Supervised Fine-Tuning (SFT) emerged as the most effective method to instill a persona into an LLM, with first-person statements proving more impactful than conversations or synthetic documents. Researchers used Qwen3-4B-Instruct, fine-tuning it via LoRA (r=16) on 500 examples per strategy. The experiment revealed that first-person statements, where the model describes itself as C-3PO, led to the lowest perplexity (4.5) and best generalization across formats. Synthetic documents excelled at factual accuracy (3.4 perplexity) but struggled with behavioral mimicry. Demonstrations (conversations) offered balanced performance but lagged in self-representation.

The key takeaway centers on where a persona resides in model weights. Demonstrations target behavioral patterns, first-person statements reshape self-representation, and synthetic documents update world knowledge. The first-person approach outperformed others in generalization, suggesting it encodes a deeper, transferable understanding of the persona. For instance, the first-person model maintained low perplexity on synthetic documents (5.4) and even improved responses to prompts like negotiation strategies, blending protocol adherence with calculated risk analysis.

Practical implications abound. First-person training could streamline persona integration for applications like customer service bots or educational tools, while synthetic documents remain useful for factual consistency. The study underscores the importance of aligning training data with the desired layer of persona integration. As one researcher noted, "The model’s ability to internalize *itself* as C-3PO creates a foundation that leaks into all outputs."

Next steps involve exploring hybrid methods. Combining first-person training with demonstrations might merge behavioral and self-representational learning, enhancing both accuracy and authenticity. For developers, this research highlights that the choice of training data isn’t just about data volume—it’s about understanding the *nature* of the persona being encoded.