acceptodds
Under review as a conference paper at ICLR 2027

Learning to Evolve Clinical Harnesses Through Entity-Graph-Guided Orchestration for Clinical Decision Making

Abstract

Large language model based agents show promise for clinical decision making but struggle to maintain coherent patient states and coordinate execution across long-horizon workflows. Harness optimization can improve this execution infrastructure without updating model parameters, yet loosely structured clinical evidence and fixed proposal strategies limit its effectiveness. Thus, we propose a framework that learns to evolve clinical harnesses by coupling entity-graph-guided harness with a self-evolving proposer. To represent evolving clinical evidence coherently and explicitly, a shared patient entity graph explicitly represents evolving clinical evidence, diagnostic hypotheses, and their dependencies, coordinating memory retrieval, clinical-skill selection, and tool execution through a consistent, stage-visible patient state. To mitigate optimization collapse arising from a fixed proposer, we introduce a self-evolving proposer that learns how to evolve the harness via Candidate-DAG memory and an evolving proposer skill library. Candidate-DAG memory preserves alternative harness designs, component inheritance, and case-level corrections and regressions, including useful interventions from lower-scoring candidates. An evolving proposer skill library distills this experience into reusable procedures for parent selection, cross-branch component reuse, avoidance of unsuccessful interventions, and targeted acquisition of training-derived medical knowledge. Extensive experiments on MedChain and ClinicalBench datasets demonstrate that our framework outperforms general-purpose and medical-specialized LLMs, multi-agent systems, and existing harness optimization methods in average performance.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Clinical Decision Making, LLM Agents, Harness Optimization, Self-Evolving Agents, Graph-Based Reasoning

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_Bh8DOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors present a harness-optimization framework for sequential clinical decision making. A frozen Qwen3.5-9B agent is wrapped in a patient entity graph that coordinates memory retrieval, clinical skills and imaging tools, and an LLM proposer iteratively edits this harness using a Candidate-DAG history and a self-updating library of proposal skills.

Critical strengths

The stage-indexed entity graph is a sensible and inspectable design, and the case-level bookkeeping of corrections and regressions is a reasonable improvement over greedy, score-only harness search.

Critical weaknesses
  1. Section 3.1.2 states that during testing, skill performance is assessed with the test reference answers and the benchmark evaluator, and that this feedback is passed to the proposer, which decides whether to update Q(s). If test labels influence skill quality, skill eligibility or any later proposer decision, the reported test scores are not held-out estimates. The authors must clarify exactly when test labels are used and whether any test-derived signal reaches a harness that is later evaluated on the same test set. If such a path exists, all headline numbers should be recomputed without it. The leakage audit in step (2) of the evolution loop is carried out by an LLM, so it does not resolve this concern either.
  2. Gains are confounded with privileged access to training data, and key baselines are missing. The framework builds a global disease graph and a contrastive skill library from 8724 MedChain and 840 ClinicalBench training cases, plus 1,000 and 200 cases for evolution. Most baselines (zero-shot LLMs, MDAgents, MedAgent) appear to have no comparable access to this data. The ablation suggests the data matters more than the optimizer. The variant without the self-adapting proposer already exceeds both harness baselines, so much of the margin seems to come from the initial harness and its training-derived memory rather than from the proposed optimization procedure.
  3. The central claim about optimization collapse is not tested. The introduction argues that fixed proposers stagnate or regress and that a self-updating proposer avoids this, yet the paper contains no search-dynamics evidence. The single ablation row does not show that the claimed failure mode was present or that it was fixed. Proposer token usage, wall-clock time and monetary cost are also not reported. This matters because the method depends on a frontier proprietary model, and recent work the authors themselves cite stresses matched optimization budgets.
  4. Generalization is not established. Both benchmarks come from Chinese online or hospital sources. The harness is evolved separately for each benchmark, and memory, skills and evolution cohorts are drawn from the same distribution as the test set. In ClinicalBench, each of the 150 diseases has ten cases, so test diseases are almost certainly represented in the global disease graph, and retrieval may function as disease-level lookup rather than reasoning.
Questions for authors

See weaknesses above.

6 comments
AuthorsOpenReview ↗

Respond to Reviewer Bh8D (1/3)

We thank the reviewer for the careful assessment and for recognizing the inspectable design of our stage-indexed entity graph and the value in harness optimization.

