# Agent Role Structure and Operating Characteristics in Large Language Model Clinical Classification: A Comparative Study of Specialist and Deliberative Multi-Agent Protocols

Preprint · Anderson C

- Archive id: doi:10.64898/2026.02.22.26346818
- Published: 2026-02-24
- Venue: medRxiv
- DOI: https://doi.org/10.64898/2026.02.22.26346818
- Citations: 0
- Source URL: https://doi.org/10.64898/2026.02.22.26346818
- Provenance: europepmc

## Abstract

Large language models (LLMs) are increasingly considered in clinical decision support, yet the architectural effects of role decomposition within multi-agent systems remain poorly isolated. Prior comparisons of single-agent and multi-agent prompting often confound workflow structure with changes in model configuration, training, or decoding. We present a controlled study evaluating whether specialist and generalist multi-agent protocols produce different operating characteristics under fixed inference conditions. Two deterministic protocols, Generic Deliberative (GD) and Feature-Specialist (FS), are evaluated under identical base weights, decoding settings, computational budget, and adjudication logic. The primary comparison isolates role structure and information routing, while feature-assignment and model-scale ablations evaluate boundary conditions. On the UCI Cleveland Heart Disease and Pima Indians Diabetes benchmarks, altering role structure produces substantial operating-characteristic shifts. Under the primary configuration ( llama3.1:8b ), Cleveland FS achieves higher accuracy (0.72 vs. 0.65) and macro-F1 (0.71 vs. 0.65), with a specificity-favoring operating point, whereas Pima GD achieves higher accuracy (0.68 vs. 0.51) and macro-F1 (0.64 vs. 0.49), while FS induces pronounced class asymmetry. Feature-pair ablations show that specialist feature assignment is a design lever, with FS accuracy ranging from 0.41 to 0.72. A model-scale ablation ( qwen2.5:14b ) shows that protocol-level effects interact with base-model capacity. Overall, internal role decomposition should not be treated as safety-neutral: it can act as a feature-conditioned architectural bias that changes evidence aggregation and shifts class-wise error profiles, indicating that prompt structure and architectural configuration should be explicitly tuned, validated, and reported in safety-sensitive LLM-based clinical decision systems.

_Abstract reproduced under CC-BY (article license as recorded by Europe PMC; abstract reproduced unmodified apart from markup removal)._

Topics: [AI in Clinical Medicine](https://space-kitty.com/topics/ai-medicine)

---
Space Kitty Archive · https://space-kitty.com · API: https://space-kitty.com/api/v1 · MCP: https://space-kitty.com/mcp · Premium: $0.20/request via x402 (Base Sepolia (testnet)). Metadata: arXiv (CC0), OpenAlex (CC0), Europe PMC (CC BY/CC0 works), ClinicalTrials.gov.