Unleashing Cognitive Synergy In Large Language Models: A Task-solving Agent Through Multi-persona Self-collaboration

“A cognitive synergist denotes an intelligent agent that works in conjunction with several minds, merging their unique abilities and expertise to improve problem-solving and overall efficacy in intricate tasks. In this work, we aim to develop a cognitive synergist based on a single LLM that can "split into" multiple personas and engage in multi-persona self-collaboration to address both knowledge-intensive and reasoning-intensive tasks. The underlying biological intuition stems from the significance of pretend play and role-playing (Pellegrini, 2009) in a child’s cognitive development. According to Piaget’s developmental theory (Piaget, 1954), engaging in pretend play and taking on different roles allows children to cultivate essential skills such as problem-solving, critical thinking, empathy, and cooperation.

The main inspiration for this work originates from recent findings (Deshpande et al., 2023; Xu et al., 2023) suggesting that assigning personas to an LLM can elicit specific behaviors. For instance, Xu et al. (2023) demonstrates that when conditioned on a task-specific expert identity, an LLM can generate superior answers compared to having no assigned persona. Another closely related line of work Park et al. (2023); Schick et al. (2022); Li et al. (2023); Cai et al. (2023) hints at the possibility of constructing an AI society with multiple LLM agents collaborating in different roles. However, some lingering limitations of these previous works include: (1) personas are typically fixed or task-specific, necessitating human supervision; (2) such collaboration often requires multiple individual LLM instances, resulting in a doubling or tripling of inference costs.

To unleash the potential of cognitive synergy in LLMs, we propose Solo Performance Prompting (SPP), which prompts a single LLM to identify, simulate, and collaborate with multiple personas to solve challenging tasks. Figure 1 provides a high-level overview of SPP. Here, a persona can represent either a domain expert, such as a movie enthusiast, or a target audience, such as a ten-year old child. Through the dynamic identification of various personas, we empower a single LLM to acquire diverse domain knowledge accurately without additional retrieval systems. By facilitating multi-turn self-collaboration, we enable self-revision and self-feedback from various perspectives without requiring additional agents.”