W1: Clarifying data splits and evaluation independence

Response: We apologize for the inaccurate wording and clarify that test set has been strictly held out throughout harness evolution. In Section 3.1.2 line 268-269, "during testing'' refers to testing on validation set in each round of harness evolution, and "test reference answers'' should refer to validation-set reference answers, which provides feedback to determine whether to update Q(s)Q(s). These evaluations and updates are performed iteratively on the training and validation sets. For clarity, we would like to correct the sentence into: During harness evolution, skill performance is assessed on the validation set using the corresponding validation reference answers and the benchmark evaluator.

As correctly stated in Section 4.1 (Lines 407–408) of our original paper "Graph memory and clinical skills are initialized from training cases, while validation feedback guides candidate evaluation and updates to the Candidate-DAG memory and proposer skill library.'' For our experiments on the two datasets, we strictly split the data into three subsets (Table R1), including a training set to initialize memory and skills, a validation set to provide feedback for iterative harness evolution, and a held-out test set exclusively for final benchmark evaluation.

Table R1. Data splits.

DatasetTrainingValidationTest
MedChain8,7241,0002,439
ClinicalBench840200460

W2: Comparable training-data access and the contributions of initialization and optimization

Response: Thank you for raising this point. Regarding training-data access, our comparisons include memory-based agents (MedChain-Agent and MedAgent-Pro) that use the same training cases as our framework, as well as harness-optimization baselines (Meta-Harness and Self-Harness) with access to the same training data. Our method consistently outperforms these baselines on both benchmarks (Table R2), indicating that the gains cannot be attributed solely to privileged access to training data.

Table R2. Average scores of memory-based agents.

MethodMedChainClinicalBench
MedChain-Agent0.52000.5344
MedAgent-Pro0.49040.4917
Meta-Harness0.54960.5640
Self-Harness0.54980.5537
Ours0.59710.6099

Regarding the contribution of harness optimization, we clarify that "w/o Self-Evolving Proposer" still performs harness optimization. This variant removes Candidate-DAG memory and the proposer skill library while retaining a fixed proposer for harness optimization. To disentangle the contributions of initialization and optimization, we additionally report the performance of the initial harness before optimization in Table R3. The overall gain of 0.0593 over initialization (w/o Harness optimization) demonstrates that the improvements cannot be explained by the initial harness and training-derived memory alone.

Table R3. Initialization and optimization on MedChain.

ConfigurationAverage score
w/o Harness optimization0.5378
w/o Self-Evolving Proposer0.5742
Full framework0.5971
AuthorsOpenReview ↗

Respond to Reviewer Bh8D (2/3)

W3: Search dynamics and optimization costs

Response: Thank you for this comment. To examine optimization behavior of different methods in our experiments, we report the validation performance across evolution rounds in Table R4. Our proposer distills reusable optimization strategies from historical successes and failures to guide subsequent modifications. The resulting harness performance shows an overall upward trend. In comparison, the two baselines with fixed proposal procedures achieve smaller overall gains and remain at relatively low performance levels for extended periods. Similar conservative reversion has been reported in general agent-harness optimization [1, 2].

Table R4. MedChain validation scores across evolution rounds.

Method123456789101112131415
Ours0.55040.55570.55780.55250.56370.56590.56550.57180.57430.57590.58160.59180.59740.61380.6169
Meta-Harness0.54050.54440.54620.54510.54750.54800.55320.55170.55380.55760.56020.55590.55970.56080.5603
Self-Harness0.54100.53970.54050.54020.54360.54500.55580.55030.55420.55810.55670.56090.55920.56140.5610

Regarding optimization budgets, Table R5 reports the number of candidates and proposer LLM calls, token consumption, execution time per optimization round, and final average score. Our method requires moderately more proposer calls and tokens than Meta-Harness, but obtains a substantially higher score. Compared with Self-Harness, our method uses fewer LLM calls, fewer tokens, and much less execution time while achieving better performance. These results suggest that the proposed self-evolving proposer improves optimization effectiveness without introducing excessive computational overhead.

Table R5. Candidate generation and optimization costs per round.

MethodCandidatesProposer LLM callsTokensExecution timeMedChain average score
Ours332.866M38m 11s0.5971
Meta-Harness312.064M34m 30s0.5496
Self-Harness455.127M2h 09m 56s0.5498

[1] Ursekar, V., Shanker, A., Chatrath, V., Xue, Y., and Denton, S. M. (2026). VeRO: A Harness for Agents to Optimize Agents. ICML.

[2] Luo, X., Wang, F., Hu, C., Xue, D., and Deng, Y. (2026). Self-Evolving Agent Harnesses via Gated Semantic Quality-Diversity. arXiv:2607.13683v1.

