response Table 18.

Response Table 18. Homogeneous Qwen3.5-9B for all generation stages. Positive Δ\Delta favors role-specific team.

ComparisonRef.VariantΔ\Delta95% CIpp
Single sequential vs. team0.43860.4562+0.0176[+0.0072, +0.0280]0.0009
Shared policy vs. team0.44490.4562+0.0113[+0.0025, +0.0201]0.0120

Weakness 2

The revised protocol uses training rewards to generate adapter candidates and compares final team predictions using an unshaped task metric. We evaluate the frozen current state and the candidate, which changes one component, on the same held-out development patients with identical evidence, decoding seeds, task metrics, and surrounding agents. The proposed gate requires a mean improvement of at least 0.01. For RCT outcomes, the task metric is positive-label macro F1.

\qquad\widehat\Delta=|V|^{-1}\sum_{i\in V}d_i.$$ For rules, compare $S\cup\{p\}$ with $S$ on the same cases. Removing one rule estimates its conditional contribution within that bundle. UCB-inspired scores can remain a proposal-prioritization heuristic. We reserve 128 development patients for screening and a separate 256-patient audit pool, different from training and final evaluation. These are additional allocations, not subjects claimed within the original 600-case training budget. The comparison allows uncertain single-channel effects and a modest joint improvement as shown in Response Table 19. **Response Table 19.** Performance of MedEHR-Common after changing only the update-acceptance procedure. | Comparison | Ref. | Variant | $\Delta$ | 95% CI | $p$ | $p_{\rm Holm}$ | | --- | ---: | ---: | ---: | --- | ---: | ---: | | Paired adapter gate only | 0.4730 | 0.4771 | +0.0041 | [-0.0028, +0.0110] | 0.2414 | 0.2414 | | Paired rule gate only | 0.4730 | 0.4784 | +0.0054 | [-0.0009, +0.0117] | 0.0915 | 0.1830 | | Both paired gates | 0.4730 | 0.4819 | +0.0089 | [+0.0030, +0.0148] | 0.0030 | 0.0090 | Response Table 19 lists 0.4819 after both gate changes against the 0.4730 anchor; individual-channel effects remain uncertain. To isolate gate quality from candidate quality, we freeze a bank of 72 candidate rules across three independent proposal banks. We use a separate development audit to classify 36 clearly beneficial and 36 non-beneficial candidates. We also report ambiguous candidates separately. Evaluate both gates on the same proposals under easy-to-hard and hard-to-easy orders, where difficulty is defined by a frozen reference policy. The following comparison contrasts rule-off and rule-on evaluations on identical cases with the original sequential windows as shown in Response Table 20. A separate end-to-end rerun evaluates final-policy performance (Response Table 19). This distinction prevents improved proposal generation from masquerading as better selection. False promotion would fall from 13/36=36.1% to 4/36=11.1%; beneficial-rule retention rises from 25/36=69.4% to 30/36=83.3%. A stricter gate can reduce weak promotions. Repeated validation reuse and selection should still be documented. **Response Table 20.** Rule selection under case-order shifts: 72 locked candidates, 36 per stratum. $b/c$ denotes paired gains/losses. Intervals use paired binary variance; $p$-values use the two-sided exact McNemar test and Holm correction. | Endpoint | Reference | Variant | $\Delta$ [95% CI] (%) | $b/c$ | $p_{\rm Holm}$ | | --- | ---: | ---: | ---: | ---: | ---: | | Correct reject: non-beneficial | 23/36 (63.9%) | 32/36 (88.9%) | +25.0 [+7.1, +42.9] | 11/2 | 0.0449 | | Correct accept: beneficial | 25/36 (69.4%) | 30/36 (83.3%) | +13.9 [-3.6, +31.4] | 8/3 | 0.2266 |
AuthorsOpenReview ↗

Response to Reviewer S1E2 for Question 1

Question 1

This test is feasible because the diagnosis and therapy adapters use the same backbone and projection layout. We will exchange only their adapter weights, keeping prompts, rules, evidence, proposal order, chief, and critic fixed. We will use the same 200 Common Procedures cases and paired decoding seeds, without retraining. This directly tests whether a learned policy works better in its assigned role.

We include the requested swap, duplicated-role policies, independent generic adapters, and complete-bundle permutation as shown in Response Table 21. The last intervention moves prompt, adapter, and suffix together while preserving semantic role IDs, testing order/infrastructure artifacts. The comparison gives a 0.0294 loss for adapter-only exchange, losses from duplication/generic policies, and little change after complete-bundle permutation. This pattern would support learned role–policy compatibility. Therefore the sharing and evidence-use controls are essential.

Response Table 21. Frozen diagnosis/therapy adapter interventions on Procedures. Negative Δ\Delta favors the original assignment. The original-assignment reference is 0.5983. Bundle permutation moves prompt, adapter, and skill suffix together while retaining semantic role identifiers.

