Role setting is a configuration technique that enables a model to speak with a specific identity, tone, and knowledge boundary through system prompts, character cards, or fine-tuning. It addresses not only the question of "realism" but also consistency: the same character must not contradict itself across multiple turns of dialogue, nor should it bypass its settings to answer questions it is not supposed to answer.

The early form of role setting was rule-based chatbots. In 1966, Weizenbaum demonstrated in the paper *ELIZA—A Computer Program For the Study of Natural Language Communication Between Man and Machine that "personality" could be shaped solely by dialogue rules, using pattern matching to simulate the tone of a psychotherapist. In the era of large models, role setting derives more from product practices and is typically written into system prompts, taking precedence over standard user inputs.

The first systematic benchmark for role setting is RoleLLM. In 2023, Wang et al. constructed character profiles for 100 roles in the paper *RoleLLM: Benchmarking, Eliciting, and Enhancing Role-Playing Abilities of Large Language Models, using Context-Instruct to extract role-specific knowledge, RoleGPT to mimic speaking styles, and RoCIT for role-conditioned instruction fine-tuning.

The evaluation vehicle for role setting is RoleBench—comprising 168,093 samples, it is the first systematic, fine-grained role-level benchmark. RoleLLaMA (English) and RoleGLM (Chinese), obtained through fine-tuning, significantly enhanced role-playing capabilities, even approaching GPT-4-driven methods. Role consistency is fundamental to AI companionship, game NPCs, brand personification, and educational simulations. However, excessive immersion may amplify biases or bypass safety boundaries, so role setting also requires evaluable standards and cannot rely solely on "feeling realistic."