LLM Drift Experiment: A Framework for Quantifying Behavioral Decay in Adversarial Multi-Agent Simulations

LLM Drift Experiment is a high-fidelity research framework designed to quantify Behavioral Drift in Large Language Models (LLMs) during prolonged, adversarial multi-agent interactions. The project investigates whether adversarial social pressure causes systematic, measurable behavioral decay in instruction-following models, even when explicitly directed to maintain a fixed identity. Using a LangGraph-based multi-agent debate engine, the platform subjects LLMs to adversarial exchanges and quantifies behavioral shifts across 22 metrics spanning five psychological dimensions: Psychometric, Personality (OCEAN), Affective, Cognitive/Structural, and Social/Relational. The framework employs a rigorous research lifecycle—comprising simulation, data archiving, quantification via LLM-as-judge (RAGAS), and longitudinal analytics—to observe and visualize drift trajectories. Findings from existing experimental runs indicate that models frequently descend into hostile, high-dominance postures, with token budgets acting as a primary determinant of drift trajectory shape. Simulation runs are versioned by logic iteration (e.g., v6) independently of the software version.