Closed-World Resolution Against Tool Hallucination in LLM Agents
An academic paper published on arXiv that systematically investigates tool hallucination in LLM agents, proposes a five-category taxonomy (H1–H5), and introduces a closed-world resolution cascade defense mechanism.
Timeline (last 90 days)
The arXiv paper 'Closed-World Resolution Against Tool Hallucination in LLM Agents' reveals the problem of Tool Hallucination.
The researchers tested ten hosted models under two calling interfaces and recorded a total of 322 real hallucination cases.
A large model with 675B parameters performed on par with small models of 7–8B on tool hallucinations, indicating that model scale cannot solve tool hallucination.
All Facts (3)
The arXiv paper 'Closed-World Resolution Against Tool Hallucination in LLM Agents' reveals the problem of Tool Hallucination.
50%UnverifiedThe researchers tested ten hosted models under two calling interfaces and recorded a total of 322 real hallucination cases.
50%UnverifiedA large model with 675B parameters performed on par with small models of 7–8B on tool hallucinations, indicating that model scale cannot solve tool hallucination.
50%