Can AI Infer Invisible Rules on Its Own? The Question Posed by 'Agentic Automata Learning'

Imagine you have just been dropped into a complex maze you have never visited before. There is no map, no compass. You are given only one “query tool” that allows you to tap on the walls at junctions or ask, after passing through a path once, if it is the correct route. How quickly could you map out the entire structure of the maze using this tool?

Recently, AI researchers conducted an intriguing experiment to test how Large Language Model (LLM)-based AI agents perform in such situations—that is, whether they can discover the invisible laws of an environment (a World Model) on their own. [Source: Can LLM Agents Infer World Models? Evidence from Agentic Automata Learning]

Why is this important?

The AI we have used until now has been more like a “student,” learning from vast, well-organized datasets and finding answers within them. But future AI agents will be different. They need to become “explorers” who are dropped into unfamiliar environments, figuring out for themselves what is right or wrong and what actions lead to what results.

This research seeks to determine whether AI can go beyond memorizing prescribed answers and autonomously infer the hidden principles of complex systems. If AI masters these “principles of learning,” it could fundamentally change the way we live, from automatically grasping operational rules in complex industrial settings to discovering new laws during scientific experiments. [Source: Global AI Weekly - Issue 155]

Easy to Understand: AI’s ‘Detective Game’

Researchers have set up a new testbed called ‘Agentic Automata Learning.’ In simple terms, ‘automata’ are a set of mechanical rules where the state changes based on input. As an easy analogy, it is similar to giving an AI agent a locked secret safe and having it figure out the password pattern for the door lock on its own. [Source: Agentic Automata Learning]

AI agents figure out the rules of this ‘safe (environment)’ through two main types of queries:

1. Membership Queries: Asking, “Does this password (string) belong to the combination that opens this safe?” and verifying.
2. Equivalence Queries: Asking, “Is this rule I have discovered so far exactly the same as the rule that opens the entire password?” and if wrong, receiving feedback to revise it. [Source: Can LLM Agents Infer World Models? Evidence from Agentic Automata Learning]

Through this process, the AI constantly goes through trial and error, refining the structure of the environment. It is just like how we complete a whole picture by piecing together puzzle pieces one by one. [Source: Can LLM Agents Infer World Models? Evidence from Agentic Automata Learning]

Current Situation: How far have we come?

The study results are quite interesting. Current AI agents have fully demonstrated their potential as ‘explorers’ that can make intriguing discoveries while interacting with their environment. However, they are not perfect yet.

The researchers pointed out that AI agents lack the robustness or efficiency compared to ‘classical automata learning algorithms’ established over decades. In particular, it was found that the agent’s performance drops sharply as soon as the environment becomes slightly more complex. This means that the intelligence AI possesses is still confined to ‘empirical guessing,’ and it needs to further develop the ability to drill down into highly precise and logical rule systems. [Source: Can LLM Agents Infer World Models? Evidence from Agentic Automata Learning]

What happens next?

This research is the first step for AI to move away from pre-determined answer sheets and find answers on its own. While AI agents cannot immediately figure out all the complex physical laws of the real world on their own, experts believe that the proposed ‘Agentic Automata Learning’ will become an important benchmark for evaluating AI intelligence. [Source: Can LLM Agents Infer World Models? Evidence from Agentic Automata Learning]

Moving forward, we will witness the process of AI agents moving beyond simple ‘conversational partners’ to become ‘true intellectual companions’ that find laws and solve problems on their own in unfamiliar environments.

MindTickleBytes’ AI Reporter Perspective

The attempt to move beyond mere data memorization and have AI autonomously infer the laws of its environment is a critical step toward true intelligence. Although currently less efficient than classical algorithms, closing this gap will be key to perfecting the era of agents.

References

1. Reef Menaged et al., Can LLM Agents Infer World Models? Evidence from Agentic Automata Learning
2. Emergent Mind, Agentic Automata Learning
3. Hacker News, Evidence from Agentic Automata Learning
4. Modern Orange, Can LLM Agents Infer World Models?
5. Agent Brief, Engineering the Agentic Reality Wall
6. Hugging Face, Can LLM Agents Infer World Models?
7. arXiv Signals, Can LLM Agents Infer World Models?
8. Reef Menaged, Can LLM Agents Infer World Models? - Agentic Automata Learning
9. Emergent Mind, Can LLM Agents Infer World Models? Evidence from Agentic Automata Learning
10. Global AI Community, Global AI Weekly - Issue 155