Self-Role Prompting: A New Paradigm for Zero-Shot Reasoning Enabling Large Models to Autonomously Choose Thinking Roles

This article proposes an innovative zero-shot reasoning strategy—Self-Role Prompting—which enables large language models to autonomously select the most suitable reasoning role before solving a problem. This strategy uses a two-stage framework (role identification + role-driven reasoning) without the need for manually designed role templates, and has achieved significant results in mathematical reasoning (AQUA-RAT), commonsense question answering (CommonsenseQA), and strategic reasoning (StrategyQA) tasks, providing a new direction for prompt engineering.

Prompt engineering is a key technology to unlock the potential of large models, evolving from simple instructions to chain-of-thought prompting. Traditional methods require preset fixed roles, but different tasks need different reasoning perspectives, so preset roles may not be applicable. Self-Role Prompting proposes a flexible solution: allowing the model to autonomously select a suitable role without manual templates, which is a zero-shot strategy.

Self-Role Prompting uses a two-stage architecture:

1. Role Identification: The model analyzes task characteristics to generate a suitable role (e.g., choosing "mathematics professor" for math problems), with the prompt template: "Please analyze the problem and describe the most suitable expert role";
2. Role-Driven Reasoning: Solve the problem as the generated role, with the prompt template: "As [role], solve the problem and show your thinking process." This strategy is completely zero-shot and has the advantages of generality, adaptability, and scalability.

The study was validated on three benchmark tests:

• AQUA-RAT (Mathematical Reasoning): Outperforms standard zero-shot, activating mathematical reasoning abilities;
• CommonsenseQA (Commonsense Question Answering): Selects roles with life experience to better utilize common sense;
• StrategyQA (Strategic Reasoning): Analyzes from multiple perspectives to improve the accuracy of implicit reasoning. Compared with standard zero-shot, chain-of-thought prompting, and manual expert role prompting, Self-Role Prompting maintains the advantages of zero-shot and achieves better results.

1. Activation of Specific Knowledge: Choosing a role activates relevant subsets of knowledge and reasoning styles (e.g., a mathematics professor uses rigorous methods);
2. Reflection of Metacognitive Ability: The model evaluates task characteristics, adjusts reasoning strategies, and demonstrates self-regulation capabilities;
3. Optimization of Contextual Learning: Role generation constructs a context more suitable for the task, stimulating the model's potential.

Basic Process:

1. Design a role generation prompt;
2. Design a reasoning prompt;
3. Two-stage calling (first generate the role, then perform role-driven reasoning). Code Example: Python implementation of a two-stage calling function (role generation + reasoning). Optimization Tips: Control the length of role descriptions (1-2 sentences), adjust temperature parameters (higher when generating roles, lower when reasoning), and integrate multiple roles (synthesize results from multiple roles for complex problems).

Limitations:

• High computational cost (two-stage reasoning);
• Role selection may not be suitable for ambiguous/cross-domain tasks;
• Dependence on the capabilities of the base model. Future Directions:
• Adaptive role switching;
• Role library learning;
• Multimodal expansion;
• Integration with retrieval-augmented generation and tool usage.

Self-Role Prompting is an important advancement in prompt engineering, demonstrating the potential of models to improve reasoning abilities through autonomous decision-making. Its simplicity and effectiveness have practical value, without the need for complex examples or templates. In the future, we can expect to see more innovative applications and expansion optimizations in complex tasks.