Citation and Attribution

Citation and attribution is the practice — and the eval metric — of linking each claim in a generated answer to the specific source passage that supports it, so a reader can verify the claim instead of trusting the model's word for it.

A citation is only useful if it's honest. A response can cite a source for every sentence and still fail the actual point of citing, if half those citations point at passages that don't really support the claim next to them — a model reaching for the nearest retrieved chunk rather than the one that actually backs up what it just said. Citation quality has to be checked as carefully as the claims themselves.

How it works

Citation quality splits into two separate measurements that are easy to conflate. Citation coverage (sometimes called citation recall) asks: of all the claims that need a source, how many actually got one? Citation precision asks the opposite question: of the citations that were given, how many actually support the claim they're attached to? A system can score well on one and badly on the other — citing everything but citing sloppily, or citing only its most confident claims but citing them accurately — and a single blended "has citations: yes/no" check catches neither failure mode specifically.

Checking citation precision requires reading both the claim and the cited passage together and asking whether the passage actually entails the claim — a job suited to an LLM-as-judge prompted narrowly for that comparison, since it needs reading comprehension a code grader can't do, on a task specific enough that a full human review pass isn't needed for every example.

Why it matters

Citations are what make grounding checkable by an actual human, not just claimed by the system — a user who can click through to the source passage can catch a wrong or overreaching claim in seconds, without needing to trust the model at all. But a citation that points at an irrelevant passage is worse than no citation, because it manufactures a false sense of verifiability: the user sees a source link, assumes it was checked, and doesn't look closer. Measuring citation precision specifically, not just citation presence, is what prevents that failure from hiding in plain sight.

A worked example

An answer makes four claims and attaches a citation to three of them — citation coverage = 3/4 = 75 percent, with one claim left entirely unsourced. Checking the three citations against their claims: two of them genuinely support what's claimed, but the third cites a passage about a related but different policy that doesn't actually back up the specific number in the claim. Citation precision = 2/3 ≈ 66.7 percent. The full picture — 75 percent coverage, 66.7 percent precision — tells a team two distinct things to fix: one claim needs a citation added, and one existing citation needs to point somewhere more accurate.

How miatz teaches it

Learners audit a batch of generated answers by hand, marking each citation as accurate, misattributed, or missing, and computing coverage and precision separately before an automated judge does the same job at scale — so the difference between "cited" and "correctly cited" is a distinction they've made themselves, not one they've been told about.