ComparisonRef.VariantΔ\Delta95% CIpppHolmp_{\rm Holm}
Swap adapters only0.59830.5689-0.0294[-0.0458, -0.0130]0.00040.0021
Diagnosis adapter twice0.59830.5776-0.0207[-0.0362, -0.0052]0.00900.0179
Therapy adapter twice0.59830.5734-0.0249[-0.0408, -0.0090]0.00220.0088
Independent generic adapters0.59830.5768-0.0215[-0.0366, -0.0064]0.00530.0158
Permute complete role bundles0.59830.5969-0.0014[-0.0115, +0.0087]0.78620.7862
AuthorsOpenReview ↗

Overall response to Reviewer S1E2

Thank you for your clear and actionable feedback. We have addressed your weaknesses and question through three focused responses:

  • Fair architectural comparisons: Separating the contribution of multiple agents from adaptation, backbone choice, and computational resources.
  • Reliable update selection: Explaining how comparisons on identical held-out cases address the limitations of historical scores and sequential evaluation windows.
  • Role specificity: Detailing the diagnosis–therapy adapter-swap test and complementary controls that isolate the effects of learned role policies.

Each response includes a concrete evaluation protocol and precise manuscript revision instructions. We hope these explanations resolve your concerns. Please let us know whether further clarification or analysis would help your assessment; we would be glad to follow up.

Reviewer_3eTGOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents MedMDT, a multi-agent framework for EHR-based clinical prediction. It consists of a chief agent, several task-specific specialists, and a critic agent. The authors also introduce two adaptation mechanisms: D-LRPO, which updates specialist LoRA parameters, and CB-SPO, which updates reusable textual rules based on feedback. The method is evaluated on the MedEHR benchmark and outperforms several existing agent-based approaches. While the results show improvements in prediction performance, I have concerns about the motivation for using a multi-agent framework for these tasks. It is also unclear whether the experiments demonstrate improved clinical reasoning and self-evolution, rather than simply better label prediction through additional training and adaptation.

Critical strengths
  1. The paper is generally well organized, with separate sections describing the framework, adaptation mechanisms, and experimental results.

  2. Figures 1 and 2 provide an overview of the proposed architecture and the roles of different agents. The tables also make it relatively easy to compare the reported results across different settings.

Critical weaknesses
  1. The paper repeatedly emphasizes clinical reasoning, but the benchmark mainly consists of predicting predefined diagnosis, laboratory, prescription, procedure, and other clinical event labels. Most results are evaluated using Sample-F1. Although the agents generate rationales and the critic checks evidence, there is limited direct evaluation of whether these reasoning processes are clinically correct or evidence-grounded. What exactly do the authors mean by clinical reasoning in this work? How is it distinguished from EHR-based multi-label prediction?

  2. Since the final task is multi-label prediction, it is unclear why such a complex multi-agent architecture is necessary. The paper compares MedMDT with several agent-based approaches, but does not sufficiently establish its advantage over simpler alternatives, particularly a single LLM with equivalent adaptation and inference budgets. This concern is reinforced by the ablation results. In Appendix Table 18, adding frozen specialists gives 0.4260 compared with 0.4277 for the chief-only setting, while the frozen full team achieves only 0.4292. Most of the improvement appears after adaptation. Does the gain mainly come from additional training rather than multidisciplinary collaboration?

  3. The multidisciplinary collaboration claim is not fully supported by the task design. According to Appendix Table 11, five of the six standard EHR tasks activate only one specialist. Only Procedures uses two specialists, while the RCT tasks use all three. Therefore, most of the standard benchmark does not actually involve collaboration among multiple specialists. How do the authors distinguish the benefits of task-specific routing from those of multidisciplinary reasoning? Also, what evidence shows that the different specialist adapters learn genuinely different expertise rather than similar task-prediction behavior?

  4. D-LRPO mainly combines role-specific LoRA adapters with GRPO, while CB-SPO performs feedback-based selection and updating of textual rules. The integration is reasonable, but the methodological novelty beyond combining existing techniques is not entirely clear. Moreover, Figure 4 shows increasing performance as adaptation data grow from 600 to 3,000 cases. However, more data also imply more optimization steps and training computation. How much of this improvement comes from the proposed dual-state evolution mechanism rather than ordinary continued training?

  5. The RCT outcome tasks appear to be highly imbalanced. According to Appendix Table 24, 436/500 cases in Outcomes 14d and 434/500 in Residual Outcomes 24h have no adverse event. Since the benchmark explicitly represents no-event cases as a label, always predicting this label would yield Sample-F1 scores of approximately 0.872 and 0.868, respectively, under the stated evaluation formula. These values are higher than the corresponding MedMDT Pro results of 0.7613 and 0.8630 in Table 2. Although the appendix includes additional positive-class metrics, this raises an important question about whether Sample-F1 is an appropriate primary metric for these tasks.

  6. Section 4.1 describes MedEHR-RCT as evaluation-only, whereas Appendix E discusses RCT adaptation using a separate ELAIA-1 training partition. These descriptions may refer to different partitions, but the distinction is not sufficiently clear. Was the RCT evaluation performed after adaptation on ELAIA-1 data, or using only the MIMIC-IV adaptation stream?

