GxP-Agent Uses Process-DAG Topology to Achieve 100% Accuracy in LLM-Based Clinical Trial Programming
GxP-Agent is an innovative multi-agent system designed to automate the programming of clinical trials, a task where leading large language models have struggled. This system converts study protocols into datasets that meet CDISC standards, which are essential for regulatory submissions. In 11 trials involving five different models, none succeeded in generating valid datasets. GxP-Agent employs a directed acyclic graph (DAG) to break down dataset creation into 15 specific nodes, managed by worker agents equipped with validation gates and retry options. When assessed on CDISC-Bench, based on the FDA's CDISCPilot01 with 254 subjects and 49 variables, GxP-Agent achieved a perfect structural match of 100% using Claude Sonnet 4.6, while the top baseline only attained 59.2% accuracy. The research is published on arXiv (identifier 2608.16890).
Key facts
- No LLM produced a valid subject-level analysis dataset across 11 single-shot attempts with five frontier models.
- GxP-Agent encodes regulatory process ordering as a directed acyclic graph (DAG).
- The system decomposes dataset generation into 15 domain-specific nodes.
- Worker agents execute nodes with pharmaverse skill context, validation gates, and conditional retry.
- CDISC-Bench is a new execution-based benchmark from the FDA pilot submission CDISCPilot01.
- CDISCPilot01 includes 254 subjects and 49 ground-truth ADSL variables.
- GxP-Agent with Claude Sonnet 4.6 achieved 100% structural match (49/49 variables, 254 correct records) across three runs.
- The best retrieval-augmented baseline achieved only 59.2% accuracy.
- The paper was announced on arXiv with ID 2608.16890.
Entities
Institutions
- arXiv
- Clinical Data Interchange Standards Consortium (CDISC)
- U.S. Food and Drug Administration (FDA)
- pharmaverse