PaperlayerView exact revision on arXiv

Source-linked AI summary

GPAgentBench-2K: Benchmarking Large Language Model Agents in Complex Clinical Action Space

Boqi Chen, Xudong Liu, Yunke Ao, Heejin Do, Jianing Qiu

arXiv:2608.30188v1cs.CL

TL;DR

Existing benchmarks simplify clinical workflows into static prediction or coarse, unconstrained MDPs. GPAgentBench-2K introduces a CMDP benchmark built from expert-validated GP records, modeling six actions and safety-informed abstention. Across 16 LLMs, clinical quality degrades as action spaces scale, while C-GRPO improves constrained optimization yet leaves substantial safety failures.

  • Problem

    Existing clinical-agent benchmarks use static prediction or coarse, unconstrained MDP formulations that do not capture complex primary-care workflows and safety requirements.

  • Method

    GPAgentBench-2K constructs a primary-care CMDP from expert-validated GP records, with six clinical actions, workflow action masking, multiple terminal outcomes, and safety costs.

  • Results

    Across 16 state-of-the-art LLMs, clinical quality degrades in complex action spaces, and frontier models violate safety constraints in over half of high-risk cases; C-GRPO improves over unconstrained RL but leaves high safety-failure rates.

  • Takeaways & Limitations

    Clinically acceptable safety in complex clinical action spaces remains an unsolved problem that GPAgentBench-2K is designed to evaluate and advance.

  • Takeaways & Limitations

    The benchmark uses a simulated patient agent, institution-specific referral norms, and expert-designed action masks whose validity against real-world practice requires further verification.

Abstract

from arXiv · show

Large Language Models (LLMs) show great potential as clinical agents, yet existing benchmarks reduce clinical workflows to static predictions or unconstrained Markov Decision Processes (MDPs) with coarse action sets. To address this, we introduce GPAgentBench-2K, the first Constrained MDP (CMDP) LLM-agent benchmark for primary-care clinical decision-making, constructed from expert-validated records of real-world GP encounters. Our environment models a full spectrum of six foundational clinical actions, imposes a topological workflow prior over the action space, and operationalizes safety-informed abstention as a first-class outcome. Evaluating 16 state-of-the-art LLMs reveals a significant performance degradation as the action space scales. Crucially, we uncover a clinical quality-safety gap: even frontier models with the highest diagnosis accuracy violate safety constraints in over half of high-risk cases. Finally, we establish a reference point using Constrained Group Relative Policy Optimization (C-GRPO), and show that while explicitly modeling constraints improves performance over unconstrained RL methods, it remains far from clinically acceptable safety.

1 Introduction

Existing clinical-agent benchmarks simplify workflows through static prediction or coarse, unconstrained MDPs. GPAgentBench-2K introduces a CMDP benchmark with six clinical actions and safety-informed abstention, revealing degraded quality and persistent safety failures in complex action spaces.

  • Existing benchmarks often assume comprehensive patient data upfront, reducing clinical decision-making to a single-step prediction task that bypasses active evidence gathering.
  • Conventional MDP environments improve interaction but typically use coarse action sets and unconstrained objectives that permit safety-accuracy tradeoffs.
  • GPAgentBench-2K models six foundational actions, expert-designed action masking, and terminal outcomes for both safe management and safety-informed referral.
  • 16 state-of-the-art LLMs show significantly degraded clinical quality in complex action spaces, while frontier models with the highest diagnostic accuracy violate safety constraints in over half of high-risk cases.
  • C-GRPO improves over unconstrained reinforcement learning baselines, but absolute safety-failure rates remain high in complex clinical action spaces.

2 Related Work

Prior clinical-agent work advances beyond static prediction through interactive MDPs, but remains limited by simplified action spaces and unconstrained terminal objectives. GPAgentBench-2K addresses both limitations with a full-spectrum CMDP formulation.

  • Non-MDP benchmarks treat diagnosis as static prediction, while some multi-agent systems divide workflows into narrow roles with oversimplified individual action spaces.
  • Its overview pairs expert-validated GP records with sampled patient personas and scores agent trajectories using diagnosis, treatment, safety, and diagnostic-cost terms.
  • MDP environments support active evidence gathering through patient dialogue or iterative testing, but commonly retain minimal action sets and a single diagnosis terminal outcome.
  • GPAgentBench-2K is presented as the first clinical-agent CMDP benchmark requiring full-spectrum action selection within strict safety boundaries.

