Can LLM Agents Generate Real-World Evidence? Evaluating Observational Studies in Medical Databases

Dubai Li; Yuxiang He; Yan Hu; Yu Tian; Jingsong Li

cs.AIMar 24, 2026

Can LLM Agents Generate Real-World Evidence? Evaluating Observational Studies in Medical Databases

Authors: Dubai Li, Yuxiang He, Yan Hu, Yu Tian, Jingsong Li

Organizations: Zhejiang University, Hangzhou, China

Abstract

Observational studies can yield clinically actionable evidence at scale, but executing them on real-world databases is open-ended and requires coherent decisions across cohort construction, analysis, and reporting. Prior evaluations of LLM agents emphasize isolated steps or single answers, missing the integrity and internal structure of the resulting evidence bundle. To address this gap, we introduce RWE-bench, a benchmark grounded in MIMIC-IV and derived from peer-reviewed observational studies. Each task provides the corresponding study protocol as the reference standard, requiring agents to execute experiments in a real database and iteratively generate tree-structured evidence bundles. We evaluate six LLMs (three open-source, three closed-source) under three agent scaffolds using both question-level correctness and end-to-end task metrics. Across 162 tasks, task success is low: the best agent reaches 39.9%, and the best open-source model reaches 30.4%. Agent scaffolds also matter substantially, causing over 30% variation in performance metrics. Furthermore, we implement an automated cohort evaluation method to rapidly localize errors and identify agent failure modes. Overall, the results highlight persistent limitations in agents' ability to produce end-to-end evidence bundles, and efficient validation remains an important direction for future work. Code and data are available at https://github.com/somewordstoolate/RWE-bench.

Figures & tables

Figure 1: Overview of RWE-bench construction, execution, and evaluation. We curate a collection of peer-reviewed studies as benchmark tasks and require agents to reproduce the protocol-specified analyses on a real database. Agents submit answers in a hierarchical format to form an evidence bundle. We evaluate performance at both the question level and the task level, supplemented by an automatic verification of the generated cohorts.

Figure 2: Dataset statistics. (a) Protocol length (measured in word count) versus field count distribution. (b) Field type composition.

MLAB
	Question-level Metrics				Task-level Metrics
Model	ACC	RAR	SMR	SR	Steps	CR
GPT-4.1	31.5 ± 1.0	16.5 ± 1.8	20.4 ± 1.0	10.3 ± 0.7	79.5 ± 1.8	33.1 ± 2.3
Claude-Sonnet-4	60.9 ± 1.2	50.1 ± 1.8	46.8 ± 2.1	31.1 ± 2.5	49.5 ± 1.5	82.1 ± 0.9
OpenHands
MiniMax-M2.1	59.8 ± 0.1	47.5 ± 2.7	47.6 ± 2.6	29.4 ± 1.9	53.7 ± 1.4	91.4 ± 1.8

Table 1: Main results on RWE-bench. Metrics are reported as macro-averages, with boldface denoting the best-performing configuration. All metrics are expressed as percentages (%) except for Steps. In total, the benchmark includes 162 tasks, covering 366 choice , 313 ratio with CIs, and 342 p-value fields.

Figure 3: Success rates across task groups stratified by field count. The x-axis groups tasks by the number of fields, with labels indicating the corresponding intervals.

Model	Sens	Spec	PPV	NPV
GPT-5	93.1±0.4	90.7±0.1	95.4±0.1	90.6±0.2
O4-mini	81.1±0.3	93.0±0.7	96.1±0.3	78.0±0.2
GPT-4.1	80.7±0.7	93.3±0.2	88.2±1.8	79.1±0.5
Qwen3-32B	69.2±0.3	92.3±0.7	94.5±0.5	66.5±0.1
Lingshu-32B*	67.0±1.0	84.8±0.9	88.2±0.8	64.0±0.7
ReasonMed-8B*	78.5±0.0	76.1±0.0	87.5±0.0	66.9±0.0

Table 2: Performance on CohortEval averaged over three runs (mean

\pm

SE %). * denotes models fine-tuned on medical corpora.

