文章背景与核心概要
临床试验编程是将研究方案转化为符合临床数据交换标准协会（CDISC）指南的分析就绪数据集的关键过程，但它长期以来一直是监管申报中的主要瓶颈。本文指出，标准的大语言模型（LLM）代码生成在此任务中会彻底失效。在对五个前沿模型进行的 11 次单次尝试中，没有一个模型能生成有效的主体级分析数据集。
为了解决这一问题，作者引入了 GxP-Agent，这是一个多智能体系统，它将监管工作流程构建为有向无环图（DAG）。通过将庞大的数据集生成过程分解为 15 个领域特定的工作节点，并辅以验证门、条件重试以及 pharmaverse 技能上下文，GxP-Agent 实现了前所未有的可靠性和合规性。该系统在 CDISC-Bench 基准测试中表现优异，最高可实现 100% 的结构匹配率，充分证明了将领域特定的监管流程知识直接编码到图拓扑中，是实现可靠且符合 GxP 规范的临床试验编程的关键催化剂。
GxP-Agent: Process-DAG Topology for Reliable Clinical Trial Programming with LLM Agents
GxP-Agent: Process-DAG Topology for Reliable Clinical Trial Programming with LLM Agents
Metadata
- arXiv ID: 2608.16890 [cs.AI]
- Authors: Jaime Yan
- Submitted: 13 May 2026
- Primary Subject: Artificial Intelligence (
cs.AI) - ACM Classes: I.2.11; J.3
- DOI: 10.48550/arXiv.2608.16890
Metadata
- arXiv ID: 2608.16890 [cs.AI]
- Authors: Jaime Yan
- Submitted: 13 May 2026
- Primary Subject: Artificial Intelligence (
cs.AI)- ACM Classes: I.2.11; J.3
- DOI: 10.48550/arXiv.2608.16890
Executive Summary
Clinical trial programming—the critical process of transforming study protocols into analysis-ready datasets compliant with Clinical Data Interchange Standards Consortium (CDISC) guidelines—remains a major bottleneck in regulatory submissions.
This paper demonstrates that standard large language model (LLM) code generation fails catastrophically at this task. Across 11 single-shot attempts using five frontier models, zero valid subject-level analysis datasets were produced. To resolve this, the author introduces GxP-Agent, a multi-agent system that structures regulatory workflows as a Directed Acyclic Graph (DAG). By breaking monolithic dataset generation down into 15 domain-specific worker nodes augmented with validation gates, conditional retries, and pharmaverse skill context, GxP-Agent achieves unprecedented reliability and compliance.
Executive Summary
Clinical trial programming—the critical process of transforming study protocols into analysis-ready datasets compliant with Clinical Data Interchange Standards Consortium (CDISC) guidelines—remains a major bottleneck in regulatory submissions.
This paper demonstrates that standard large language model (LLM) code generation fails catastrophically at this task. Across 11 single-shot attempts using five frontier models, zero valid subject-level analysis datasets were produced. To resolve this, the author introduces GxP-Agent, a multi-agent system that structures regulatory workflows as a Directed Acyclic Graph (DAG). By breaking monolithic dataset generation down into 15 domain-specific worker nodes augmented with validation gates, conditional retries, and pharmaverse skill context, GxP-Agent achieves unprecedented reliability and compliance.
Key Findings & Benchmark Performance
The author evaluates the approach using CDISC-Bench, a novel execution-based benchmark derived from the FDA pilot submission CDISCPilot01 (comprising 254 subjects and 49 ground-truth ADSL variables).
- Claude Sonnet 4.6 (with GxP-Agent): Achieved a 100% structural match (49/49 variables, 254 correct records) across three independent runs.
- Baselines:
- Best retrieval-augmented baseline: 59.2% structural match.
- All single-agent and flat multi-agent approaches: 0%.
- Model Generalization (GPT-4.1): While scoring 0% under every other architecture, GPT-4.1 achieved a 59.2% mean structural match when governed by the same Process-DAG topology.
- Generalizability to Adverse Events (ADAE): Tested on a 9-node branching DAG (55 variables, 1,191 records), GxP-Agent achieved a 100% structural match on the first attempt.
Key Findings & Benchmark Performance
The author evaluates the approach using CDISC-Bench, a novel execution-based benchmark derived from the FDA pilot submission
CDISCPilot01(comprising 254 subjects and 49 ground-truth ADSL variables).
- Claude Sonnet 4.6 (with GxP-Agent): Achieved a 100% structural match (49/49 variables, 254 correct records) across three independent runs.
- Baselines:
- Best retrieval-augmented baseline: 59.2% structural match.
- All single-agent and flat multi-agent approaches: 0%.
- Model Generalization (GPT-4.1): While scoring 0% under every other architecture, GPT-4.1 achieved a 59.2% mean structural match when governed by the same Process-DAG topology.
- Generalizability to Adverse Events (ADAE): Tested on a 9-node branching DAG (55 variables, 1,191 records), GxP-Agent achieved a 100% structural match on the first attempt.
Conclusion
The findings underscore that encoding domain-specific regulatory process knowledge directly into a graph topology—rather than depending solely on internal LLM reasoning—is a vital catalyst for achieving reliable, GxP-compliant clinical trial programming.
Conclusion
The findings underscore that encoding domain-specific regulatory process knowledge directly into a graph topology—rather than depending solely on internal LLM reasoning—is a vital catalyst for achieving reliable, GxP-compliant clinical trial programming.
Full-Text & Supplemental Links
Full-Text & Supplemental Links