Can large language models reason and plan?

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Natural Language Processing

Background:

• Humans possess self-correction abilities, refining their own erroneous judgments through critical self-assessment.
• The capacity for self-critiquing in artificial intelligence, particularly in Large Language Models (LLMs), remains an open question.

Purpose of the Study:

• To investigate whether Large Language Models (LLMs) exhibit self-critiquing and error correction capabilities analogous to human cognitive functions.
• To determine if current LLM architectures provide a basis for self-generated error correction.

Main Methods:

• Analysis of LLM outputs for instances of self-identified errors.
• Comparative studies of human versus LLM error correction mechanisms.
• Examination of internal LLM processes for self-monitoring functionalities.

Main Results:

• No evidence was found to support the assumption that LLMs can self-critique and correct their own erroneous outputs.
• LLM behavior does not currently mirror human self-correction processes.

Conclusions:

• The capability for self-critiquing and error correction, common in humans, is not demonstrated in current Large Language Models (LLMs).
• Further research is needed to explore potential future developments or alternative architectures that might enable such capabilities in AI.