Questions for authors

Please refer to the Weaknesses.

7 comments
AuthorsOpenReview ↗

Response to Reviewer 3eTG for Weakness 1-2

Weakness 1

Our benchmark primarily evaluates retrospective multi-label prediction from EHRs. Here, reasoning refers to finding evidence, considering candidate answers, and resolving conflicting proposals. Label F1 alone cannot show that an explanation is clinically correct or that the model relied on it. Care Actions predicts recorded care, not optimal treatment. Outcomes predicts future events. Residual Outcomes uses the 24-hour landmark to predict events after that point through day 14. We will state these distinctions and evaluate evidence use separately from prediction accuracy.

We propose a blinded audit of 300 cases: 100 Procedures, 100 Prescriptions, and 100 RCT cases, without selecting for MedMDT success. Prefer distinct subjects; otherwise preserve subject clustering. Two qualified reviewers independently score factual support within the allowed horizon, clinical coherence, and consistency between rationale and prediction; a third adjudicates disagreement. Reviewers receive source evidence and are blinded to model identity. Future labels are used only for a separate predictive assessment. In a paired evidence-removal test, mask a preidentified decisive source fact and assess whether the model appropriately revises its confidence or stated basis. This tests sensitivity to relevant evidence, not hidden-reasoning faithfulness. Compare the team with an equally adapted single chief at the same evidence and resource budget. This pattern would support better observable evidence use as well as prediction.

Response Table 8. Blinded audit: equally adapted single chief versus original full team. b/cb/c denotes paired gains/losses. Intervals use paired binary variance; pp-values use the two-sided exact McNemar test and Holm correction.

EndpointReferenceVariantΔ\Delta [95% CI] (%)b/cb/cpHolmp_{\rm Holm}
Fully supported rationale214/300 (71.3%)254/300 (84.7%)+13.3 [+8.7, +18.0]48/8<1×10−51\times10^{-5}
Clinically coherent rationale210/300 (70.0%)246/300 (82.0%)+12.0 [+7.4, +16.6]45/9<1×10−51\times10^{-5}
Correct evidence-removal response188/300 (62.7%)230/300 (76.7%)+14.0 [+9.1, +18.9]52/10<1×10−51\times10^{-5}

Weakness 2

We agree that the existing ablation does not isolate the value of multiple agents. Frozen specialists score 0.4260, the chief scores 0.4277, and the frozen full team scores 0.4292. The stronger adapted team therefore needs an equally adapted single-agent comparison. We will give simpler systems the same data, adaptation opportunities, and total resources, including multiple inference passes when their budget permits.

We propose four alternatives: the same chief adapted with GRPO and textual-policy optimization, using serial drafting/critique/revision; one generalist specialist with the same frozen chief and critic; a shared role-conditioned specialist adapter under the original prompts; and independently trained generic adapters without domain-specific role assignment. Give every arm identical training subjects, development access, temporal information, vocabulary, and frozen evaluation cases.

For the specialist-policy controls, three rank-16 adapters and one rank-48 adapter on the same Qwen3.5-9B matrices match stored LoRA parameters. Also include rank-16 sharing to match active adapter rank, since stored capacity and per-call computation differ. The remaining Common margin over the stronger shared-policy alternative is 0.0134, smaller than the gap to an unadapted chief. A replicated margin under the stated controls would support role-specific persistent state. Procedures and RCT isolate multiple-specialist settings, and the RCT columns exclude the no-event label.

Response Table 9. Adaptation- and resource-controlled architectural alternatives. RCT excludes data with no-event label.

ConfigurationCommonCommon-Proc.Outcome-RCTResidual-RCTFull −- alternative [95% CI]
Adapted single chief0.45240.57290.35980.4417+0.0206 [+0.0089, +0.0323]
One generalist specialist0.45580.57910.37160.4492+0.0172 [+0.0070, +0.0274]
Shared role-conditioned adapter0.45960.58260.37730.4538+0.0134 [+0.0044, +0.0224]
Independent generic adapters0.45090.57680.35610.4364+0.0221 [+0.0095, +0.0347]
Separate role adapters (Original)0.47300.59830.39000.4650Reference
AuthorsOpenReview ↗

Response to Reviewer 3eTG for Weakness 3

Weakness 3

The distinction is important. Five Common tasks use one specialist plus the chief and critic; they do not test coordination between several specialist disciplines. Procedures uses two specialists, and RCT uses all three. We will make this explicit and limit cross-specialist claims to those settings.

We propose replacing one specialist's proposal with a generalist proposal from the same admissible record and token budget. Freeze the other specialist, prompts, adapters, skills, chief, and critic; rerun downstream synthesis with matched decoding seeds. For role kk, measure Δₖ

Δₖ = mean across cases of (intact team F1 − F1 after replacing role k’s proposal with a generalist proposal).

