LLM Agents Struggle with World Model Inference in Automata Learning

New research introduces 'agentic automata learning' to rigorously assess the capacity of tool-calling LLM agents to infer hidden environmental structures through interaction. By tasking agents with uncovering a deterministic finite automaton (DFA) via membership and equivalence queries, the study establishes a scalable testbed for evaluating interactive discovery. The core finding reveals a sharp decline in LLM performance as DFA size increases, indicating a fundamental scalability challenge in their ability to construct internal 'world models'. While reasoning models demonstrate superior performance over non-reasoning counterparts, detailed trajectory analyses pinpoint recurring failures in critical cognitive processes such as query planning, evidence integration, and hypothesis construction. This work underscores that despite advancements, current LLM agents remain significantly less robust and efficient than classical algorithms for this type of interactive learning. This research is timely given the increasing focus on autonomous AI agents that must operate and learn in dynamic, unknown environments. The ability to infer and adapt to underlying system logic, or a 'world model,' is paramount for such agents to perform complex tasks beyond rote execution. The methodology provides a controlled, measurable framework to benchmark this capability, moving beyond subjective evaluations of agentic behavior. By comparing LLM agents against established automata-learning algorithms, the study provides a clear, objective measure of their current limitations and potential. The findings suggest that while LLMs excel at language generation and pattern recognition, their capacity for systematic, interactive discovery and model building is still nascent. The implications for AI development are substantial. The identified weaknesses in query planning and evidence integration highlight areas where current LLM architectures and training paradigms fall short for true autonomous learning. For AI agents to move beyond pre-trained knowledge and effectively navigate novel situations, they must develop more sophisticated mechanisms for active information gathering, coherent synthesis of observations, and iterative hypothesis refinement. Overcoming these limitations will be crucial for deploying AI agents in complex real-world scenarios requiring genuine understanding and adaptive intelligence, rather than just sophisticated pattern matching or retrieval.