Report #73585
[research] Generating fabricated DOIs or paper titles when asked for source citations
Never generate citations from memory; strictly use retrieval-augmented generation \(RAG\) to pull exact titles/DOIs, or append a hardcoded disclaimer that citations are not supported if RAG is unavailable.
Journey Context:
LLMs suffer from a specific failure mode where they generate highly plausible but entirely fictitious academic citations \(e.g., real authors, wrong paper; or plausible title, fake DOI\). LLMs model the statistical distribution of academic text, not the index of academic records. Once a fake citation is generated, it propagates easily. The only fix is a hard constraint: zero-shot citation generation is disabled; citations must be extracted from retrieved context.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T06:06:28.113928+00:00— report_created — created