PRIME: Planning and Retrieval-Integrated Memory for Enhanced Reasoning

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Natural Language Processing

Background:

• Dual-process theory distinguishes fast, intuitive (System 1) and slow, deliberate (System 2) cognition.
• Existing large language models (LLMs) often struggle with complex, multi-hop, knowledge-grounded reasoning.
• There is a need for advanced reasoning frameworks that mimic human cognitive processes for improved LLM accuracy and efficiency.

Purpose of the Study:

• Introduce PRIME (Planning and Retrieval-Integrated Memory for Enhanced Reasoning), a multi-agent framework.
• Dynamically integrate System 1 and System 2 thinking within an LLM.
• Enhance the reasoning capabilities of open-source LLMs to match state-of-the-art closed-source models.

Main Methods:

• PRIME employs a Quick Thinking Agent (System 1) for initial response generation.
• A System 2 reasoning pipeline with specialized agents (planning, hypothesis generation, retrieval, information integration, decision-making) is triggered upon detected uncertainty.
• The framework was experimentally evaluated using LLaMA 3 models.

Main Results:

• PRIME successfully mimics human cognitive processes, enhancing both efficiency and accuracy in LLM reasoning.
• Experiments demonstrated that PRIME-enhanced LLaMA 3 models achieve performance competitive with GPT-4 and GPT-4o.
• The framework shows significant improvements on benchmarks requiring multi-hop and knowledge-grounded reasoning.

Conclusions:

• PRIME offers a scalable solution for improving LLM reasoning in complex, knowledge-intensive domains.
• The dynamic integration of System 1 and System 2 thinking represents a significant advancement in LLM architecture.
• This research paves the way for more sophisticated and human-like AI reasoning capabilities.