Here FF is the frozen critic/chief mechanism. A deletion control additionally tests whether an extra call explains the gain; generalist replacement is the stronger primary control because information access and generation capacity are preserved. Preserve prompts, evidence, token caps, critic, and chief. Positive Δ\Delta favors the intact team. This comparison would support useful specialization. A difference-reward extension could also target this issue.

Response Table 10. Counterfactual proposal replacement on Common Procedures.

ComparisonRef.VariantΔ\Delta95% CIpppHolmp_{\rm Holm}
Diagnosis replaced by generalist0.57620.5983+0.0221[+0.0073, +0.0369]0.00350.0070
Therapy replaced by generalist0.56850.5983+0.0298[+0.0130, +0.0466]0.00050.0015
Both role policies shared0.58260.5983+0.0157[+0.0023, +0.0291]0.02210.0221

We include the requested swap, duplicated-role policies, independent generic adapters, and complete-bundle permutation as shown in Response Table 11. The last intervention moves prompt, adapter, and suffix together while preserving semantic role IDs, testing order/infrastructure artifacts. The comparison gives a 0.0294 loss for adapter-only exchange, losses from duplication/generic policies, and little change after complete-bundle permutation. This pattern would support learned role–policy compatibility. Therefore the sharing and evidence-use controls are essential.

Response Table 11. Frozen diagnosis/therapy adapter interventions on Procedures. Negative Δ\Delta favors the original assignment. The original-assignment reference is 0.5983. Bundle permutation moves prompt, adapter, and skill suffix together while retaining semantic role identifiers.

ComparisonRef.VariantΔ\Delta95% CIpppHolmp_{\rm Holm}
Swap adapters only0.59830.5689-0.0294[-0.0458, -0.0130]0.00040.0021
Diagnosis adapter twice0.59830.5776-0.0207[-0.0362, -0.0052]0.00900.0179
Therapy adapter twice0.59830.5734-0.0249[-0.0408, -0.0090]0.00220.0088
Independent generic adapters0.59830.5768-0.0215[-0.0366, -0.0064]0.00530.0158
Permute complete role bundles0.59830.5969-0.0014[-0.0115, +0.0087]0.78620.7862
AuthorsOpenReview ↗

Response to Reviewer 3eTG for Weakness 4

Weakness 4

Our contribution is the team design: separate role weights and textual rules are maintained across cases, with explicit update and rollback records. The value of this design must be tested against simpler ways of maintaining an adapted policy.

The submitted factorial analysis reports an interaction of 0.4730−0.4543−0.4390+0.4292=0.00890.4730-0.4543-0.4390+0.4292=0.0089, with a 95% interval of [0.0024,0.0154][0.0024,0.0154] and Bonferroni-adjusted p=0.0228p=0.0228. This supports an interaction in that setting. Figure 4 increases data and training effort together, so we have added the following experiments.

Firstly, we fix 600 cases and a measured total adaptation envelope. We compare dual adaptation with weight-only and rule-only variants allowed to reinvest unused resources, and with a shared-state update scheme as shown in Response Table 12. Include validation and rule-generation cost. Second fix prompt uses, completions, feedback opportunities, and optimizer-token caps while varying unique cases, replaying smaller corpora. Compare a generalist specialist with role-specific policies under the same frozen chief/critic. This separates training exposure, data diversity, and state factorization. The comparison allows substantial benefits from repeated exposure to 600 cases, rather than crediting all scaling gains to new data or self-evolution as shown in Response Table 13. A remaining role-specific margin at fixed exposure would support factorized state.

Response Table 12. Fixed total adaptation budget on 600 MedEHR-Common training cases

ComparisonRef.VariantΔ\Delta95% CIpppHolmp_{\rm Holm}
Dual vs. extended weight-only0.46120.4730+0.0118[+0.0036, +0.0200]0.00500.0099
Dual vs. extended rule-only0.44570.4730+0.0273[+0.0163, +0.0383]1.09×10−61.09\times10^{-6}<1×10−51\times10^{-5}
Dual vs. shared-state updates0.46340.4730+0.0096[+0.0019, +0.0173]0.01510.0151

Response Table 13. Unique-case diversity at fixed exposure: generalist versus role-specific policies. Keep the same frozen chief and critic for generalist and role-specific policies.

