/p/2026-10-26 · explainer
Paper explainer · 2610.30612 · Chauhan and Agarwal

Fusion wins the average. It loses the query.

Blending a keyword index into vector search is the default advice, and the averages back it: the best blend beat vector-only retrieval on 33 of 35 encoder-and-corpus pairs. Scored one query at a time, the same blend ranked the relevant document below whichever retriever was better alone on 13–48% of queries, and below plain vector-only on 8–21% — for an average gain of 0.008–0.051 on a 0-to-1 ranking score. Which queries pay is close to unguessable, so the decision has to be made per corpus, from a measurement, before you blend.

01 · Two numbers, one system

change the lens — the same seven corpora, three ways to score the same blend

02 · Where the tax comes from

score blend   sf(q) = α · norm(dense) + (1 − α) · norm(sparse)  · α = 0.7 by default, scores min-max normalised over each top 100
rank blend   srrf(q) = ∑ 1 / (60 + rank)  · score magnitude discarded
the tax   τ(q) = max( sparse, dense ) − blend  · positive means you should not have blended

step through which retriever found the answer

03 · Can you see it coming?

04 · The reranker does not refund it

05 · Your own retrieval stack illustrative

pick the corpus your own index behaves most like

Results

What the paper actually measured

What it does not show

In practice