# Thought Cloning: Learning to Think while Acting by Imitating Human Thinking

**Authors:** Shengran Hu, Jeff Clune  
**arXiv:** [2306.00323](https://arxiv.org/abs/2306.00323)  
**Date:** 2023-06-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.00323v1/

## Key points

- Language is a defining characteristic of human thinking
- AI agents have yet to achieve the same level of language use as humans
- Thought Cloning framework aims to train AI agents to think like humans do
- Thought Cloning focuses on cloning thoughts and reasoning processes, not just actions
- Researchers tested the effectiveness of Thought Cloning in a simulated environment called BabyAI BossLevel
- BabyAI BossLevel presents several challenges for AI agents, including partial observability, complex missions described in natural language, hard-to-explore mazes with multiple closed rooms and locked doors, and long-horizon planning.
- Results showed that Thought Cloning outperformed traditional Behavioral Cloning methods by learning much faster and exhibiting better performance on out-of-distribution test tasks.
- The agent's thoughts are observable in the Thought Cloning framework which provides important benefits for AI safety and interpretability.
- By training agents how to think as well as behave using the novel Imitation Learning framework of Thought Cloning creates safer and more powerful AI agents capable of handling novel situations.

## Summary

The ability to use language is often considered a defining characteristic of human thinking, enabling us to generalize, plan, adapt and explore in ways that artificial intelligence (AI) agents have yet to achieve. To address this cognitive gap, researchers have proposed the Thought Cloning framework for training AI agents to think like humans do. Unlike traditional behavioral cloning methods which focus on replicating human actions, Thought Cloning aims to clone the thoughts and reasoning processes behind those actions. In a recent study, researchers tested the effectiveness of Thought Cloning in a simulated environment called BabyAI BossLevel. This environment presents several challenges for AI agents including partial observability, complex missions described in natural language, hard-to-explore mazes with multiple closed rooms and locked doors as well as long-horizon planning. The results showed that Thought Cloning outperformed traditional Behavioral Cloning methods by learning much faster and exhibiting better performance on out-of-distribution test tasks. Additionally, because the agent's thoughts are observable in the Thought Cloning framework it provides important benefits for AI safety and interpretability by allowing for easier diagnosis of problems and steering or preventing unsafe behavior. Overall, by training agents how to think as well as behave using the novel Imitation Learning framework of Thought Cloning creates safer and more powerful AI agents capable of handling novel situations.