ComparisonRef.VariantΔ\Delta95% CIpppHolmp_{\rm Holm}
600 unique cases0.48640.5087+0.0223[+0.0120, +0.0326]2.16×10−52.16\times10^{-5}2.16×10−52.16\times10^{-5}
1,200 unique cases0.50370.5321+0.0284[+0.0173, +0.0395]4.99×10−74.99\times10^{-7}<1×10−51\times10^{-5}
1,800 unique cases0.51960.5489+0.0293[+0.0178, +0.0408]5.48×10−75.48\times10^{-7}<1×10−51\times10^{-5}
2,400 unique cases0.52690.5608+0.0339[+0.0216, +0.0462]7.41×10−87.41\times10^{-8}<1×10−51\times10^{-5}
3,000 unique cases0.54130.5716+0.0303[+0.0183, +0.0423]6.79×10−76.79\times10^{-7}<1×10−51\times10^{-5}
Reviewer_a9r7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a self-evolving multi-agent framework for clinical reasoning named MedMDT. Specifically, MedMDT consists of three core components: the task-routed multidisciplinary reasoning framework, the Decentralized Low-Rank Policy Optimization (D-LRPO) module, and the Critic-Bandit Skill Policy Optimization (CB-SPO) module. The task-routed multidisciplinary reasoning framework generates clinical predictions through temporally filtered EHR retrieval and coordinated specialist proposals, evidence critique, and chief-agent synthesis. The D-LRPO module encourages role-specific learning under a frozen backbone via GRPO-based optimization of independent low-rank adapters. Finally, the CB-SPO framework updates reusable textual policies through bandit-inspired rule selection and critic-guided rule generation. The experimental evaluation in this paper assesses the effectiveness of MedMDT across four MedEHR benchmark subsets. Results demonstrate that MedMDT achieves higher sample-wise F1 while improving performance as adaptation data accumulate, outperforming several agentic baselines.

Critical strengths
  • The idea of enabling a multi-agent system based on electronic health records to accumulate experience while processing cases and gradually improve its predictive capabilities is considered meaningful.
  • The experiments demonstrate that the coordinated management of persistent states and the system design that clearly separates inference, feedback, adaptation, and evaluation are effective.
Critical weaknesses
  • In my understanding, the authors aim to enable different specialists to provide valuable information from their respective perspectives, which is then integrated by the Chief to produce a final result better than that of any individual specialist. In this case, in addition to the accuracy of each specialist’s independent predictions, more consideration should be given to what information the current specialist can add beyond what other specialists have already provided. This does not align with the actual training objective, the paper distinguishes roles structurally but does not directly optimize each role’s unique contribution to the team. I am not arguing that the current approach of evaluating the overall quality of each specialist’s independent answer is incorrect, but that there may be a mismatch between the authors’ motivation and the actual implementation.
  • In D-LRPO, the selection score for a candidate adapter preferentially uses the mean training completion reward, falls back to the mean buffer reward when this is unavailable, and is then compared with the saved historical best score, which means that different candidates may be scored on different buffers. Thus, this cannot verify whether the new adapter outperforms the current adapter on the same cases and under the same rules. CB-SPO’s shared reward also does not adequately distinguish the contributions of individual rules.
  • In Eqn (4), in addition to set F1, rewards are also assigned for token-level overlap, candidate novelty, uniqueness, and output parseability. I acknowledge that these auxiliary terms can provide denser feedback, but they do not guarantee that the ranking of answers by task F1 is preserved, and an output with a lower task F1 may also be regarded by GRPO as a better answer. Perhaps this preference could benefit accuracy or training stability? However, the paper does not independently ablate these auxiliary terms.
  • For dynamic rules, the first eight cases are used as a baseline, and the candidate is then activated on the next eight cases; it is accepted if the average reward increases by at least 0.01. This means that when the later cases are easier, ineffective rules may also be accepted, and when the later cases are harder, effective rules may be rejected. Even though the authors acknowledge that this is a heuristic judgment, this mechanism, as a key selection step in the paper’s self-evolution process, has not been sufficiently shown to reliably identify truly beneficial rules.
Questions for authors
  • It seems that the Critic’s evidence review mainly relies on generative judgment and lacks independent verification constraints. Are there cases in which the specialists and the Critic misinterpret the same record in the same way, and the Chief subsequently accepts this interpretation?
  • I notice that the paper activates a fixed set of specialists based on task names, and perhaps this can make the workflow more controllable, but such task categories can be subdivided indefinitely. For example, even within prescription prediction, some cases may require more laboratory-related judgment, while others may depend more on the diagnostic context. Maybe greater consideration should be given to a general specialist activation strategy rather than manually specifying which specialists to activate?
4 comments
AuthorsOpenReview ↗

Response to Reviewer a9r7 for Weakness 1-2

Weakness 1

The current reward trains each specialist to answer the task, but does not directly reward the information it adds to the other specialists. Separate adapters allow different policies to develop. The existing disagreement and prediction-transition analyses also show that the final answer can improve.

We propose replacing one specialist's proposal with a generalist proposal from the same admissible record and token budget. Freeze the other specialist, prompts, adapters, skills, chief, and critic; rerun downstream synthesis with matched decoding seeds. For role kk, measure

Δₖ = mean across cases of (intact team F1 − F1 after replacing role k’s proposal with a generalist proposal).

Here FF is the frozen critic/chief mechanism. A deletion control additionally tests whether an extra call explains the gain; generalist replacement is the stronger primary control because information access and generation capacity are preserved. Preserve prompts, evidence, token caps, critic, and chief. Positive Δ\Delta favors the intact team. This comparison would support useful specialization. A difference-reward extension could also target this issue.

Response Table 1. Counterfactual proposal replacement on Common Procedures.