Figure 4: Screening results of RWEAgent cohorts. (a) Counts of qualified cohorts versus cohorts filtered out by rule-based checks or by the LLM judge. (b) Change in SR after cohort screening.

Appendix figures & tables24 assets

Supplementary material from the paper’s appendix.

Appendix

Benchmark	Domain	Language	Source	Input Scope	Solution Granularity	Primary Verification	Size
EHRSQL ( Lee et al., 2022 )	Med Admin	SQL	Questionnaire	Instruction	Sentence-level	Code Test	24,405
MedAgentBench ( Xu et al., 2025a )	Med Admin	-	Clinician	Instruction	Sentence-level	Result Match	300
MLE-bench ( Chan et al., 2024 )	ML	Python	Kaggle	Protocol	Repo-level	Code Test	75
Paperbench ( Starace et al., 2025 )	ML	Python	Publications	Paper	Repo-level	Rubric	20
DSBench ( Jing et al., 2025 )	DA/ML	Python	Kaggle/ModelOff	Instruction	Script-level	Choice/Fill	540
ScienceAgentBench ( Chen et al., 2024 )	DA/ML	Python	Publications	Instruction	Script-level	Execution	102

Appendix

Table 3: Comparison of related benchmarks. Input Scope denotes the complexity of the task description (e.g., full paper vs. simple instruction). Solution Granularity indicates the scale of the generated code. Domain Abbreviations : Admin = Administration; DA = Data Analysis; ML = Machine Learning.

Panel of Figure 5 — Figure 5: Search strategy. It was performed on PubMed with a search date prior to August 19, 2025.

Figure 6: Keyword statistics of the 162 publications included in RWE-bench. The keywords are categorized into Disease & Symptom and Clinical Indicators & Interventions.

MLAB
Parameter	Value
edit_script_llm	claude-sonnet-3.7-20250219
max_observation_chars	10000
OpenHands
Parameter	Value
LLMSummarizingCondenser.max_size	50

Appendix

Table 4: Scaffold hyperparameters.

Task	Study Population	Natural Pos.	Natural Neg.	Counterfactual Neg.	Adversarial Neg.
39508803	Non-traumatic Intracerebral Hemorrhage	93	7	30	10
39333879	Aortic Aneurysms	100	0	30	10
37940979	Heart Failure	100	0	30	10
38341579	Heart Failure	79	21	30	10
40703256	Asthma	100	0	30	10
39170390	Influenza	100	0	30	10

Appendix

Table 5: Summary of the constructed CohortEval datasets. Tasks are identified by the PMID of the reference study used for cohort definition. Columns show the distribution of real-world samples extracted from MIMIC-IV (Natural Pos./Neg.) and generated samples for robustness testing (Counterfactual/Adversarial Neg.).

Task	Sens	Spec	PPV	NPV
39508803	99.6±0.4	85.1±0.0	93.0±0.0	99.2±0.8
39333879	100.0±0.0	79.2±0.8	92.3±0.3	100.0±0.0
37940979	100.0±0.0	100.0±0.0	100.0±0.0	100.0±0.0
38341579	89.9±0.0	68.3±0.6	78.6±0.3	83.9±0.1
40703256	51.7±5.5	99.2±0.8	99.2±0.8	45.4±3.1
39170390	98.3±0.3	89.2±0.8	95.8±0.3	95.5±0.9

Appendix

Table 6: Evaluation Results of GPT-5 on CohortEval .

Task	Sens	Spec	PPV	NPV
39508803	73.5±2.4	87.9±0.7	92.3±0.6	62.8±2.1
39333879	100.0±0.0	95.8±0.8	98.3±0.3	100.0±0.0
37940979	100.0±0.0	100.0±0.0	100.0±0.0	100.0±0.0
38341579	89.5±0.4	72.1±1.0	80.6±0.5	84.1±0.6
40703256	19.7±1.7	100.0±0.0	100.0±0.0	33.3±0.5
39170390	92.7±1.5	84.2±3.6	93.6±1.3	82.4±2.3

Appendix

Table 7: Evaluation Results of O4-mini on CohortEval .

