Definition
Plain language
A document that looks relevant to a question but doesn't actually contain the answer.
As stated in the literature
In retrieval and contrastive learning, a candidate that is superficially similar to the target but is not a correct match; used to stress-test whether a system genuinely verifies evidence.
Also called: hard negatives
Why it matters: These tempting look-alikes test whether a system really verifies its evidence or just grabs whatever superficially matches.
For example, a search returns an article titled almost exactly like the question but which never actually states the answer.
Heard on the show
“This paper builds its test corpora from hard negatives, documents a strong embedding model thinks look relevant.”Episode 198 — The Model That Knows the Answer and Can't Say It