In recent years, large language models (LLMs) have become much better at handling many different tasks. More recently, researchers have noticed that LLMs may show reasoning skills, especially when they are very large. Reasoning is a key part of intelligence, but we still don't fully understand how AI models learn to reason or use it to solve complex problems. This has become a major focus for many research teams.

How LLMs Reason with Foundation Models

A study by Sun et al. (2023) looks at the latest progress in how foundation models (like LLMs) tackle reasoning tasks. This research also looks into reasoning in models that work with multiple types of data (like images and text) and AI agents that can operate on their own.

Reasoning tasks can include things like solving math problems, understanding logic, figuring out cause-and-effect relationships, and even reasoning with visual information. The study outlines various ways to improve reasoning in foundation models, such as using alignment training (where the model is guided towards correct answers) and in-context learning (where the model learns from examples given in the prompt).

How Do LLMs Learn to Reason?

LLMs can develop reasoning skills using different methods of prompting. Qiao et al. (2023) divided research on reasoning methods into two main types: reasoning-enhanced strategies and knowledge-enhancement reasoning. 

- Reasoning-enhanced strategies include prompt engineering (designing prompts that lead to better reasoning), process optimization (improving how the model works), and using external tools to help the model reason better. Examples of this are methods like Chain-of-Thought and Active-Prompt.
  
- Knowledge-enhancement reasoning focuses on giving the model more information to improve reasoning.

Huang et al. (2023) also summarized techniques to improve reasoning in LLMs, like GPT-3. These techniques range from fine-tuning the models on datasets that explain reasoning to using methods like chain-of-thought prompting (where the model is guided through each step of reasoning), breaking down problems into smaller parts, and using in-context learning (learning from examples provided).

Can LLMs Really Reason and Plan?

There is an ongoing debate about whether LLMs can truly reason and plan. Both reasoning and planning are key skills that could help LLMs handle complex tasks in areas like robotics and AI agents.

Subbarao Kambhampati (2024) wrote a position paper about this issue. In it, he explains that, based on his research and experiences, he does not believe LLMs actually "reason" or "plan" in the way humans understand these skills. Instead, what LLMs do is retrieve information from their vast training data, which can sometimes look like reasoning but isn’t really the same thing.