Language models rely on pretrained understanding over provided definitions
Research report. Peer-review status is not recorded in the source records below; consult the linked publication.
Researchers tested whether large language models actually apply definitions provided in prompts or rely on their pretrained understanding by having models annotate the same texts twice with different definitions of concepts.
Across six concepts and three large language models, switching between competing definitions produced greater changes in annotations than paraphrasing a single definition while keeping its requirements the same.
Sources behind this report
Analysis
The context
Models annotated the same texts twice with different definitions, with a swap of competing definitions changing more annotations on average than a paraphrase across six concepts and three LLMs.
Report history
No earlier revision dates are retained here. This does not establish that the report has never changed; read any dated updates and corrections.
Corrections
No corrections have been appended.