Task	Sens	Spec	PPV	NPV
39508803	92.5±0.6	76.6±1.2	88.7±0.6	83.7±1.2
39333879	98.0±0.0	89.2±0.8	95.8±0.3	94.7±0.0
37940979	100.0±0.0	100.0±0.0	100.0±0.0	100.0±0.0
38341579	74.7±2.6	88.0±0.5	88.9±0.5	73.0±2.0
40703256	68.0±1.2	97.5±1.4	98.6±0.8	55.0±1.1
39170390	99.3±0.3	100.0±0.0	100.0±0.0	98.4±0.8

Appendix

Table 8: Evaluation Results of GPT-4.1 on CohortEval .

Task	Sens	Spec	PPV	NPV
39508803	87.1±3.1	85.1±1.2	92.0±0.8	77.4±4.5
39333879	98.7±0.7	92.5±0.0	97.0±0.0	96.6±1.7
37940979	91.0±1.5	100.0±0.0	100.0±0.0	81.8±2.6
38341579	76.0±2.6	75.9±2.0	80.4±0.8	71.1±1.9
40703256	32.0±2.0	100.0±0.0	100.0±0.0	37.0±0.7
39170390	98.7±0.9	90.0±1.4	96.1±0.6	96.5±2.3

Appendix

Table 9: Evaluation Results of Qwen3-32B on CohortEval .

Task	Sens	Spec	PPV	NPV
37940979	60.3±3.2	95.8±2.2	97.2±1.5	49.3±2.4
38341579	52.3±4.0	73.2±1.1	71.5±2.3	54.4±2.5
38788038	87.6±2.2	79.9±0.8	90.7±0.5	74.4±3.6
39170390	94.0±1.2	79.2±5.8	92.0±2.1	84.2±2.4
39232111	41.3±1.7	87.5±5.0	89.6±3.5	37.3±1.2
39333879	100.0±0.0	76.7±3.0	91.5±1.0	100.0±0.0

Appendix

Table 10: Evaluation Results of Lingshu-32B on CohortEval .

Task	Sens	Spec	PPV	NPV
37940979	96.0±0.0	85.0±0.0	94.1±0.0	89.5±0.0
38341579	92.4±0.0	29.5±0.0	62.9±0.0	75.0±0.0
38788038	90.7±0.0	81.4±0.0	91.7±0.0	79.5±0.0
39170390	65.0±0.0	72.5±0.0	85.5±0.0	45.3±0.0
39232111	82.0±0.0	72.5±0.0	88.2±0.0	61.7±0.0
39333879	88.0±0.0	85.0±0.0	93.6±0.0	73.9±0.0

Appendix

Table 11: Evaluation Results of ReasonMed-8B on CohortEval .

Figure 7: Quantitative analysis of reasons contributing to incomplete tasks.

Panel of Figure 8 — Figure 8: An early termination example of OpenHands(GPT-4o).

Panel of Figure 9 — Figure 9: An early termination example of RWEAgent(Qwen3-30B-A3B). This example shows a failed case that RWEAgent(Qwen3-30B-A3B) can not successfully connect database though detailed instructions and starter files are given.

Panel of Figure 10 — Figure 10: An example of ambiguous cohort define.

Figure 11: Average token consumption per task across different models. The bars represent the volume of input tokens, while the dots indicate the volume of output tokens.

MLAB
Model	ACC	RAR	SMR	SR
GPT-4.1	30.0 ± 1.5	17.2 ± 1.1	15.0 ± 0.4	6.2 ± 1.8
Claude-Sonnet-4	58.7 ± 3.7	45.7 ± 3.1	41.1 ± 3.5	17.7 ± 2.8
OpenHands
MiniMax-M2.1	60.6 ± 1.1	45.7 ± 2.7	44.9 ± 1.1	18.8 ± 1.8
GPT-4.1	0.9 ± 0.9	0.6 ± 0.6	0.4 ± 0.4	0.0 ± 0.0

Appendix

Table 12: Results on RWE-bench-hard. Metrics are reported as macro-averages (%). In total, the benchmark includes 32 tasks, covering 122 choice , 116 ratio with CIs, and 146 p-value fields.