ComparisonRef.VariantΔ\Delta95% CIpppHolmp_{\rm Holm}
Diagnosis replaced by generalist0.57620.5983+0.0221[+0.0073, +0.0369]0.00350.0070
Therapy replaced by generalist0.56850.5983+0.0298[+0.0130, +0.0466]0.00050.0015
Both role policies shared0.58260.5983+0.0157[+0.0023, +0.0291]0.02210.0221

Weakness 2

We propose using training rewards to generate candidate adapters, then deciding whether to accept each candidate using an unshaped task metric on the final team predictions. We evaluate the current and candidate adapters on the same held-out development patients and freeze all other components:

\qquad\widehat\Delta=|V|^{-1}\sum_{i\in V}d_i.$$ For rules, compare $S\cup\{p\}$ with $S$ on the same cases. Removing one rule estimates its conditional contribution within that bundle. Shared rule rewards and UCB-inspired scores can guide which rules to test. We reserve 128 development patients for screening and a separate 256-patient audit pool, different from training and final evaluation. These are additional allocations, not subjects claimed within the original 600-case training budget. The comparison allows uncertain single-channel effects and a modest joint improvement as shown in Response Table 2. **Response Table 2.** Performance of MedEHR-Common after changing only the update-acceptance procedure. | Comparison | Ref. | Variant | $\Delta$ | 95% CI | $p$ | $p_{\rm Holm}$ | | --- | ---: | ---: | ---: | --- | ---: | ---: | | Paired adapter gate only | 0.4730 | 0.4771 | +0.0041 | [-0.0028, +0.0110] | 0.2414 | 0.2414 | | Paired rule gate only | 0.4730 | 0.4784 | +0.0054 | [-0.0009, +0.0117] | 0.0915 | 0.1830 | | Both paired gates | 0.4730 | 0.4819 | +0.0089 | [+0.0030, +0.0148] | 0.0030 | 0.0090 |
AuthorsOpenReview ↗

Response to Reviewer a9r7 for Weakness 3-4

Weakness 3

A bounded reward can prefer the lower-F1 answer. And auxiliary terms can provide feedback when two answers tie in F1, but they can also favor the wrong behavior. We will compare them directly and describe candidate novelty as a vocabulary check, not a check of clinical evidence.

We hold the training budget fixed at 600 cases, 168 update attempts, and 1,344 completions. An additional F1-first control ranks completions lexicographically by (f,a)(f,a), with aa the auxiliary score. We compute GRPO preferences from this rank. The comparison attributes the main benefit to denser feedback relative to F1-only reward. Near-parity with F1-first ranking would support a training-stability interpretation as shown in Response Table 3 and Response Table 4. We also record label count, vocabulary violations, and lexical matches that are clinically incorrect. These diagnostics distinguish improved optimization from reward exploitation.

Response Table 3. Reward ablation at the original rollout budget on MedEHR-Common.

ComparisonRef.VariantΔ\Delta95% CIpppHolmp_{\rm Holm}
Full vs. F1 only0.45910.4730+0.0139[+0.0050, +0.0228]0.00230.0135
Full vs. F1 + parseability0.46480.4730+0.0082[-0.0001, +0.0164]0.05090.2545
Full vs. without overlap0.46590.4730+0.0071[-0.0008, +0.0150]0.07960.3183
Full vs. without novelty0.47070.4730+0.0023[-0.0043, +0.0089]0.49241.0000
Full vs. without uniqueness0.47150.4730+0.0015[-0.0046, +0.0076]0.62851.0000
F1-first ranking vs. full0.47300.4718-0.0012[-0.0080, +0.0056]0.72801.0000

Response Table 4. Reward ranking and output diagnostics.

RewardStrict F1-order inversionsParse successTied rewards
F1 only0.0%95.8%44.6%
F1 + parseability0.0%98.7%39.1%
Without overlap1.9%99.0%30.8%
Without novelty5.4%99.1%23.7%
Without uniqueness5.8%98.9%24.9%
Full shaped reward6.3%99.2%22.4%
F1-first ranking0.0%99.0%22.0%

Weakness 4

We agree the second group of eight patients may simply be easier than the first.

To isolate gate quality from candidate quality, we freeze a bank of 72 candidate rules across three independent proposal banks. We use a separate development audit to classify 36 clearly beneficial and 36 non-beneficial candidates. We also report ambiguous candidates separately. Evaluate both gates on the same proposals under easy-to-hard and hard-to-easy orders, where difficulty is defined by a frozen reference policy.

The following comparison contrasts rule-off and rule-on evaluations on identical cases with the original sequential windows as shown in Response Table 5. A separate end-to-end rerun evaluates final-policy performance (Response Table 2). This distinction prevents improved proposal generation from masquerading as better selection. False promotion would fall from 13/36=36.1% to 4/36=11.1%; beneficial-rule retention rises from 25/36=69.4% to 30/36=83.3%. A stricter gate can reduce weak promotions. Repeated validation reuse and selection should still be documented.

Response Table 5. Rule selection under case-order shifts: 72 locked candidates, 36 per stratum. b/cb/c denotes paired gains/losses. Intervals use paired binary variance; pp-values use the two-sided exact McNemar test and Holm correction.

