A Machine That Learns From Its Own Experience

Key Points

• This research explores how an active-inference agent, using a language model, learns from its own experiences.
• Developed an active-inference agent with a language model as its core component.
• Examined the agent's ability to discover new actions and override previous beliefs based on experiences.
• Analyzed the agent’s learning process through seven mutually-falsifying senses.
• The agent discovered actions not originally programmed, indicating flexibility in learning.
• It successfully recalibrated its beliefs, paying costs for incorrect assumptions.
• The agent transitioned from inability to solve tasks to successfully solving them through self-discovery.