In the work titled LLMs with Chain-of-Thought Are Non-Causal Reasoners, the application of Chain of Thought (CoT) prompts in LLMs is critically analyzed. The research indicates an unexpected frequency of correct answers despite incorrect CoTs, leading to questions about the causality in AI reasoning.

Highlights include:

• Investigation into the cause-effect relationships in LLMs through causal analysis.
• Identification of discrepancies between LLM reasoning and human cognition.
• Influence of in-context learning, supervised fine-tuning, and feedback on the causal structure.

The findings provide insightful perspectives on the nature of CoT in LLMs and suggest that the mechanistic nature of AI reasoning diverges significantly from human reasoning, indicating a further need for exploration in AI’s reasoning capabilities.