Necessary and sufficient knowledge enhanced collaborative logical reasoning in LLMs.

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Natural Language Processing

Background:

• Large language models (LLMs) possess extensive knowledge but struggle with accurate reasoning due to insufficient or unnecessary information utilization.
• Failures in knowledge utilization lead to incorrect conclusions and flawed reasoning paths in LLMs.
• Existing logical reasoning paradigms in LLMs have inherent limitations impacting their reliability.

Purpose of the Study:

• To propose a novel collaborative logical reasoning (CLR) framework to address the reasoning limitations of LLMs.
• To enhance the accuracy and reliability of LLM-generated conclusions through improved knowledge utilization.
• To lay the groundwork for modeling human cognitive thinking processes in AI.

Main Methods:

• The CLR framework integrates deductive reasoning (evidence retrieval) to generate initial reasoning paths.
• Abductive reasoning (knowledge attribution) is employed to identify necessary conditions for validating reasoning paths.
• Reliable inductive reasoning is used to derive final conclusions after verifying reasoning paths.

Main Results:

• CLR demonstrates superior performance compared to existing baselines across multiple datasets.
• The framework shows effectiveness in identifying and self-correcting errors within LLM reasoning processes.
• CLR successfully combines multiple logical reasoning paradigms for enhanced outcomes.

Conclusions:

• The CLR framework effectively remedies inherent limitations in LLM logical reasoning paradigms.
• CLR enhances LLM accuracy, reliability, and error-correction capabilities.
• This work advances the development of AI systems that better model human cognitive reasoning.