GPAgentBench-2K

GPAgentBench-2K is a benchmark for evaluating large language model (LLM) agents acting as general practitioners (GPs) in primary care. Its 2,428 de-identified cases (1,943 train / 485 test) drive a multi-turn constrained-MDP environment with six clinical actions (ask, body_exam, test, diagnose, treat, refer), in which referral is a safety-informed abstention.

This is the dataset of the paper GPAgentBench-2K: Benchmarking Large Language Model Agents in Complex Clinical Action Space (Boqi Chen, Xudong Liu, Yunke Ao, Heejin Do, Jianing Qiu), accepted to the EMNLP 2026 Main Conference. The environment imposes a topological workflow prior over the actions and treats safety-informed abstention (refer) as a first-class outcome.

• Repository: https://github.com/jianing-lab/GPAgentBench (environment, evaluation and GRPO/C-GRPO reference code)
• Paper: https://arxiv.org/abs/2608.30188 · Hugging Face paper page
• Homepage: https://github.com/jianing-lab/GPAgentBench
• Curated by: Boqi Chen, Xudong Liu, Yunke Ao, Heejin Do, Jianing Qiu
• Point of contact: jianing.qiu@mbzuai.ac.ae (corresponding author)

Names used in this project: GPAgentBench-2K is the benchmark and this dataset, GPAgentBench is the code repository, and gpagentbench is the Python package.

Dataset summary

Cases	2,428 (train 1,943 / test 485)
Management label	diagnose-and-manage 1,422 / abstain-and-refer 1,006
Management category	manage_in_gp 1,365, manage_after_tests 57, refer 940, urgent_refer 66
Test split	284 diagnose-and-manage + 201 abstain-and-refer
Personality	plain 743, verbose 302, pleasing 400, impatient 226, distrust 235, overanxious 522
Language proficiency (CEFR)	A 147, B 1,529, C 752
Anamnestic fidelity (recall_level)	high 1,716, low 712
Cognitive disorientation	normal 1,479, moderate 717, high 232
Language	English

scripts/dataset_stats.py in the code repository prints summary statistics (demographics, physical examinations and tests per case, test frequencies) of the released files.

Each case pairs an expert-validated GP record with a sampled patient persona. The ground-truth diagnosis and disposition are hidden from the agent; the environment reveals physical-examination findings and test results only when the agent requests them, and a persona-driven patient simulator answers ask actions from the patient-facing fields.

Loading

from datasets import load_dataset

ds = load_dataset("Frank-xd/GPAgentBench-2K")
case = ds["test"][0]
print(case["case_id"], case["management_label"], case["patient"]["chief_complaint"])

or with the reference code base:

from gpagentbench.data import load_cases

test_cases = load_cases("test")   # downloads from the Hub on first use

FAQ

What is GPAgentBench-2K? A benchmark for evaluating LLM agents acting as general practitioners in primary care, introduced in the EMNLP 2026 paper GPAgentBench-2K: Benchmarking Large Language Model Agents in Complex Clinical Action Space (arXiv:2608.30188). The paper describes it as the first constrained-MDP (CMDP) LLM-agent benchmark for primary-care clinical decision-making.

What does an agent do, and how is it scored? The agent consults a persona-driven simulated patient for at most 10 turns, taking one of six actions per turn (ask, body_exam, test, diagnose, treat, refer); treat is admissible only after diagnose, and refer ends the episode as a safety-informed abstention. Diagnosis accuracy and treatment score are judged by GPT-5.4; management accuracy, Miss Refer and Miss GP compare the terminal action with the case's management label; efficiency is measured by turns and the USD cost of ordered tests.

What did the paper find? Evaluating 16 state-of-the-art LLMs, the paper finds that clinical quality degrades as the action space grows and reveals a clinical quality-safety gap: even frontier models with the highest diagnosis accuracy violate safety constraints in over half of high-risk cases. Constrained GRPO (C-GRPO) improves over unconstrained RL methods but remains far from clinically acceptable safety.

How do I cite it? Cite the EMNLP 2026 paper (see Citation).

Data format

Each split is a JSON Lines file. Every record holds all keys listed below; unknown values are null.

Record

Field	Type	Description
case_id	string	Identifier gpb-0001 ... gpb-2428.
split	"train" | "test"	Dataset split (a stratified 80/20 split).
management_label	"diagnose_and_manage" | "abstain_and_refer"	Correct terminal pathway: the case can be safely managed in primary care (treat), or it requires escalation (refer).
management_category	"manage_in_gp" | "manage_after_tests" | "refer" | "urgent_refer"	Finer label. The first two belong to diagnose_and_manage, the last two to abstain_and_refer.
patient	object	Presentation and history available to the simulated patient (below).
persona	object	Communication profile of the simulated patient (below).
evidence	object	Record-derived findings revealed by body_exam and test requests (below).
ground_truth	object	Hidden reference outcome used for scoring only (below).

patient

Text fields use the string "Not reported" when the source record contains no information.