3 Benchmark Construction

GPAgentBench-2K is constructed from real-world primary-care records and represents multiple valid management outcomes rather than forcing every case into diagnosis alone. Cases undergo physician validation and structured preprocessing before simulation.

  • The dataset uses real-world GP records spanning 9 primary departments and 62 secondary departments, including demographics, symptoms, history, examinations, tests, and management plans.
  • Cases support both diagnose-and-manage trajectories for safely manageable conditions and abstain-and-refer trajectories for cases requiring specialist escalation.
  • Two senior physicians validate retained cases for evidence completeness, prevalence, diagnostic correctness, treatment suitability, referral necessity, and specialty selection.
  • Clinical records are transformed into standardized JSON linking free-text descriptions to patient-simulation and clinical-observation facets.

4 Environment Design

The environment simulates primary-care interaction through persona-driven patients and a constrained six-action workflow. It scores clinical quality while separately tracking safety and efficiency costs, with physician-audited evaluation components.

  • 4.1 Persona-Driven Patient Simulator: The patient simulator incorporates personality, language, recall, and cognitive characteristics that shape what evidence an agent can elicit.
  • 4.2 Constrained Markov Decision Process: Figure 3 reports a monotonic decrease in diagnosis accuracy and treatment score as the parsed action set expands from two actions to six.
  • 4.1 Persona-Driven Patient Simulator: An independent physician audit evaluates simulator trajectories for factual consistency, clinical realism, and persona adherence, with strong inter-rater agreement.
  • 4.2 Constrained Markov Decision Process: The CMDP objective maximizes expected return subject to feasibility thresholds for safety and diagnostic-cost constraints.
  • 4.2 Constrained Markov Decision Process: The CMDP state records the initial patient presentation and accumulated dialogue, physical observations, and diagnostic results.
  • 4.2 Constrained Markov Decision Process: The six-action space is governed by expert-designed state-dependent action masks, enabling recursive evidence collection followed by diagnosis, treatment, or referral.
  • 4.2 Constrained Markov Decision Process: Rewards measure diagnosis, treatment, and management accuracy, while costs capture safety failures, dialogue turns, and diagnostic-test spending.

5 Experiments

Experiments show that expanding the clinical action space consistently reduces model quality and exposes a gap between diagnostic performance and safety. Constrained RL improves quality and safety relative to unconstrained RL, but substantial safety failures remain.

  • Clinical Quality Drops as Action Space Scales: Dx accuracy and Tx score decrease monotonically as the admissible action set expands from two actions to six.The test cases and patient personas remain fixed, and omitted-action evidence is provided upfront, isolating action-space complexity as the varying factor.
  • Clinical Quality Drops as Action Space Scales: Introducing refer causes the largest single-step Tx score drop, averaging 2.9 percentage points.This decline indicates difficulty calibrating safety-informed abstention, including when not to act.
  • Clinical Quality-Safety Gap: 70.0% Dx accuracy for Claude Opus 4.7 and 53.2% Tx score for GPT-5.4 coexist with missed required referrals exceeding half in high-risk cases.Claude Opus 4.7 and Sonnet 4.6 miss 67.2% and 77.1% of required referrals, respectively, while Kimi K2.6 over-refers and misses 71.8% of local-management cases.
  • Scaling Model Parameters: Scaling Qwen2.5 from 7B to 72B raises Dx Accuracy from 40.4% to 54.1% and Tx Score from 18.4% to 25.4%, but does not reliably improve management safety or cost.Qwen2.5 (72B) misses 54.7% of high-risk referrals and incurs the highest diagnostic cost, $473, through over-ordering tests.
  • Reinforcement Learning Results: C-GRPO improves Dx accuracy, Tx score, and management accuracy over GRPO by 11.8 pp, 2.1 pp, and 3.6 pp, respectively.It also requires 6.24 average turns and costs $435 in diagnostics, while improving both safety dimensions relative to GRPO.
  • Reinforcement Learning Results: C-GRPO reduces Miss-Refer to 82.1% and Miss-GP to 26.1%, yet still misses 82.1% of required referrals.The improvement is relative to baseline GRPO and does not close the absolute safety gap.
  • Case Study: GPAgent safely refers a high-risk patient after detecting an irregular rectal mass, whereas Qwen2.5-7B locally tests, diagnoses, and treats the cancer.Both agents uncover the same critical evidence; their trajectories differ at the safety-informed abstention boundary.
  • Step-Level Decision Quality: GPAgent-7B trained with C-GRPO duplicates questioning in 1.0% of turns, matching frontier-level behavior and improving over its Qwen2.5 base.DoctorAgent-RL (7B) duplicates known information in 36.7% of turns and reaches 48.9% when duplicated exams and tests are included.

