Can LLM Agents Infer World Models? Evidence from Agentic Automata Learning

2026-06-15

AI Summary

This work proposes an automata learning–based evaluation framework to assess whether tool-augmented large language model (LLM) agents can infer hidden environments through interaction. The approach models environments as deterministic finite automata (DFAs) and introduces membership and equivalence queries to construct a controllable, quantifiable, and algorithmically grounded interactive testing platform. By integrating classical automata learning theory into LLM agent evaluation for the first time, the study reveals that current models perform reasonably well on small DFAs but exhibit significant performance degradation as automaton size increases. Although reasoning-capable models outperform their non-reasoning counterparts, they still display systematic deficiencies in query planning, evidence integration, and hypothesis formation.

Abstract

We propose agentic automata learning to evaluate the extent to which tool-calling LLM agents can uncover hidden environments through interaction. In our setup, an agent should uncover a hidden deterministic finite automaton (DFA) by interacting with an oracle through (1) membership queries ("Does this string belong to the target language?") and (2) equivalence queries ("Is this the target DFA?"). This yields a scalable testbed with controlled task complexity, measurable interaction efficiency, and strong baselines (classic automata-learning algorithms). Evaluating state-of-the-art LLMs, we find that performance drops sharply as DFA size increases. Reasoning models are markedly stronger than non-reasoning models, yet trajectory analyses reveal recurring failures in query planning, evidence integration, and hypothesis construction. Overall, our results show that current LLM agents can sometimes perform non-trivial interactive discovery, but remain far less robust and efficient than classic algorithms for the task.

Problem

• world models
• LLM agents
• automata learning
• interactive discovery
• deterministic finite automaton

Innovation

• agentic automata learning
• world model inference
• tool-calling LLM agents
• deterministic finite automaton
• interactive discovery