Addressing Tool Hallucination in LLM Agents with Closed-World Resolution

The study identifies a structural blind spot where agents call non-existent tools or use invalid arguments. It proposes a training-free Closed-World Resolver that performs registry membership and signature checks to prevent hallucinated calls from bypassing existing security gates.