6 Conclusion

GPAgentBench-2K models clinical agents in a complex, safety-aware action space and exposes substantial quality-safety failures. C-GRPO improves over unconstrained reinforcement learning, but clinically acceptable safety remains unresolved.

  • GPAgentBench-2K models six foundational clinical actions and treats safety-informed abstention as a first-class clinical outcome.
  • 16 LLMs show significant performance degradation as the clinical action space becomes more complex.
  • Frontier models with high diagnostic accuracy violate safety constraints in over half of high-risk cases, revealing a clinical quality-safety gap.
  • C-GRPO improves diagnostic accuracy and reduces the safety trade-off incurred by scalarized reinforcement learning, but the residual safety gap remains substantial.

Limitations

The benchmark’s conclusions are bounded by language, simulation, institutional, and expert-designed workflow assumptions. It is intended for research rather than direct clinical deployment.

  • The benchmark uses English-language clinical records, so transfer to other languages and healthcare systems remains an open question.
  • The benchmark uses a simulated patient agent and healthcare-institution-specific referral norms, limiting direct interpretation as clinical guidance.
  • Expert-designed action masks improve reproducibility and reduce hallucinated observations, but their validity relative to real-world practice requires further verification.

A.3 Examples of Processed Data

The processed examples demonstrate how unstructured clinical narratives are mapped into a structured schema while preserving encounter-specific examinations, tests, dispositions, referrals, treatments, and follow-up.

  • The extraction pipeline maps raw clinical records into predefined JSON outputs while filtering out explicit diagnostic labels.
  • A routine outpatient example records discharge without specialist referral and includes conservative, non-pharmacologic management with routine follow-up.
  • The structured schema includes demographics, vital signs, physical examination, diagnostic tests, referral details, treatments, and follow-up instructions.
  • An urgent example records hospital admission, abnormal urinary findings, and urgent referral to the Emergency Department.

A.4 Statistics of the Dataset

GPAgentBench-2K covers diverse primary-care cases, spanning general management and specialist referral across varied demographics, symptoms, examinations, tests, and clinical-agent formulations.

  • A.4 Statistics of the Dataset: 2,434 cases divide into 45.1% manageable by general practitioners and 54.9% requiring specialist referral.
  • A.4 Statistics of the Dataset: The dataset contains 756 standardized disease categories and a long-tail distribution of specialist referral destinations.
  • A.4 Statistics of the Dataset: Cases average 4.9 symptoms overall versus 2.2 in chief complaints, indicating that complete presentations are not usually articulated upfront.
  • A.4 Statistics of the Dataset: The most frequent symptoms include cough, dizziness, nausea, and fever, with gastrointestinal and systemic complaints prominent among the top 20.
  • A.4 Statistics of the Dataset: Patients average 4.1 physical examinations and 2.9 diagnostic tests drawn from 584 unique test items across four test types.
  • B Extended Related Work: Existing clinical-agent frameworks commonly simplify individual action spaces through static prediction, specialized roles, limited test interaction, or soft resource penalties.

C.2 Reward, Cost, and Evaluation Metrics

The environment evaluates diagnosis, treatment, terminal management, safety, and diagnostic efficiency using rewards and costs grounded in clinician-validated primary-care cases.

