Can LLM Agents Infer World Models? Evidence from Agentic Automata Learning (arxiv.org)

Researchers have introduced the concept of agentic automata learning to assess whether large language model (LLM) agents can infer hidden world models through interaction with an oracle. The study evaluates LLMs’ capabilities in uncovering hidden deterministic finite automata (DFA) using two types of queries: membership and equivalence. The findings revealed that while LLMs can sometimes successfully engage in interactive discovery, their performance declines as the complexity of the DFA increases. Notably, reasoning models outperformed non-reasoning ones, but both types exhibited significant shortcomings in query planning and evidence integration. This research is significant for the AI/ML community as it highlights the limitations of current LLMs in tasks requiring non-trivial interactive problem-solving and model inference. The performance gap when compared to classic automata-learning algorithms underscores the continuing challenges in making LLMs more robust and efficient in their reasoning processes. By establishing a structured testbed with controlled complexity, the study provides valuable insights into the potential and current deficiencies of LLMs in learning and uncovering complex world models, paving the way for future advancements in AI interactions.