New Paper: Tool-Calling LLM Agents Struggle to Infer World Models From Interaction (Agentic Automata Learning)

arXiv (Stanovsky et al.)·low signal

An arXiv paper submitted June 15, 2026 (2606.16576, Menaged/Lior/Ravfogel/Aharoni/Stanovsky) tests whether LLM agents can uncover a hidden deterministic finite automaton via membership and equivalence queries — a scalable, controllable testbed with classic automata-learning algorithms as baselines. Performance drops sharply as DFA size grows, and while reasoning models markedly beat non-reasoning ones, trajectory analysis exposes recurring failures in query planning, evidence integration, and hypothesis construction. For builders, it's concrete evidence that current agents are weak at active environment exploration — relevant when designing agents expected to discover system behavior on their own.