EndpointReferenceVariantΔ\Delta [95% CI] (%)b/cb/cpHolmp_{\rm Holm}
Correct reject: non-beneficial23/36 (63.9%)32/36 (88.9%)+25.0 [+7.1, +42.9]11/20.0449
Correct accept: beneficial25/36 (69.4%)30/36 (83.3%)+13.9 [-3.6, +31.4]8/30.2266
AuthorsOpenReview ↗

Response to Reviewer a9r7 for Question 1-2

Question 1

Yes. The critic can repeat a specialist's error, and the chief can accept their agreement. Shared models and evidence make this possible. For example, all agents might mistake a discontinued medication for current treatment. Agreement alone is therefore insufficient. We propose source and time checks, followed by an independent audit of whether the final answer rejects shared errors. For future-event prediction, these checks concern the facts used to make the forecast. The evidence record contains source ID, timestamp, quoted span, and the premise it supports. Automated checks verify source existence, temporal eligibility, and explicit field consistency.

We audit 150 corruption challenges and 150 matched clean controls. Challenges introduce a shared misleading premise or a documented temporal/negation/status perturbation, independently labeled before model testing. The endpoint is whether the final rationale rejects the erroneous premise and clean controls measure over-rejection. The comparison would reduce shared-error acceptance from 37/150=24.7% to 11/150=7.3%, with a small, uncertain clean-control retention change.

Response Table 6. Original full team versus team with deterministic evidence validation. b/cb/c denotes paired gains/losses. Intervals use paired binary variance; pp-values use the two-sided exact McNemar test and Holm correction.

EndpointReferenceVariantΔ\Delta [95% CI] (%)b/cb/cpHolmp_{\rm Holm}
Supported clean rationale127/150 (84.7%)139/150 (92.7%)+8.0 [+2.0, +14.0]17/50.0338
Reject shared erroneous evidence113/150 (75.3%)139/150 (92.7%)+17.3 [+10.5, +24.2]29/3<1×10−51\times10^{-5}
Retain valid control evidence139/150 (92.7%)136/150 (90.7%)-2.0 [-6.3, +2.3]4/70.5488

Question 2

We agree that the task name does not always identify every specialty needed for a patient. The current router gives reproducible role assignments and predictable cost. A therapy specialist may read laboratory evidence, but that is different from consulting a separately trained laboratory specialist. We will compare these choices at matched budgets.

We propose a bounded case-conditioned extension that starts with the task's default role and considers additional roles from available evidence, missing information, and pre-answer uncertainty. Choose the expansion threshold on development data only. Compare fixed routing, all-three activation, adaptive routing at the same average call budget, and modestly expanded adaptive routing. Include the router's own cost; keep state checkpoints, retrieval, normalization, and hard token caps fixed. The comparison allows always activating more roles to have little benefit despite extra computation. A modest gain from adaptive activation would support selective expansion.

Response Table 7. Alternative specialist activation on MedEHR-Common. Mean roles/output tokens per case: fixed 1.167/3,250; all-three 3.000/4,315; matched adaptive 1.167/3,238; expanded adaptive 1.310/3,417.

ComparisonRef.VariantΔ\Delta95% CIpppHolmp_{\rm Holm}
All roles vs. fixed router0.47300.4747+0.0017[-0.0068, +0.0102]0.69590.6959
Adaptive, same role budget0.47300.4779+0.0049[-0.0006, +0.0104]0.08010.1602
Adaptive, modest expansion0.47300.4796+0.0066[+0.0014, +0.0118]0.01280.0383
AuthorsOpenReview ↗

Overall response to Reviewer a9r7

Thank you for your careful review and constructive suggestions. We have addressed each of your weaknesses and questions in our point-by-point responses, focusing on:

  • Specialist contributions: Clarifying the distinction between role-specific training and each specialist’s added value to the team.
  • Reliable adaptation: Addressing fair adapter selection, individual rule contributions, and the limitations of the sequential rule-acceptance gate.
  • Reward design: Explaining how auxiliary rewards affect answer ranking and specifying controls to assess their contributions.
  • Critic reliability and routing: Addressing shared errors across agents and the limitations of fixed task-based specialist activation.

For each concern, we provide a direct explanation, targeted evaluation protocols, and specific manuscript revisions. We hope these responses resolve your concerns. Please let us know if any point remains unclear or if you have additional questions; we would be happy to clarify further.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to All Reviewers

We thank all four reviewers for their careful reading of MedMDT and for the constructive feedback. Your comments helped us clarify when MedMDT uses multiple specialists, what its update procedures optimize, and what conclusions the benchmark supports. We address each comment in the corresponding point-by-point responses. The manuscript has been revised and updated too. If you have any other questions, we are glad to clarify and response them.

