Also known as: pointwise reranking, independent scoring
Pointwise scoring evaluates each (query, document) pair independently, producing one relevance score per pair. The dominant pattern for cross-encoder rerankers because it's simple, parallelizable, and produces calibrated scores.
Pointwise scoring is the simplest reranking shape: take a query and one document, output a number indicating how relevant that document is to that query. To rerank a candidate set of N documents, run the model N times in parallel — once per pair — and sort by the resulting scores.
Most production cross-encoder rerankers are pointwise (zerank-2 included). The interface is just score = model(query, document).
The pointwise model evaluates each pair in isolation. It can’t:
For those, you compose pointwise scoring with a downstream MMR or diversity step, or use a listwise approach (with all the costs that entails).
A classic challenge: pointwise targets (a single relevance score per (q, d)) are noisy because annotators disagree on absolute relevance. The zELO methodology sidesteps this by training on pairwise preferences and recovering pointwise targets via a Thurstone fit — same statistical trick that powers chess Elo rankings.
Ask two annotators “how relevant is doc D to query Q on a scale of 1-5?” and you’ll get correlations as low as 0.4-0.6 in real datasets. Humans anchor differently: one annotator’s “4” is another’s “3”, and a single annotator drifts across a session. Worse, the meaning of “3” depends on what they’ve seen recently — a mediocre doc looks great after a batch of irrelevant ones, terrible after a batch of perfect ones. Ask the same two annotators “is doc A or doc B more relevant to Q?” and inter-annotator agreement jumps to 0.85+. Pairwise comparisons cancel the absolute-scale drift. zELO collects pairwise judgments at scale, then uses a Thurstone-style maximum-likelihood fit to recover continuous pointwise scores that are consistent — and trains the production reranker via MSE regression on those fitted scores. Result: pointwise model with calibrated outputs, trained on data that was never directly pointwise.
Pointwise scoring trades global awareness for the one property production retrieval cannot give up: a calibrated, thresholdable, per-pair number.
Calibration requires that the score has meaning across queries — '0.7 means 70% relevant' should hold for query A and query B alike. Pointwise produces a per-pair score you can anchor to a target distribution; listwise output is just an ordering, with no cross-query signal.
They don't, directly. [zELO](/concepts/zelo/) collects much-less-noisy pairwise judgments, runs a [Thurstone fit](/concepts/thurstone-model/) to recover continuous Elo-style pointwise scores, then trains the production reranker via MSE regression on those fitted scores. Pairwise in, pointwise out.
Pointwise can't see other candidates, so it can't deduplicate near-duplicates or enforce diversity. The standard pattern is to compose pointwise scores with a downstream MMR-style diversity step, or fall back to a [listwise reranker](/concepts/listwise-reranking/) when ordering coherence dominates.