AuthorsOpenReview ↗

Respond to Reviewer Bh8D (3/3)

W4: Generalization across clinical benchmarks and cross-benchmark adaptation

Response: To further assess generalization beyond separately optimizing the harness for each benchmark, we transfer the candidate with the highest MedChain performance to ClinicalBench. We rebuild its clinical memory and skills using ClinicalBench training cases and perform one additional evolution round on the ClinicalBench validation set to adapt to the target tasks and scoring protocol. We then freeze the adapted harness and evaluate it on the 460 clinicalbench test cases. It achieves an average score of 0.5778 (Table R6). This score exceeds those of Meta-Harness (0.5640) and Self-Harness (0.5537) on ClinicalBench, supporting our framework’s ability to generalize across clinical benchmarks.

Table R6. MedChain-to-ClinicalBench adaptation with one target-side evolution round.

FrameworkAccDIFRPDDBDDFDPTTPImaging DiagnosisCDRAcceptabilityAvg.
Meta-Harness0.70210.99890.84560.46170.34110.63910.30840.29020.48850.44340.16760.5640
Self-Harness0.65260.99890.84010.46380.33430.60810.31010.28500.49080.38460.14490.5537
MedChain → ClinicalBench0.71300.99570.86520.46680.33580.68910.30910.30090.52420.48260.18690.5778

Regarding disease overlap and suggesting broader evaluation of generalization, although training and test cases may share disease categories, patients with the same disease can differ substantially in their symptoms, examination findings, diagnostic evidence, and treatment needs. Making accurate clinical decisions therefore requires integrating patient-specific evidence with relevant medical knowledge, rather than simply retrieving disease-level information. Our graph-based memory is designed to provide structured medical knowledge and associations derived from training cases, enabling the harness to perform case-specific clinical reasoning for diagnosis, differential diagnosis, and treatment planning. We acknowledge that evaluation on clinical datasets from outside China would further strengthen the evidence of generalization and consider this an important direction for future work.

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

This paper proposes a framework for improving long-horizon clinical decision making through entity-graph-guided harness orchestration and a self-evolving proposer. The entity graph maintains a structured patient state to coordinate memory retrieval, clinical skills, and tool execution. Experiments on MedChain and ClinicalBench show improved average performance over general-purpose LLMs, medical multi-agent systems, and existing harness optimization methods.

Critical strengths
  1. The paper addresses two relevant challenges in long-horizon clinical decision making: maintaining consistent patient states across clinical stages and overcoming stagnation in automatic harness optimization. The proposed solutions are reasonably aligned with these challenges.

  2. The idea of combining Candidate-DAG memory with an evolving proposer skill library is interesting for harness optimization. In particular, preserving useful modifications from lower-scoring candidates and converting past optimization experience into reusable strategies are potentially valuable design choices.

  3. The proposed framework achieves the highest average performance among the evaluated methods on both MedChain and ClinicalBench, which illustrate the effectiveness of the proposed method.

Critical weaknesses
  1. Section 3.1.2 states that skill performance is evaluated using test reference answers and that the resulting feedback is passed to the proposer to update skill quality. This raises concerns about whether test-set information influences the optimization process. The authors should clarify how test feedback is handled and whether the final evaluation remains independent of harness optimization.

  2. The current experiments only remove the entire Self-Evolving Proposer, without separately evaluating its two key components: Candidate-DAG memory and the evolving proposer skill library. Additional ablations are needed to establish their individual contributions and determine whether both components are necessary for the reported improvements.

  3. Although the paper describes how the proposer learns from previous candidates, it provides limited concrete evidence of this process. Representative examples would make the proposed mechanism more convincing.

  4. The paper uses GPT-5.6-Sol in Codex as its proposer but does not compare against directly using Codex or other general-purpose coding agents to optimize the same initial harness. Moreover, the paper lacks a detailed analysis of harness optimization costs, such as LLM calls, token consumption, and execution time, making it difficult to assess the efficiency of the proposed approach.

  5. The existing ablations do not establish whether the graph structure itself is responsible for these improvements. A comparison with non-graph memory using equivalent clinical information would better demonstrate the advantages of the proposed entity-graph representation.

Questions for authors

See weakness.

6 comments
AuthorsOpenReview ↗

Respond to Reviewer sqny (1/2)

We thank the reviewer for recognizing the long-horizon clinical decision-making challenges we address, the value of combining Candidate-DAG memory with an evolving proposer skill library, and the effectiveness of our framework on both MedChain and ClinicalBench benchmarks.

