Unleashing the Emergent Cognitive Synergy in Large Language Models: A Task-Solving Agent through Multi-Persona Self-Collaboration (2307.05300v4)

Abstract: Human intelligence thrives on cognitive synergy, where collaboration among different minds yield superior outcomes compared to isolated individuals. In this work, we propose Solo Performance Prompting (SPP), which transforms a single LLM into a cognitive synergist by engaging in multi-turn self-collaboration with multiple personas. A cognitive synergist is an intelligent agent that collaboratively combines multiple minds' strengths and knowledge to enhance problem-solving in complex tasks. By dynamically identifying and simulating different personas based on task inputs, SPP unleashes the potential of cognitive synergy in LLMs. Our in-depth analysis shows that assigning multiple fine-grained personas in LLMs improves problem-solving abilities compared to using a single or fixed number of personas. We evaluate SPP on three challenging tasks: Trivia Creative Writing, Codenames Collaborative, and Logic Grid Puzzle, encompassing both knowledge-intensive and reasoning-intensive types. Unlike previous works, such as Chain-of-Thought, that solely enhance the reasoning abilities in LLMs, experimental results demonstrate that SPP effectively reduces factual hallucination, and maintains strong reasoning capabilities. Additionally, comparative experiments show that cognitive synergy only emerges in GPT-4 and does not appear in less capable models, such as GPT-3.5-turbo and Llama2-13b-chat, which draws an interesting analogy to human development. Code, data, and prompts can be found at: https://github.com/MikeWangWZHL/Solo-Performance-Prompting.git.

• The paper introduces Solo Performance Prompting (SPP), enabling LLMs to internally simulate multi-persona collaboration for enhanced task-solving.
• The paper shows that SPP outperforms standard methods by reducing factual errors and boosting reasoning across diverse tasks such as logic puzzles and coding challenges.
• The paper reveals that emergent cognitive synergy is evident in advanced models like GPT-4, emphasizing the value of dynamic persona identification.

Unleashing Cognitive Synergy in LLMs

The paper "Unleashing the Emergent Cognitive Synergy in LLMs: A Task-Solving Agent through Multi-Persona Self-Collaboration" introduces a novel method called Solo Performance Prompting (SPP), which aims to enhance the task-solving abilities of LLMs by engaging them in self-collaboration with dynamically identified personas. This research addresses the persistent challenges LLMs face, such as factual hallucination and limited reasoning capabilities, by simulating multi-agent collaboration within a single model.

Key Contributions

1. Solo Performance Prompting (SPP): The authors propose SPP as a mechanism to transform LLMs into cognitive synergists. This involves the LLM adopting multiple personas and collaborating internally to solve complex tasks. The approach seeks to stimulate cognitive synergy, where collective interactions surpass individual efforts in problem-solving.
2. Evaluation Across Diverse Tasks: The effectiveness of SPP is evaluated on three challenging tasks: Trivia Creative Writing, Codenames Collaborative, and Logic Grid Puzzle. These tasks encompass both knowledge-intensive and reasoning-intensive challenges, offering a robust testing ground for the proposed method.
3. Comparative Analysis: Experimental results indicate that SPP outperforms standard prompting methods, notably reducing factual errors while maintaining reasoning capabilities. Unlike methods such as Chain-of-Thought (CoT), SPP achieves significant improvements in tasks requiring diverse information synthesis.
4. Emergent Ability in GPT-4: The paper highlights that cognitive synergy emerges prominently in GPT-4 but is absent in less capable models like GPT-3.5 and Llama2-13b, drawing parallels to developmental milestones in human cognition. This insight suggests that model capabilities are crucial for realizing the potential of multi-persona collaboration.
5. Dynamic Persona Identification: Unlike previous works that use fixed or task-specific personas, SPP dynamically identifies personas based on task inputs, demonstrating superior performance. This dynamic approach allows the model to adapt its knowledge acquisition strategies effectively.

Implications and Future Directions

The proposed SPP framework has significant implications for enhancing LLM performance across diverse applications, including creative industries and logic-based reasoning tasks. By facilitating internal collaboration within a single model, SPP reduces computational overhead compared to multi-agent systems and does not rely on external retrieval mechanisms, showcasing an efficient strategy for improving both knowledge accuracy and reasoning strength.

Future research could explore further optimizing persona identification and developing more sophisticated collaboration strategies to enhance performance. Additionally, extending SPP to integrate with external data sources might address tasks requiring up-to-date information.

In summary, this paper contributes important advancements to the field of LLMs by demonstrating that carefully crafted internal collaborations can elevate model performance to new heights. The findings emphasize the importance of dynamic, context-aware persona simulation, paving the way for more intelligent and adaptable AI systems.