Closed-World Resolution Against Tool Hallucination in LLM Agents

arXiv:2609.19425v1 Announce Type: new Abstract: Tool-augmented large language model (LLM) agents fail in a way no tool-selection or tool-security method addresses: they call tools that do not exist and pass arguments no schema declares. Existing defenses either pick the right tool (selection) or constrain what an agent may do with real tools (gating), both of which presuppose the emitted call refers to a real tool at all. We show this is a structural blind spot: a hallucinated call is by construction not a decision any gate made, so no gate can reject it. This paper is primarily a measurement and benchmark study. We give a five-class taxonomy of tool hallucination (H1-H5) and, as a reference point, the Resolution Rung: a training-free, closed-world resolver (registry membership plus a signature check) whose interest is where it must sit, not what it computes. We prove hallucination defense must precede any causal gate, and characterize the one irreducible residue (borrowed arguments schema-indistinguishable from a valid call). Across ten hosted models under two invocation surfaces we measure 322 genuine hallucinations; fabricated-tool calls concentrate on the unconstrained raw-JSON surface (34 vs. 3), and model scale does not help (a 675B model matches a 7-8B one). We then extend to the Model Context Protocol, where merging several servers into one namespace creates hallucination surfaces a single registry cannot express (a second taxonomy, M1-M5); on the live MCP surface we measure 154 hallucinations, including from frontier models that were clean on the single-registry su

Study measures tool hallucination in LLM agents.

This paper identifies and measures hallucinated tool calls by LLM agents, a technical blind spot leading to errors. It proposes a benchmark and a resolution approach to detect hallucinations, enabling improvement. The direct societal impact is limited but points to better reliability in AI tool use. Evidence is primary but specialized.

This is a rigorous measurement study documenting a real, quantified failure mode in deployed agentic AI systems, showing scale does not fix it and multi-server merges make it worse. The implication is materially concerning for current agent deployments, though the paper also offers a partial mitigation and benchmark, tempering the severity.

This paper identifies and characterizes a technical limitation in LLM agents related to tool hallucination. It proposes a measurement framework and a conceptual defense, but does not make claims about societal implications. The findings are primarily technical contributions to AI safety and reliability.

The paper identifies a significant and unaddressed failure mode in tool-augmented LLMs. This failure mode could lead to unreliable or dangerous behavior. The problem is structural and exacerbated by merging models.

SOURCE     arXiv cs.AI (tier 1)
   ↓
DOCUMENT   c4177923-5cde-4699-a6ca-f6b29ffa3f99
           https://arxiv.org/abs/2609.19425
   ↓
EVIDENCE   2 extracted excerpts
   ↓
MODEL RUN  5 runs, methodology 1.0.1
   ↓
SCORE      -37  (Adverse)
   ↓
CONFIDENCE 83%