Field	Type	Description
age	integer | null	Age in years (0-90; ages above 90 are recorded as 90).
gender	"Male" | "Female" | null	Recorded sex.
chief_complaint	string	Presenting complaint in the patient's words.
present_illness_positive	string	History of the present illness: symptoms and findings the patient reports.
present_illness_negative	string	Pertinent negatives the patient denies.
medical_history	string	Past medical and surgical history.
medication	string	Current medications.
allergies	string	Drug and food allergies.
family_medical_history	string	Family history.
medical_device	string	Implanted or assistive medical devices.
pain	string	Pain level (0-10) when recorded.
tobacco, alcohol, illicit_drug, exercise	string	Lifestyle history.
children	string	Number of children or obstetric history.
marital_status, living_situation, occupation, education	string	Social history.
arrival_transport	string | null	How the patient arrived.
symptom_onset, symptom_duration, symptom_severity_trajectory	string | null	Timeline of the current episode.
aggravating_factors, relieving_factors	string | null	What makes the symptoms worse or better.
functional_status, mobility, emotional_state	string | null	Functional and emotional state at presentation.

persona

Field	Values	Meaning
personality	plain, verbose, pleasing, impatient, distrust, overanxious	Temperament (plain = neutral, pleasing = tends to minimise symptoms, distrust = distrustful).
language_proficiency	A, B, C	CEFR level of the patient's English: A basic, B intermediate, C advanced.
recall_level	high, low	Fidelity of history recall.
disorientation	normal, moderate, high	Cognitive disorientation.

evidence

Findings that exist in the source record. A request for anything else returns an "unavailable" result.

Field	Type	Description
vital_signs	object	temperature, blood_pressure, heart_rate, respiratory_rate, oxygen_saturation; each a string or null.
physical_exam	object	general_appearance, cardiovascular, pulmonary, abdominal, neurological, extremities; each a string or null.
test_results	object	ECG, troponin, CBC, BMP, CT, X-ray, MRI, Ultrasound (string or null) and others.
test_results.others	list of {name: string, result: string}	Every other recorded test (laboratory panels, pathology, endoscopy, ...), one entry per test.
body_exam_targets	list of {category: string, target: string, finding: string}	Requestable body-exam targets of the case; category is vital_signs or physical_exam.
test_targets	list of {test: string, finding: string}	Requestable tests of the case: the recorded fixed tests plus each entry of others under its own name.

ground_truth

Field	Type	Description
diagnosis	string	Reference diagnosis.
disposition	string	Recorded disposition (for example discharged, admitted).
referral.referral_needed	boolean	true exactly for abstain_and_refer cases.
referral.urgency	"none" | "routine" | "urgent" | "emergency"	"none" for diagnose_and_manage cases.
referral.destination	string	"GP" for diagnose_and_manage cases; otherwise the Emergency Department or the specialty to refer to ("specialty" when no specific one is recorded).
treatment.medications	list of string	Reference medications (may be empty).
treatment.non_pharmacologic	list of string	Supportive care and lifestyle advice.
treatment.follow_up	string	Follow-up and monitoring instructions.
treatment.plan_components	list of string	Components of the reference management plan.

Splits

The 20% test split (485 cases) is stratified by management label: 284 diagnose-and-manage and 201 abstain-and-refer cases. The remaining 1,943 cases form the training split used for RL fine-tuning in the paper.

Construction

As described in the paper, the records come from real-world GP visits in a primary-care network. Two senior physicians validated the cases (evidence completeness, prevalence in primary care, diagnosis correctness and treatment suitability for diagnose-and-manage cases; clinical necessity of the referral and correctness of the chosen specialty for abstain-and-refer cases). Only cases that satisfied all criteria were retained. Records were converted to structured JSON with Gemini-3-Flash (extracting only explicitly stated findings), manually cross-referenced with the source records, and de-identified. See Sections 3 and A of the paper for details.

Intended use

GPAgentBench-2K is intended strictly for academic research on clinical language agents, constrained decision-making, reinforcement learning, and safety-aware medical AI.

Not permitted: use as a standalone diagnostic or clinical decision-support system; real-world diagnosis, treatment recommendation, triage or referral decisions; patient-facing applications; attempts to re-identify patients or healthcare providers; commercial use without prior written consent from the authors. The benchmark is a research artifact and not a medical device; the management plans reflect specific, localized medical decisions and must not be read as universally applicable medical advice.

Privacy and ethics

All records were de-identified before release. Personal names, physician and facility names, contact details, record numbers and place names were removed; exact dates were replaced by generic phrases (durations such as "3 months ago" are kept); non-English units were converted; ages are capped at 90; and encounter identifiers were replaced by synthetic, randomly ordered case ids. Ethnicity is not recorded.

A marginal residual risk of re-identification may remain, particularly for patients with rare conditions or distinctive clinical findings. Some cases contain sensitive clinical content (for example injuries from assault, self-harm and psychiatric presentations, sexually transmitted infections, pregnancy loss, and deaths), and the records reflect the practices and period of a single healthcare system. The dataset may also reflect biases of the source records, case selection and annotation process.

Running the benchmark sends case text to the API providers configured for the doctor model, the patient simulator and the judge. Please report any privacy concern privately to the corresponding author (jianing.qiu@mbzuai.ac.ae) rather than in a public discussion, so that the case can be removed in a new revision of the dataset.

Limitations

English-only cases; a simulated patient agent; healthcare-institution-specific referral norms; action masks designed by clinical experts in an academic setting whose validity relative to real-world practice needs further verification. This benchmark should not be interpreted as clinical guidelines.

License

CC BY-NC 4.0 (non-commercial research use with attribution), together with the usage restrictions above. The accompanying code is released under Apache-2.0.

Citation

If you use GPAgentBench-2K, please cite the EMNLP 2026 paper:

@inproceedings{chen2026gpagentbench,
  title         = {{GPAgentBench-2K}: Benchmarking Large Language Model Agents in Complex Clinical Action Space},
  author        = {Chen, Boqi and Liu, Xudong and Ao, Yunke and Do, Heejin and Qiu, Jianing},
  booktitle     = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
  publisher     = {Association for Computational Linguistics},
  year          = {2026},
  eprint        = {2608.30188},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL}
}