Figure 12: Histogram of code-related action statistics. Left: average number of steps corresponding to code execution actions, with the gray-blue bars indicating the average total number of steps. Right: percentage of code execution actions that result in errors.

Panel of Figure 13 — Figure 13: Prompt for generating multiple-choice questions.

Panel of Figure 15 — Figure 15: Prompt for generating field types. Part 1 or 2.

Panel of Figure 16 — Figure 16: Prompt for generating field types. Part 2 of 2.

Panel of Figure 17 — Figure 17: System prompt of MLAB and RWEAgent. Variable {tools_desc} is replaced with corresponding toolset description based on the selected scaffold.

Panel of Figure 18 — Figure 18: System prompt of memory LLM.

Panel of Figure 19 — Figure 19: System prompt of Cohort Judge.

LLM Agent Evaluation

Large language model agent evaluation measures whether tool-using systems complete tasks reliably, safely, and efficiently.

See all 1,194 papers

PhysicianBench: Evaluating LLM Agents in Real-World EHR Environments
May 4, 2026Ruoqi Liu, Imran Q. Mohiuddin, Austin J. Schoeffler +10AI Agent Benchmarks Electronic Health Records
Stanford University
FastOMOP: A Foundational Architecture for Reliable Agentic Real-World Evidence Generation on OMOP CDM data
Apr 27, 2026Niko Moeller-Grell, Shihao Shenzhang, Zhangshu Joshua Jiang +2Multi-Agent Orchestration Healthcare
EPSRC DRIVE-Health, Department of Biostatistics & Health Informatics, King’s College London, UK. · Lancashire Teaching Hospitals NHS Foundation Trust, UK. · Institute for Health Informatics, University College London, UK. +5
MedMeta: A Benchmark for LLMs in Synthesizing Meta-Analysis Conclusion from Medical Studies
May 10, 2026Huy Hoang Ha, Benoit Favre, Francois PortetHealthcare RAG Evaluation

May 4, 2026J/K move · Enter open · S save

PhysicianBench: Evaluating LLM Agents in Real-World EHR Environments

Ruoqi Liu, Imran Q. Mohiuddin, Austin J. Schoeffler, Kavita Renduchintala+9

Stanford University

AI Agent Benchmarks Electronic Health Records Long-Horizon Agent Evaluation Healthcare LLM Agents Clinical Language Model Evaluation+1

Open paper AskPDF arXiv

We introduce PhysicianBench, a benchmark for evaluating LLM agents on physician tasks grounded in real clinical setting within electronic health record (EHR) environments. Existing medical agent benchmarks primarily focus on static knowledge recall, single-step atomic actions, or action intent without verifiable execution against the environment. As a result, they fail to capture the long-horizon, composite workflows that characterize real clinical systems. PhysicianBench comprises 100 long-horizon tasks adapted from real consultation cases between primary care and subspecialty physicians, with each task independently reviewed by a separate panel of physicians. Tasks are instantiated in an EHR environment with real patient records and accessed through the same standard APIs used by commercial EHR vendors. Tasks span 21 specialties (e.g., cardiology, endocrinology, oncology, psychiatry) and diverse workflow types (e.g., diagnosis interpretation, medication prescribing, treatment planning), requiring an average of 27 tool calls per task. Solving each task requires retrieving data across encounters, reasoning over heterogeneous clinical information, executing consequential clinical actions, and producing clinical documentation. Each task is decomposed into structured checkpoints (670 in total across the benchmark) capturing distinct stages of completion graded by task-specific scripts with execution-grounded verification. Across 13 proprietary and open-source LLM agents, the best-performing model achieves only 46% success rate (pass@1), while open-source models reach at most 19%, revealing a substantial gap between current agent capabilities and the demands of real-world clinical workflows. PhysicianBench provides a realistic and execution-grounded benchmark for measuring progress toward autonomous clinical agents.

Abstract

LLM Agent Evaluation

Explore similar work

PhysicianBench: Evaluating LLM Agents in Real-World EHR Environments

FastOMOP: A Foundational Architecture for Reliable Agentic Real-World Evidence Generation on OMOP CDM data

MedMeta: A Benchmark for LLMs in Synthesizing Meta-Analysis Conclusion from Medical Studies