Can Multi-Agent LLMs Identify Their Peers? Stylometric Fingerprinting in Role-Constrained Political Analysis

This paper investigates the ability of multi-agent LLMs to identify their peers in political analysis texts despite prompt-level anonymization, revealing that stylometric fingerprints persist. The study evaluates three classifiers—Claude Sonnet 4.6, Llama-3.3-70B, and a fine-tuned T5-base model—on a five-class attribution task, with T5 achieving a Macro F1 score of 0.991 under a novel statement-disjoint cross-validation protocol. The findings underscore the inadequacy of anonymization for preventing model identity detection, which has significant implications for compliance with AI regulations and validation in multi-agent systems.