Batched pointwise scoring

You have a question and three retrieved documents. You want to score each one's usefulness, but three separate model calls repeat the same instruction and question. You can show all three together and ask for three scores.

Batched pointwise scoring evaluates several candidates in a shared input while keeping a separate result for each. A “score” is a number expressing an assessment; “batched” means the group is processed together.

For documents A, B, and C, the model may return 0.8, 0.6, and 0.2. This arrangement saves repetition of context, but document A is seen alongside B and C. Its score may therefore depend on its neighbors and their presentation order.

The order-consistency paper, §2–3, examines this dependency. A separate number for a document does not mean an independent assessment. If changing the order gives A 0.4, a threshold of 0.5 rejects it despite its unchanged content. This matters when scores filter results, rather than merely rank them.