W1: Clarifying data splits and evaluation independence

Response: We apologize for the inaccurate wording and clarify that test set has been strictly held out throughout harness evolution. In Section 3.1.2 line 268-269, "during testing'' refers to testing on validation set in each round of harness evolution, and "test reference answers'' should refer to validation-set reference answers, which provides feedback to determine whether to update Q(s)Q(s). These evaluations and updates are performed iteratively on the training and validation sets. For clarity, we would like to correct the sentence into: During harness evolution, skill performance is assessed on the validation set using the corresponding validation reference answers and the benchmark evaluator.

As correctly stated in Section 4.1 (Lines 407–408) of our original paper "Graph memory and clinical skills are initialized from training cases, while validation feedback guides candidate evaluation and updates to the Candidate-DAG memory and proposer skill library.'' For our experiments on the two datasets, we strictly split the data into three subsets (Table R1), including a training set to initialize memory and skills, a validation set to provide feedback for iterative harness evolution, and a held-out test set exclusively for final benchmark evaluation.

Table R1. Data splits.

DatasetTrainingValidationTest
MedChain8,7241,0002,439
ClinicalBench840200460

W2: Proposer component ablations

Response: Thank you for raising this point. To assess the contribution of each self-evolving proposer component, we ablate the Candidate-DAG memory and the proposer skill library separately. As shown in Table R2, removing either component reduces performance, confirming that both contribute to the effectiveness of our method.

Table R2. Proposer ablations on MedChain.

Proposer configurationAverage score
Ours0.5971
w/o Candidate-DAG memory0.5850
w/o Proposer Skill Library0.5806
w/o both0.5742

W3: Concrete examples of learning from candidate trajectories and evolving proposer skills

Response: Thank you for this suggestion. To provide more concrete evidence of how the proposer learns from previous candidates, we add a representative example in Table R3. The table traces the evolution from CtC_t to Ct+1C_{t+1}, showing the corresponding performance change and the proposer skill updated after each successful transition. This example illustrates that the proposer does not merely select higher-scoring candidates, instead, it distills case-level corrections and regressions into reusable proposal skills, such as tightening skill eligibility conditions, replacing misleading symptom-based retrieval with stage- and intent-aware distinctions, and retiring overly broad skills while preserving their validated subpatterns.

Table R3. Proposer evolution with skills

Candidate transitionPerformance changeUpdated proposer skill
C0 → C10.5504 → 0.5557Expand training-derived entity-skill coverage only when current-case evidence satisfies the skill conditions.
C1 → C20.5557 → 0.5578When symptom-similar retrieval is misleading, build the skill around the disease stage and care intent that actually distinguish the routes.
C2 → C30.5578 → 0.5637Retire a regressive broad symptom skill and preserve only its train-supported subpattern with explicit competing-route negatives.
AuthorsOpenReview ↗

Respond to Reviewer sqny (2/2)

W4: 1) Comparison across proposer models and 2) analysis of optimization costs

Response: 1) To compare with other general-purpose coding agents, we use Opus 5 in Claude Code as the proposer in our framework. As shown in Table R4, Opus 5 achieves performance close to GPT-5.6-Sol, supporting the robustness of our framework across these two proposer configurations.

Table R4. Comparison of proposer configurations on MedChain.

Coding agentProposer modelAverage score
CodexGPT-5.6-Sol0.5971
Claude CodeOpus 50.5926

2) We also add a cost analysis of the harness optimization process. Table R5 reports the number of candidates and proposer LLM calls, token consumption, execution time per optimization round, and final average score. Our method requires moderately more proposer calls and tokens than Meta-Harness, but obtains a substantially higher score. Compared with Self-Harness, our method uses fewer LLM calls, fewer tokens, and much less execution time while achieving better performance. These results suggest that the proposed self-evolving proposer improves optimization effectiveness without introducing excessive computational overhead.

Table R5. Optimization costs per round.

MethodCandidatesProposer LLM callsTokensExecution timeMedChain average score
Ours332.866M38m 11s0.5971
Meta-Harness312.064M34m 30s0.5496
Self-Harness455.127M2h 09m 56s0.5498

W5: Isolating the contribution of graph structure with equivalent clinical information

Response: To analyse the effectiveness of graph structure in our method, we replace the entity-graph memory with the non-graph memory containing the same clinical entities and evidence, but without explicit relational edges. In Table R6, average score drops from 0.5971 to 0.5717, indicating that the graph structure itself contributes to performance by preserving clinical relations and dependencies.

Table R6. Graph-structure ablation on MedChain.

Memory representationAverage score
Entity-graph memory0.5971
Non-graph memory0.5717

Thank you for your insightful comments. As the discussion period is coming to an end, we would like to kindly check whether our rebuttal has addressed your concerns, particularly regarding the additional proposer-component ablations and the comparison with non-graph memory. If there are any remaining questions, we would be happy to clarify them promptly.

