Limitation or failure · Computer science · General science

Reference hallucination becomes a measurable LLM failure mode

Reference hallucination treated as a measurable failure mode in LLM-generated research support.

Summary

The arXiv paper studies hallucinated book and article references as a tractable form of language-model hallucination. It shows that consistency checks, such as asking about cited authors or content, can help identify hallucinated references without consulting external databases.

AI role

Generated book and article references that could be checked for consistency and existence.

Narrative role

This caution event matters because literature reliability is a bottleneck for trusting AI research assistants.

Caveat

The paper addresses a specific citation-fabrication failure mode rather than all research-tool reliability problems.