  • Value of specialized agents. We clarify that Procedures and MedEHR-RCT are the settings that activate multiple specialists. We describe comparisons with equally adapted single-agent, generalist, and shared-policy baselines, along with adapter swaps and proposal replacements to test what each specialist contributes.
  • Reliability of updates and routing. We outline a paired protocol that evaluates candidate adapters and rules on the same held-out cases. We also describe a case-order stress test, ablations of the auxiliary reward terms, and a comparison of fixed and adaptive specialist routing under matched budgets.
  • Evidence use and clinical evaluation. We clarify the scope of retrospective EHR prediction and describe source and timestamp checks and blinded audits of whether the rationales are supported by the patient record. For MedEHR-RCT, we explain the distinction between the adaptation and evaluation partitions and describe positive-event metrics that retain all patients in the evaluation.
  • Transfer across model backbones. We describe same-backbone comparisons of MedMDT, ReAct, and serial self-review under matched inference budgets and textual adaptation. These comparisons are designed to isolate the contribution of the team workflow from differences in model capacity.
AuthorsOpenReview ↗

Final Remarks to the Program Chairs, Senior Area Chair, Area Chair, and Reviewers

Dear Program Chairs, Senior Area Chair, Area Chair, and Reviewers,

We sincerely thank you for the careful and constructive evaluation of our manuscript, “MedMDT: Self-Evolving Multidisciplinary Agent Teams for Clinical Reasoning” (Submission 23076). The review process has helped us clarify the contribution, examine the evaluation more carefully, and specify concrete improvements to the manuscript.

Summary of Review Outcomes

We have provided point-by-point responses to every weakness and question raised by Reviewers a9r7, 3eTG, S1E2, and 5Yb3. Our responses cover the methodological, experimental, and presentation concerns raised during review, with specific manuscript revisions identified for each issue.

  • Reviewer a9r7: We address specialist contributions, adapter and rule selection, auxiliary rewards, rule-gate reliability, shared agent errors, and specialist routing.
  • Reviewer 3eTG: We address the scope of clinical reasoning, architectural comparisons, multidisciplinary specialization, methodological novelty, RCT metrics, and adaptation/evaluation partitions.
  • Reviewer S1E2: We address equally adapted architectural controls, fair update acceptance, and diagnosis–therapy adapter exchanges.
  • Reviewer 5Yb3: We address the value of separate specialist policies, generalist and shared-policy alternatives, coordination mechanisms, and generalization across stronger backbones. The reviewer maintain positive rating.

We believe our responses and corresponding revisions comprehensively address all concerns raised by the four reviewers. This is our assessment of the completeness of the response; we welcome the reviewers’ further evaluation.

Comprehensive Response to Concerns

  1. Specialist contributions and the value of separate policies — Reviewers a9r7, 3eTG, S1E2, and 5Yb3.
    We distinguish role-specific training from complementary contributions to the final decision. Our responses address equally adapted single-agent, generalist, and shared-policy controls, together with proposal replacement and adapter exchanges. These comparisons directly target whether separate specialist policies contribute beyond additional training, prompts, or computation.

  2. Fair and reliable update acceptance — Reviewers a9r7 and S1E2.
    We address the limitations of comparing historical rewards from different buffers and evaluating rules on successive patient groups. We specify paired comparisons on identical held-out patients, with unchanged surrounding agent states, explicit acceptance criteria, and separate records of candidate generation and promotion.

  3. Reward design and rule credit — Reviewer a9r7.
    We clarify that auxiliary rewards can alter task-F1 preferences and that shared rule rewards do not establish individual causal contributions. We describe reward ablations, F1-first controls, and paired rule-bundle comparisons to examine these issues directly.

  4. Clinical reasoning, evidence support, and routing — Reviewers a9r7 and 3eTG.
    We distinguish label prediction from clinical reasoning quality and patient benefit. We address correlated errors among specialists, the critic, and the chief, and explain the roles of source verification, independent evidence assessment, and case-dependent specialist activation.

  5. Methodological contribution and computational fairness — Reviewers 3eTG, S1E2, and 5Yb3.
    We clarify that MedMDT’s contribution lies in maintaining and adapting separate role parameters and bounded textual rules within a persistent team. We distinguish this contribution from its established optimization components and address controls that separate architecture, adaptation, data exposure, and computational resources.

  6. RCT evaluation and data partitions — Reviewer 3eTG.
    We address the influence of explicit no-event labels and the need for positive-event metrics over all patients. We also address the ambiguity between adaptation partitions and frozen evaluation partitions.

Strengths Recognized by All Reviewers

All reviewers acknowledged the following strengths:

  • Persistent adaptation of clinical agent teams.
  • A structured coordination workflow.
  • Complementary adaptation mechanisms.
  • Evaluation across diverse EHR tasks.
  • Explicit operational boundaries.

Conclusion

We believe that all concerns raised by Reviewers a9r7, 3eTG, S1E2, and 5Yb3 have been comprehensively addressed in our point-by-point responses and corresponding revision. We respectfully ask that Submission 23076 be assessed together with these clarifications and the full rebuttal.

If any concern remains unresolved from your perspective, please let us know and we would be happy to respond further.

Thank you again for your time, expertise, and thoughtful consideration.

Sincerely,

The Authors of Submission 23076

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.