  • Terminal outcomes: Each case ends with either treat using a predicted diagnosis and treatment plan or refer, matched against a ground-truth management label.Cases are labeled as safely manageable in primary care or requiring specialist referral.
  • Clinical quality metrics: Diagnosis accuracy is a binary semantic score, with clinical equivalence allowing synonymous wording and minor specificity differences while penalizing vague diagnoses.GPT-5.4 serves as the automated judge for this score.
  • Clinical quality metrics: Treatment score uses episode-level values of 0, 0.5, or 1 for incorrect, partially aligned, or clinically equivalent treatment plans.The score compares predicted and ground-truth treatment plans with GPT-5.4.
  • Safety metrics: Management accuracy equals 1 when the terminal action matches whether primary-care management or specialist referral is appropriate, and 0 otherwise.During reinforcement learning, terminal-action correctness is driven by the safety cost rather than a separate reward.
  • Safety metrics: Safety cost is a binary indicator of incorrect terminal action, covering both missed referral and unnecessary escalation.Missed referral is a false negative for referral, whereas missed local management is a false positive.
  • Efficiency constraint: Diagnostic efficiency is modeled as a thresholded constraint on monetary test costs, while physical exams and vital-sign checks incur no cost.Laboratory and imaging tests receive monetary penalties based on a USD-grounded lookup table.

C.3 Automatic Judge Validation

The paper validates its automated evaluation and patient simulator with physician audits and independent judging, then describes constrained optimization and its remaining safety boundary.

  • Automatic judge validation: GPT-5.4 and physician ratings show high agreement, with Cohen’s κ of 0.92 for diagnosis accuracy and 0.89 for treatment score.Two senior physicians independently audited 200 outputs using the same rubric as the automated judge.
  • Automatic judge validation: All cross-judge score differences are below 1 percentage point, and model rankings remain unchanged when Claude Opus 4.7 replaces GPT-5.4.The comparison uses identical prompts and rubric for three frontier models.
  • Patient simulator validation: The simulator produced highly consistent and clinically plausible patient turns while reliably reflecting assigned challenging personas.The reported results also describe strong agreement across all audited dimensions.
  • Patient simulator validation: Physicians audited 32 simulator trajectories, with two trajectories per evaluated model contrasting standard and challenging personas.The audit assessed factual consistency, clinical realism, and persona adherence against source GP records.
  • Constrained optimization: C-GRPO combines policy masking, component-wise Lagrangian relaxation, and iterative primal-dual updates over safety and diagnostic-cost constraints.Masking restricts actions to the admissible set, while dual multipliers increase when empirical constraints are exceeded.
  • Constrained optimization: C-GRPO’s safety multiplier is capped at λ_safety ≤ p_mgmt = 6, so constraint satisfaction is not guaranteed and resulting safety levels must be reported directly.The cap prevents missed escalations from destabilizing training.

D.4.2 Specification of Intended Use for Created Artifacts

GPAGENTBENCH-2K is released as a research-only benchmark and codebase for studying primary-care clinical agents, constrained decision-making, and safety-aware evaluation.

  • Created artifacts: The release contains a CMDP benchmark environment, dataset, evaluation code, patient simulator, and reinforcement-learning scripts.These artifacts support primary-care clinical-agent decision-making research.
  • Intended use: The benchmark is intended for academic research on clinical language agents, constrained decision-making, reinforcement learning, and safety-aware medical AI.Its scope includes evidence gathering, diagnosis, treatment planning, and safety-informed referral.
  • Risks and boundaries: The benchmark and environment are not medical devices and must not be deployed as standalone clinical decision-support systems.High-stakes use requires further clinical validation, regulatory assessment, ethical review, and expert oversight.
  • Risks and boundaries: The dataset uses de-identified clinical records, but residual re-identification risk may remain for rare conditions or distinctive findings.The authors also note possible bias from source records, healthcare systems, case selection, and annotation.
  • Intended use: The codebase supports reproducible agent-environment interactions, structured-action parsing, and diagnostic, management, safety, and efficiency evaluation.The accompanying dataset remains restricted to non-commercial research use.
  • Scope: The benchmark targets primary-care and general-practice diagnosis, management, and referral from clinical presentations.Records, prompts, and interactions are in English and require structured XML-style action outputs.

D.7.6 Data Consent

The study describes physician consent and evaluation procedures for reviewing anonymized clinical cases and selecting records for the final dataset.

  • Consent: Participation was voluntary, with secure and confidential handling of responses and withdrawal permitted without penalty.The study information was designated for academic research only.
  • Case inclusion: Only cases receiving unanimous Pass ratings across individual criteria and unanimous Retain Case decisions were included in the final dataset.The criteria covered diagnosis correctness, evidence completeness, and disease prevalence.
Loading 2608.30188v1…