Fine-Tuned Multi-Agent Framework for Detecting OCEAN in Life Narratives

Assessing personality traits from long-form text is difficult because traits are latent, context-dependent, and expressed through subtle linguistic cues. While Large Language Models (LLMs) can process extensive narratives, they often suffer from pretraining-induced biases and inconsistent inferences. This paper investigates whether a structured multi-agent framework—where specialized sub-agents are conditioned on psychometric definitions—can improve the accuracy and interpretability of personality detection compared to standard single-agent approaches.

The authors propose a hierarchical multi-agent architecture to predict the Big Five (OCEAN) personality traits. For each trait, three sub-agents are fine-tuned using Low-Rank Adaptation (LoRA) and a Masked Language Modeling (MLM) objective to adopt 'High', 'Low', or 'Neutral' perspectives. These sub-agents are grounded in IPIP-NEO facet-level keys, which provide specific behavioral definitions for each trait. During inference, each sub-agent analyzes the narrative and produces a binary decision, supporting evidence, and a confidence score. A judge LLM then aggregates these outputs to generate a final, holistic trait prediction.

The multi-agent framework achieved an average macro-F1 score improvement of approximately 8% over the best-performing single-agent baseline. The authors found that the combination of Mistral-7B sub-agents (for localized evidence extraction) and a Qwen-7B judge (for long-context aggregation) yielded the most robust performance. Ablation studies confirmed that the inclusion of IPIP-NEO facet keys and MLM fine-tuning were critical components, with the removal of facet keys causing the largest performance degradation. While the model excels at identifying extreme (High/Low) trait expressions, it remains less accurate at identifying neutral trait levels.