Comments on the paper

5 comments
AuthorsOpenReview ↗

Respond to Reviewer 3up3

We thank the reviewer for recognizing the value of our entity graph in coordinating sequential clinical workflows, the benefits of retaining alternative candidate branches and distilling case-level experience into reusable proposal strategies.

W1 & Q1 Clarifying data splits and test protocol

Response: We apologize for the inaccurate wording and clarify that test set has been strictly held out throughout harness evolution. In Section 3.1.2 line 268-269, "during testing'' refers to testing on validation set in each round of harness evolution, and "test reference answers'' should refer to validation-set reference answers, which provides feedback for the proposer to determine whether to update Q(s)Q(s). These evaluations and updates are performed iteratively on the training and validation sets. For clarity, we would like to correct the sentence into: During harness evolution, skill performance is assessed on the validation set using the corresponding validation reference answers and the benchmark evaluator.

As correctly stated in Section 4.1 (Lines 407–408) of our original paper "Graph memory and clinical skills are initialized from training cases, while validation feedback guides candidate evaluation and updates to the Candidate-DAG memory and proposer skill library.'' For our experiments on the two datasets, we strictly split the data into three subsets (Table R1), including a training set to initialize memory and skills, a validation set to provide feedback for iterative harness evolution, and a held-out test set exclusively for final benchmark evaluation. Before the final evaluation on the held-out test set, the candidate with the highest validation score is selected as the final harness and frozen. Test reference answers and evaluator outputs are used only to compute final scores and they are never fed back to the proposer or used to update Q(s)Q(s) or any other harness component.

Table R1. Data splits.

DatasetTrainingValidationTest
MedChain8,7241,0002,439
ClinicalBench840200460

W2.1 & Q2 Optimization budget

Response: Thank you for the comment. To analyse the harness optimization budget, all methods run 15 rounds on the same data splits and Table R2 reports the number of candidates and proposer LLM calls, token consumption, execution time per optimization round, and final average score. Both our method and Meta-Harness use GPT-5.6-Sol as proposer, while Self-Harness uses its underlying harness model (Qwen3.5-9B) following its original design for self-optimization. Per round, our method requires three LLM calls for specification, auditing, and implementation, while Meta-Harness uses one direct call. Self-Harness requires five calls including one for error diagnosis and four for generating optimization routes.

Our method requires moderately more proposer calls and tokens than Meta-Harness, but obtains a substantially higher score. Compared with Self-Harness, our method uses fewer LLM calls, fewer tokens, and much less execution time while achieving better performance. These results suggest that the proposed self-evolving proposer improves optimization effectiveness without introducing excessive computational overhead.

Table R2. Optimization costs per round.

MethodCandidatesProposer LLM callsTokensExecution timeMedChain average score
Ours332.866M38m 11s0.5971
Meta-Harness312.064M34m 30s0.5496
Self-Harness455.127M2h 09m 56s0.5498

W2.2–2.3 & Q3 mechanism attribution: 1) Ablation studies on self-evolving proposer components; 2) Effectiveness of graph structure.

Response: Thank you for the suggestion. 1) To assess the contribution of each self-evolving proposer component, we conduct ablations by removing the Candidate-DAG memory and the Proposer Skill Library individually and jointly, with the latter resulting in a fixed proposer. As shown in Table R3, removing either component reduces performance, confirming that both contribute to the effectiveness of our method.

Table R3. Proposer ablations on MedChain.

Proposer configurationAverage score
Ours0.5971
w/o Candidate-DAG memory0.5850
w/o Proposer Skill Library0.5806
w/o both0.5742

2) To analyse the effectiveness of graph structure in our method, we replace the entity-graph memory with the non-graph memory containing the same clinical entities and evidence, but without explicit relational edges. In Table R4, average score drops from 0.5971 to 0.5717, indicating that the graph structure itself contributes to performance by preserving clinical relations and dependencies.

Table R4. Graph-structure ablation on MedChain.

Memory representationAverage score
Entity-graph memory0.5971
Non-graph memory0.5717

Thank you for your insightful comments. As the discussion period is coming to an end, we would like to kindly check whether our rebuttal has addressed your concerns about the test protocol and optimization budgets, as you noted that these clarifications could materially change your rating. If there are any remaining questions, we would be happy to clarify them promptly.

We would like to draw your attention to our follow-up discussion with Reviewer Bh8D, who acknowledged our clarified evaluation protocol. These clarifications are closely related to the evaluation-independence questions you highlighted as important to your assessment. We would greatly appreciate your perspective on these clarifications and any remaining concerns, and would be happy to respond promptly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.