Intel / article

Closed-World Resolution Against Tool Hallucination in LLM Agents

Source: arxiv.org
Author: Laxmipriya Ganesh Iyer
Date: Published Sep 18, 2026

Context

Before deciding whether an may use a tool, a system needs to establish that the requested tool and its arguments actually exist. This study examines invented tool names and undeclared arguments, and presents as a reference point a resolver that checks calls against a known registry and the tool's declared input structure before later permission checks. Research abstract.

The study also examines what happens when several servers contribute tools to a shared namespace. Colliding names and one definition hiding another introduce errors that a single-server registry does not capture. The resulting benchmark treats valid tool identity as a distinct reliability problem, including a residual case where inappropriate borrowed arguments still satisfy the declared input structure. Findings and limitation.

Key takeaways · AI-distilled

Tool-calling agents sometimes invoke tools that don't exist or pass arguments no schema declares -- a blind spot no tool-selection or tool-gating method can catch, since a hallucinated call was never a real decision a gate made.
Across ten hosted models and two invocation surfaces, the study measured 322 genuine tool hallucinations; fabricated-tool calls clustered heavily on the unconstrained raw-JSON surface (34) versus a more constrained surface (3).
Model scale didn't reduce rate: a 675-billion-parameter model hallucinated tools about as often as a 7-8B model.
Merging multiple MCP servers into one tool namespace creates new hallucination surfaces from name collisions and shadowing -- the study measured 154 such hallucinations, including from frontier models that were clean when tested against a single registry.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
hallucination — When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.

Read the source arxiv.org