LLMs Still Can’t Plan And Reason

I’ve written many times about LLMs' reasoning and planning capabilities. But I feel that people are still confused and keep telling me that LLMs can plan, sometimes even better than humans. In this blog, I will try to answer why LLMs can’t plan and also give you a little background on the formal definition of planning and reasoning. Most disagreements stem from the fact that a lot of people just jumped on the bandwagon of AI and have not studied AI formally, that’s why they keep mixing definitions and drawing up wrong conclusions about LLMs' capabilities, strengths, and weaknesses.

LLMs are good idea-generating machines but far from planning and reasoning.

Topics Covered

• The Source Of Confusion About LLMs Capabilities
• Defining Planning, Reasoning, And Formal Logic
• Debunking LLMs Reasoning and Planning Capabilities

1. LLMs can Plan And Reason, and that’s why they are good at code generation.
2. What about the emergence capabilities of LLMs?
3. What about Chain-of-thought, ReACT, and other agentic frameworks?
4. In-context learning surely helps
5. What if we finetuned LLMs with successful plans in the domain?
6. But LLMs won a silver medal in the Math Olympiad and are reaching close to human performance even in the ARC-AGI challenge
7. But LLMs can self-critique and that surely increases the performance