Large language models contain hidden persona subnetworks that shape behavior without external fine-tuning, according to new research. The study shows that LLMs naturally develop distinct activation patterns for different personas—like introvert or extrovert—within their existing parameter space. Using small calibration datasets, researchers identified these patterns and developed a masking strategy to isolate lightweight persona subnetworks.

Traditional approaches rely on prompting, retrieval-augmented generation, or fine-tuning to adapt model behavior. This research challenges that assumption by demonstrating that persona-specific knowledge already exists in the model's weights. The team created a contrastive pruning strategy to enhance separation between opposing personas, such as introvert-extrovert pairs, by identifying parameters responsible for statistical divergence.

The method requires no training and outperforms baselines that depend on external knowledge while being more efficient. These findings suggest that diverse human-like behaviors aren't just induced in LLMs—they're already embedded in the parameter space. This points toward new possibilities for controllable and interpretable personalization in large language models.